NNDesign PDF

Neural
Network
Design
Hagan
Demuth
2nd Edition Beale
De Jess
Neural Network Design
2nd Edtion
Martin T. Hagan
Oklahoma State University
Stillwater, Oklahoma
Howard B. Demuth
University of Colorado
Boulder, Colorado
Mark Hudson Beale

MHB Inc.
Hayden, Idaho
Orlando De Jess
Consultant
Frisco, Texas
Copyright by Martin T. Hagan and Howard B. Demuth. All rights reserved. No part of the book
may be reproduced, stored in a retrieval system, or transcribed in any form or by any means -
electronic, mechanical, photocopying, recording or otherwise - without the prior permission of
Hagan and Demuth.
MTH
To Janet, Thomas, Daniel, Mom and Dad
HBD
To Hal, Katherine, Kimberly and Mary
MHB
To Leah, Valerie, Asia, Drake, Coral and Morgan
ODJ
To: Marisela, Mara Victoria, Manuel, Mam y Pap.
Neural Network Design, 2nd Edition, eBook
OVERHEADS and DEMONSTRATION PROGRAMS can be found at the following website:

hagan.okstate.edu/nnd.html
A somewhat condensed paperback version of this text can be ordered from Amazon.
Contents
Preface
Introduction
Objectives 1-1
History 1-2
Applications 1-5
Biological Inspiration 1-8
Further Reading 1-10
2 Neuron Model and Network Architectures

Objectives 2-1
Theory and Examples 2-2
Notation 2-2
Neuron Model 2-2
Single-Input Neuron 2-2
Transfer Functions 2-3
Multiple-Input Neuron 2-7
Network Architectures 2-9
A Layer of Neurons 2-9
Multiple Layers of Neurons 2-10
Recurrent Networks 2-13
Summary of Results 2-16
Solved Problems 2-20
Epilogue 2-22
Exercises 2-23
i
An Illustrative Example
3 Objectives 3-1
Problem Statement 3-2
Perceptron 3-3
Two-Input Case 3-4
Pattern Recognition Example 3-5
Hamming Network 3-8
Feedforward Layer 3-8
Recurrent Layer 3-9
Hopfield Network 3-12
Epilogue 3-15
Exercises 3-16
4 Perceptron Learning Rule

Objectives 4-1
Learning Rules 4-2
Perceptron Architecture 4-3
Single-Neuron Perceptron 4-5
Multiple-Neuron Perceptron 4-8
Perceptron Learning Rule 4-8
Test Problem 4-9
Constructing Learning Rules 4-10
Unified Learning Rule 4-12
Training Multiple-Neuron Perceptrons 4-13
Proof of Convergence 4-15
Notation 4-15
Proof 4-16
Limitations 4-18
Epilogue 4-33
Exercises 4-36
ii
Signal and Weight Vector Spaces
5 Objectives 5-1
Linear Vector Spaces 5-2
Linear Independence 5-4
Spanning a Space 5-5
Inner Product 5-6
Norm 5-7
Orthogonality 5-7
Gram-Schmidt Orthogonalization 5-8
Vector Expansions 5-9
Reciprocal Basis Vectors 5-10
Epilogue 5-26
Exercises 5-28
6 Linear Transformations for Neural Networks

Objectives 6-1
Linear Transformations 6-2
Matrix Representations 6-3
Change of Basis 6-6
Eigenvalues and Eigenvectors 6-10
Diagonalization 6-13
Epilogue 6-28
Exercises 6-30
iii
Supervised Hebbian Learning
7 Objectives 7-1
Linear Associator 7-3
The Hebb Rule 7-4
Performance Analysis 7-5
Pseudoinverse Rule 7-7
Application 7-10
Variations of Hebbian Learning 7-12
Epilogue 7-29
Exercises 7-31
8 Performance Surfaces and Optimum Points

Objectives 8-1
Taylor Series 8-2
Vector Case 8-4
Directional Derivatives 8-5
Minima 8-7
Necessary Conditions for Optimality 8-9
First-Order Conditions 8-10
Second-Order Conditions 8-11
Quadratic Functions 8-12
Eigensystem of the Hessian 8-13
Epilogue 8-34
Exercises 8-36
iv
Performance Optimization
9 Objectives 9-1
Steepest Descent 9-2
Stable Learning Rates 9-6
Minimizing Along a Line 9-8
Newtons Method 9-10
Conjugate Gradient 9-15
Epilogue 9-37
Exercises 9-39
10 Widrow-Hoff Learning
Objectives 10-1
ADALINE Network 10-2
Single ADALINE 10-3
Mean Square Error 10-4
LMS Algorithm 10-7
Analysis of Convergence 10-9
Adaptive Filtering 10-13
Adaptive Noise Cancellation 10-15
Echo Cancellation 10-21
Epilogue 10-40
Exercises 10-42
v
Backpropagation
11 Objectives 11-1
Multilayer Perceptrons 11-2
Pattern Classification 11-3
Function Approximation 11-4
The Backpropagation Algorithm 11-7
Performance Index 11-8
Chain Rule 11-9
Backpropagating the Sensitivities 11-11
Summary 11-13
Example 11-14
Batch vs. Incremental Training 11-17
Using Backpropagation 11-18
Choice of Network Architecture 11-18
Convergence 11-20
Generalization 11-22
Epilogue 11-41
Exercises 11-44
12 Variations on Backpropagation
Objectives 12-1
Drawbacks of Backpropagation 12-3
Performance Surface Example 12-3
Convergence Example 12-7
Heuristic Modifications of Backpropagation 12-9
Momentum 12-9
Variable Learning Rate 12-12
Numerical Optimization Techniques 12-14
Levenberg-Marquardt Algorithm 12-19
Epilogue 12-46
Exercises 12-50
vi
Generalization
13 Objectives 13-1
Methods for Improving Generalization 13-5
Estimating Generalization Error 13-6
Early Stopping 13-6
Regularization 13-8
Bayesian Analysis 13-10
Bayesian Regularization 13-12
Relationship Between Early Stopping
and Regularization 13-19
Epilogue 13-44
Exercises 13-47
14 Dynamic Networks D
Objectives 14-1
Layered Digital Dynamic Networks 14-3
Example Dynamic Networks 14-5
Principles of Dynamic Learning 14-8
Dynamic Backpropagation 14-12
Preliminary Definitions 14-12
Real Time Recurrent Learning 14-12
Backpropagation-Through-Time 14-22
Summary and Comments on
Dynamic Training 14-30
Epilogue 14-46
Exercises 14-48
vii
Associative Learning
15 Objectives 15-1
Simple Associative Network 15-3
Unsupervised Hebb Rule 15-5
Hebb Rule with Decay 15-7
Simple Recognition Network 15-9
Instar Rule 15-11
Kohonen Rule 15-15
Simple Recall Network 15-16
Outstar Rule 15-17
Epilogue 15-34
Exercises 15-37
16 Competitive Networks
Objectives 16-1
Hamming Network 16-3
Layer 1 16-3
Layer 2 16-4
Competitive Layer 16-5
Competitive Learning 16-7
Problems with Competitive Layers 16-9
Competitive Layers in Biology 16-10
Self-Organizing Feature Maps 16-12
Improving Feature Maps 16-15
Learning Vector Quantization 16-16
LVQ Learning 16-18
Improving LVQ Networks (LVQ2) 16-21
Epilogue 16-37
Exercises 16-39
viii
Radial Basis Networks
17 Objectives 17-1
Radial Basis Network 17-2
Global vs. Local 17-9
Training RBF Networks 17-10
Linear Least Squares 17-11
Orthogonal Least Squares 17-18
Clustering 17-23
Nonlinear Optimization 17-25
Other Training Techniques 17-26
Epilogue 17-35
Exercises 17-38
18 Grossberg Network
Objectives 18-1
Biological Motivation: Vision 18-3
Illusions 18-4
Vision Normalization 18-8
Basic Nonlinear Model 18-9
Two-Layer Competitive Network 18-12
Layer 1 18-13
Layer 2 18-17
Choice of Transfer Function 18-20
Learning Law 18-22
Relation to Kohonen Law 18-24
Epilogue 18-42
Exercises 18-45
ix
Adaptive Resonance Theory
19 Objectives 19-1
Overview of Adaptive Resonance 19-2
Layer 1 19-4
Steady State Analysis 19-6
Layer 2 19-10
Orienting Subsystem 19-13
Learning Law: L1-L2 19-17
Subset/Superset Dilemma 19-17
Learning Law 19-18
ART1 Algorithm Summary 19-21
Initialization 19-21
Algorithm 19-21
Other ART Architectures 19-23
Epilogue 19-45
Exercises 19-48
20 Stability
Objectives 20-1
Stability Concepts 20-3
Definitions 20-4
Lyapunov Stability Theorem 20-5
Pendulum Example 20-6
LaSalles Invariance Theorem 20-12
Definitions 20-12
Theorem 20-13
Example 20-14
Comments 20-18
Epilogue 20-28
Exercises 30
x
Hopfield Network
21 Objectives 21-1
Hopfield Model 21-3
Lyapunov Function 21-5
Invariant Sets 21-7
Example 21-7
Hopfield Attractors 21-11
Effect of Gain 21-12
Hopfield Design 21-16
Content-Addressable Memory 21-16
Hebb Rule 21-18
Lyapunov Surface 21-22
Epilogue 21-36
Exercises 21-40
22 Practical Training Issues

Objectives 22-1
Pre-Training Steps 22-3
Selection of Data 22-3
Data Preprocessing 22-5
Training the Network 22-13
Weight Initialization 22-13
Choice of Training Algorithm 22-14
Stopping Criteria 22-14
Choice of Performance Function 22-16
Committees of Networks 22-18
Post-Training Analysis 22-18
Fitting 22-18
Pattern Recognition 22-21
Clustering 22-23
Prediction 22-24
Overfitting and Extrapolation 22-27
Sensitivity Analysis 22-28
Epilogue 22-30
xi
Case Study 1:Function Approximation
23 Objectives 23-1
Description of the Smart Sensor System 23-2
Data Collection and Preprocessing 23-3
Selecting the Architecture 23-4
Validation 23-7
Data Sets 23-10
Epilogue 23-11
24 Case Study 2:Probability Estimation

Objectives 24-1
Description of the CVD Process 24-2
Validation 24-9
Data Sets 24-12
Epilogue 24-13
25 Case Study 3:Pattern Recognition

Objectives 25-1
Description of Myocardial Infarction Recognition 25-2
Validation 25-7
Data Sets 25-10
Epilogue 25-11
xii
Case Study 4: Clustering
26 Objectives 26-1
Description of the Forest Cover Problem 26-2
Validation 26-7
Data Sets 26-11
Epilogue 26-12
27 Case Study 5: Prediction

Objectives 27-1
Description of the Magnetic Levitation System 27-2
Validation 27-8
Data Sets 27-13
Epilogue 27-14
xiii
Appendices
A Bibliography
B Notation
C Software
I Index
xiv
Preface
This book gives an introduction to basic neural network architectures and
learning rules. Emphasis is placed on the mathematical analysis of these
networks, on methods of training them and on their application to practical
engineering problems in such areas as nonlinear regression, pattern recog-
nition, signal processing, data mining and control systems.
Every effort has been made to present material in a clear and consistent
manner so that it can be read and applied with ease. We have included
many solved problems to illustrate each topic of discussion. We have also
included a number of case studies in the final chapters to demonstrate
practical issues that arise when using neural networks on real world prob-
lems.
Since this is a book on the design of neural networks, our choice of topics
was guided by two principles. First, we wanted to present the most useful
and practical neural network architectures, learning rules and training
techniques. Second, we wanted the book to be complete in itself and to flow
easily from one chapter to the next. For this reason, various introductory
materials and chapters on applied mathematics are included just before
they are needed for a particular subject. In summary, we have chosen some
topics because of their practical importance in the application of neural
networks, and other topics because of their importance in explaining how
neural networks operate.
We have omitted many topics that might have been included. We have not,
for instance, made this book a catalog or compendium of all known neural
network architectures and learning rules, but have instead concentrated
on the fundamental concepts. Second, we have not discussed neural net-
work implementation technologies, such as VLSI, optical devices and par-
allel computers. Finally, we do not present the biological and psychological
foundations of neural networks in any depth. These are all important top-
ics, but we hope that we have done the reader a service by focusing on those
topics that we consider to be most useful in the design of neural networks
and by treating those topics in some depth.
This book has been organized for a one-semester introductory course in
neural networks at the senior or first-year graduate level. (It is also suit-
able for short courses, self-study and reference.) The reader is expected to
have some background in linear algebra, probability and differential equa-
tions.
P-1
Preface
Each chapter of the book is divided into the following sections: Objectives,
Theory and Examples, Summary of Results, Solved Problems, Epilogue,
Further Reading and Exercises. The Theory and Examples section compris-
es the main body of each chapter. It includes the development of fundamen-
tal ideas as well as worked examples (indicated by the icon shown here in
2 the left margin). The Summary of Results section provides a convenient
+2
listing of important equations and concepts and facilitates the use of the
book as an industrial reference. About a third of each chapter is devoted to
the Solved Problems section, which provides detailed examples for all key
concepts.
The following figure illustrates the dependencies among the chapters.
1 7
Introduction Performance 8 Supervised
Surfaces Hebb
2
Architectures
Peformance 9 Associative 15
Optimization Learning
Illustrative 3
Example 10 Competitive 16
Widrow-Hoff
Learning
Perceptron 4 11
Learning Rule Backpropagation 18
Grossberg
Signal and 5 Variations on 12 19

Weight Vector Backpropagation ART
Spaces
13 20
Generalization Stability
Linear 6
Transformations 14 21
for Neural Dynamic Hopfield
Networks Networks
Radial Basis 17
22
Networks Practical Training
Case Study 23 Case Study 24 Case Study 25 Case Study 27 Case Study 26
Function Probability Pattern Prediction Clustering
Approximation Estimation Recognition
Chapters 1 through 6 cover basic concepts that are required for all of the
remaining chapters. Chapter 1 is an introduction to the text, with a brief
historical background and some basic biology. Chapter 2 describes the ba-
P-2
sic neural network architectures. The notation that is introduced in this
chapter is used throughout the book. In Chapter 3 we present a simple pat-
tern recognition problem and show how it can be solved using three differ-
ent types of neural networks. These three networks are representative of
the types of networks that are presented in the remainder of the text. In
addition, the pattern recognition problem presented here provides a com-
mon thread of experience throughout the book.
Much of the focus of this book will be on methods for training neural net-
works to perform various tasks. In Chapter 4 we introduce learning algo-
rithms and present the first practical algorithm: the perceptron learning
rule. The perceptron network has fundamental limitations, but it is impor-
tant for historical reasons and is also a useful tool for introducing key con-
cepts that will be applied to more powerful networks in later chapters.
One of the main objectives of this book is to explain how neural networks
operate. For this reason we will weave together neural network topics with
important introductory material. For example, linear algebra, which is the
core of the mathematics required for understanding neural networks, is re-
viewed in Chapters 5 and 6. The concepts discussed in these chapters will
be used extensively throughout the remainder of the book.
Chapters 7, and 1519 describe networks and learning rules that are
heavily inspired by biology and psychology. They fall into two categories:
associative networks and competitive networks. Chapters 7 and 15 intro-
duce basic concepts, while Chapters 1619 describe more advanced net-
works.
Chapters 814 and 17 develop a class of learning called performance learn-
ing, in which a network is trained to optimize its performance. Chapters 8
and 9 introduce the basic concepts of performance learning. Chapters 10
13 apply these concepts to feedforward neural networks of increasing pow-
er and complexity, Chapter 14 applies them to dynamic networks and
Chapter 17 applies them to radial basis networks, which also use concepts
from competitive learning.
Chapters 20 and 21 discuss recurrent associative memory networks. These
networks, which have feedback connections, are dynamical systems. Chap-
ter 20 investigates the stability of these systems. Chapter 21 presents the
Hopfield network, which has been one of the most influential recurrent net-
works.
Chapters 2227 are different than the preceding chapters. Previous chap-
ters focus on the fundamentals of each type of network and their learning
rules. The focus is on understanding the key concepts. In Chapters 2227,
we discuss some practical issues in applying neural networks to real world
problems. Chapter 22 describes many practical training tips, and Chapters
2327 present a series of case studies, in which neural networks are ap-
plied to practical problems in function approximation, probability estima-
tion, pattern recognition, clustering and prediction.
P-3
Preface
Software
MATLAB is not essential for using this book. The computer exercises can
be performed with any available programming language, and the Neural
Network Design Demonstrations, while helpful, are not critical to under-
standing the material covered in this book.
However, we have made use of the MATLAB software package to supple-
ment the textbook. This software is widely available and, because of its ma-
trix/vector notation and graphics, is a convenient environment in which to
experiment with neural networks. We use MATLAB in two different ways.
First, we have included a number of exercises for the reader to perform in
MATLAB. Many of the important features of neural networks become ap-
parent only for large-scale problems, which are computationally intensive
and not feasible for hand calculations. With MATLAB, neural network al-
gorithms can be quickly implemented, and large-scale problems can be
2+2
tested conveniently. These MATLAB exercises are identified by the icon
ans = shown here to the left. (If MATLAB is not available, any other program-
4 ming language can be used to perform the exercises.)
The second way in which we use MATLAB is through the Neural Network
Design Demonstrations, which can be downloaded from the website
hagan.okstate.edu/nnd.html. These interactive demonstrations illustrate
important concepts in each chapter. After the software has been loaded into
the MATLAB directory on your computer (or placed on the MATLAB path),
it can be invoked by typing nnd at the MATLAB prompt. All
demonstrations are easily accessible from a master menu. The icon shown
here to the left identifies references to these demonstrations in the text.
The demonstrations require MATLAB or the student edition of MATLAB,
version 2010a or later. See Appendix C for specific information on using the
demonstration software.
Overheads
As an aid to instructors who are using this text, we have prepared a
companion set of overheads. Transparency masters (in Microsoft
Powerpoint format or PDF) for each chapter are available on the web at
hagan.okstate.edu/nnd.html.
P-4
Acknowledgments
Acknowledgments
We are deeply indebted to the reviewers who have given freely of their time
to read all or parts of the drafts of this book and to test various versions of
the software. In particular we are most grateful to Professor John Andreae,
University of Canterbury; Dan Foresee, AT&T; Dr. Carl Latino, Oklahoma
State University; Jack Hagan, MCI; Dr. Gerry Andeen, SRI; and Joan Mill-
er and Margie Jenks, University of Idaho. We also had constructive inputs
from our graduate students in ECEN 5733 at Oklahoma State University,
ENEL 621 at the University of Canterbury, INSA 0506 at the Institut Na-
tional des Sciences Appliques and ECE 5120 at the University of Colo-
rado, who read many drafts, tested the software and provided helpful
suggestions for improving the book over the years. We are also grateful to
the anonymous reviewers who provided several useful recommendations.
We wish to thank Dr. Peter Gough for inviting us to join the staff in the
Electrical and Electronic Engineering Department at the University of
Canterbury, Christchurch, New Zealand, and Dr. Andre Titli for inviting
us to join the staff at the Laboratoire d'Analyse et d'Architecture des
Systms, Centre National de la Recherche Scientifique, Toulouse, France.
Sabbaticals from Oklahoma State University and a years leave from the
University of Idaho gave us the time to write this book. Thanks to Texas
Instruments, Halliburton, Cummins, Amgen and NSF, for their support of
our neural network research. Thanks to The Mathworks for permission to
use material from the Neural Network Toolbox.
P-5
Objectives
1 Introduction
Objectives 1-1
History 1-2
Applications 1-5
Biological Inspiration 1-8
Objectives
As you read these words you are using a complex biological neural network.
You have a highly interconnected set of some 1011 neurons to facilitate your
reading, breathing, motion and thinking. Each of your biological neurons,
a rich assembly of tissue and chemistry, has the complexity, if not the
speed, of a microprocessor. Some of your neural structure was with you at
birth. Other parts have been established by experience.
Scientists have only just begun to understand how biological neural net-
works operate. It is generally understood that all biological neural func-
tions, including memory, are stored in the neurons and in the connections
between them. Learning is viewed as the establishment of new connections
between neurons or the modification of existing connections. This leads to
the following question: Although we have only a rudimentary understand-
ing of biological neural networks, is it possible to construct a small set of
simple artificial neurons and perhaps train them to serve a useful func-
tion? The answer is yes. This book, then, is about artificial neural net-
works.
The neurons that we consider here are not biological. They are extremely
simple abstractions of biological neurons, realized as elements in a pro-
gram or perhaps as circuits made of silicon. Networks of these artificial
neurons do not have a fraction of the power of the human brain, but they
can be trained to perform useful functions. This book is about such neu-
rons, the networks that contain them and their training.
1-1
1 Introduction
History
The history of artificial neural networks is filled with colorful, creative in-
dividuals from a variety of fields, many of whom struggled for decades to
develop concepts that we now take for granted. This history has been doc-
umented by various authors. One particularly interesting book is Neuro-
computing: Foundations of Research by John Anderson and Edward
Rosenfeld. They have collected and edited a set of some 43 papers of special
historical interest. Each paper is preceded by an introduction that puts the
paper in historical perspective.
Histories of some of the main neural network contributors are included at
the beginning of various chapters throughout this text and will not be re-
peated here. However, it seems appropriate to give a brief overview, a sam-
ple of the major developments.
At least two ingredients are necessary for the advancement of a technology:
concept and implementation. First, one must have a concept, a way of
thinking about a topic, some view of it that gives a clarity not there before.
This may involve a simple idea, or it may be more specific and include a
mathematical description. To illustrate this point, consider the history of
the heart. It was thought to be, at various times, the center of the soul or a
source of heat. In the 17th century medical practitioners finally began to
view the heart as a pump, and they designed experiments to study its
pumping action. These experiments revolutionized our view of the circula-
tory system. Without the pump concept, an understanding of the heart was
out of grasp.
Concepts and their accompanying mathematics are not sufficient for a
technology to mature unless there is some way to implement the system.
For instance, the mathematics necessary for the reconstruction of images
from computer-aided tomography (CAT) scans was known many years be-
fore the availability of high-speed computers and efficient algorithms final-
ly made it practical to implement a useful CAT system.
The history of neural networks has progressed through both conceptual in-
novations and implementation developments. These advancements, how-
ever, seem to have occurred in fits and starts rather than by steady
evolution.
Some of the background work for the field of neural networks occurred in
the late 19th and early 20th centuries. This consisted primarily of interdis-
ciplinary work in physics, psychology and neurophysiology by such scien-
tists as Hermann von Helmholtz, Ernst Mach and Ivan Pavlov. This early
work emphasized general theories of learning, vision, conditioning, etc.,
and did not include specific mathematical models of neuron operation.
1-2
History
The modern view of neural networks began in the 1940s with the work of
Warren McCulloch and Walter Pitts [McPi43], who showed that networks
of artificial neurons could, in principle, compute any arithmetic or logical
function. Their work is often acknowledged as the origin of the neural net-
work field.
McCulloch and Pitts were followed by Donald Hebb [Hebb49], who pro-
posed that classical conditioning (as discovered by Pavlov) is present be-
cause of the properties of individual neurons. He proposed a mechanism for
learning in biological neurons (see Chapter 7).
The first practical application of artificial neural networks came in the late
1950s, with the invention of the perceptron network and associated learn-
ing rule by Frank Rosenblatt [Rose58]. Rosenblatt and his colleagues built
a perceptron network and demonstrated its ability to perform pattern rec-
ognition. This early success generated a great deal of interest in neural net-
work research. Unfortunately, it was later shown that the basic perceptron
network could solve only a limited class of problems. (See Chapter 4 for
more on Rosenblatt and the perceptron learning rule.)
At about the same time, Bernard Widrow and Ted Hoff [WiHo60] intro-
duced a new learning algorithm and used it to train adaptive linear neural
networks, which were similar in structure and capability to Rosenblatts
perceptron. The Widrow-Hoff learning rule is still in use today. (See Chap-
ter 10 for more on Widrow-Hoff learning.)
Unfortunately, both Rosenblatts and Widrows networks suffered from the
same inherent limitations, which were widely publicized in a book by Mar-
vin Minsky and Seymour Papert [MiPa69]. Rosenblatt and Widrow were
aware of these limitations and proposed new networks that would over-
come them. However, they were not able to successfully modify their learn-
ing algorithms to train the more complex networks.
Many people, influenced by Minsky and Papert, believed that further re-
search on neural networks was a dead end. This, combined with the fact
that there were no powerful digital computers on which to experiment,
caused many researchers to leave the field. For a decade neural network re-
search was largely suspended.
Some important work, however, did continue during the 1970s. In 1972
Teuvo Kohonen [Koho72] and James Anderson [Ande72] independently
and separately developed new neural networks that could act as memories.
(See Chapter 15 and Chapter 16 for more on Kohonen networks.) Stephen
Grossberg [Gros76] was also very active during this period in the investi-
gation of self-organizing networks. (See Chapter 18 and Chapter 19.)
Interest in neural networks had faltered during the late 1960s because of
the lack of new ideas and powerful computers with which to experiment.
During the 1980s both of these impediments were overcome, and research
in neural networks increased dramatically. New personal computers and
1-3
1 Introduction
workstations, which rapidly grew in capability, became widely available. In

addition, important new concepts were introduced.
Two new concepts were most responsible for the rebirth of neural networks.
The first was the use of statistical mechanics to explain the operation of a
certain class of recurrent network, which could be used as an associative
memory. This was described in a seminal paper by physicist John Hopfield
[Hopf82]. (Chapter 20 and Chapter 21 discuss these Hopfield networks.)
The second key development of the 1980s was the backpropagation algo-
rithm for training multilayer perceptron networks, which was discovered
independently by several different researchers. The most influential publi-
cation of the backpropagation algorithm was by David Rumelhart and
James McClelland [RuMc86]. This algorithm was the answer to the criti-
cisms Minsky and Papert had made in the 1960s. (See Chapter 11 for a de-
velopment of the backpropagation algorithm.)
These new developments reinvigorated the field of neural networks. Since
the 1980s, thousands of papers have been written, neural networks have
found countless applications, and the field has been buzzing with new the-
oretical and practical work.
The brief historical account given above is not intended to identify all of the
major contributors, but is simply to give the reader some feel for how
knowledge in the neural network field has progressed. As one might note,
the progress has not always been slow but sure. There have been periods
of dramatic progress and periods when relatively little has been accom-
plished.
Many of the advances in neural networks have had to do with new con-
cepts, such as innovative architectures and training rules. Just as impor-
tant has been the availability of powerful new computers on which to test
these new concepts.
Well, so much for the history of neural networks to this date. The real ques-
tion is, What will happen in the future? Neural networks have clearly
taken a permanent place as important mathematical/engineering tools.
They dont provide solutions to every problem, but they are essential tools
to be used in appropriate situations. In addition, remember that we still
know very little about how the brain works. The most important advances
in neural networks almost certainly lie in the future.
The large number and wide variety of applications of this technology are
very encouraging. The next section describes some of these applications.
1-4
Applications
1
Applications
A newspaper article described the use of neural networks in literature re-
search by Aston University. It stated that the network can be taught to
recognize individual writing styles, and the researchers used it to compare
works attributed to Shakespeare and his contemporaries. A popular sci-
ence television program documented the use of neural networks by an Ital-
ian research institute to test the purity of olive oil. Google uses neural
networks for image tagging (automatically identifying an image and as-
signing keywords), and Microsoft has developed neural networks that can
help convert spoken English speech into spoken Chinese speech. Research-
ers at Lund University and Skne University Hospital in Sweden have
used neural networks to improve long-term survival rates for heart trans-
plant recipients by identifying optimal recipient and donor matches. These
examples are indicative of the broad range of applications that can be found
for neural networks. The applications are expanding because neural net-
works are good at solving problems, not just in engineering, science and
mathematics, but in medicine, business, finance and literature as well.
Their application to a wide variety of problems in many fields makes them
very attractive. Also, faster computers and faster algorithms have made it
possible to use neural networks to solve complex industrial problems that
formerly required too much computation.
The following note and Table of Neural Network Applications are repro-
duced here from the Neural Network Toolbox for MATLAB with the per-
mission of the MathWorks, Inc.
A 1988 DARPA Neural Network Study [DARP88] lists various neural net-
work applications, beginning with the adaptive channel equalizer in about
1984. This device, which is an outstanding commercial success, is a single-
neuron network used in long distance telephone systems to stabilize voice
signals. The DARPA report goes on to list other commercial applications,
including a small word recognizer, a process monitor, a sonar classifier and
a risk analysis system.
Thousands of neural networks have been applied in hundreds of fields in
the many years since the DARPA report was written. A list of some of those
applications follows.
Aerospace
High performance aircraft autopilots, flight path simulations,
aircraft control systems, autopilot enhancements, aircraft com-
ponent simulations, aircraft component fault detectors
1-5
1 Introduction
Automotive
Automobile automatic guidance systems, fuel injector control,
automatic braking systems, misfire detection, virtual emission
sensors, warranty activity analyzers
Banking
Check and other document readers, credit application evalua-
tors, cash forecasting, firm classification, exchange rate fore-
casting, predicting loan recovery rates, measuring credit risk
Defense
Weapon steering, target tracking, object discrimination, facial
recognition, new kinds of sensors, sonar, radar and image sig-
nal processing including data compression, feature extraction
and noise suppression, signal/image identification
Electronics
Code sequence prediction, integrated circuit chip layout, pro-
cess control, chip failure analysis, machine vision, voice syn-
thesis, nonlinear modeling
Entertainment
Animation, special effects, market forecasting
Financial
Real estate appraisal, loan advisor, mortgage screening, corpo-
rate bond rating, credit line use analysis, portfolio trading pro-
gram, corporate financial analysis, currency price prediction
Insurance
Policy application evaluation, product optimization
Manufacturing
Manufacturing process control, product design and analysis,
process and machine diagnosis, real-time particle identifica-
tion, visual quality inspection systems, beer testing, welding
quality analysis, paper quality prediction, computer chip qual-
ity analysis, analysis of grinding operations, chemical product
design analysis, machine maintenance analysis, project bid-
ding, planning and management, dynamic modeling of chemi-
cal process systems
1-6
Applications
Medical
1
Breast cancer cell analysis, EEG and ECG analysis, prosthesis
design, optimization of transplant times, hospital expense re-
duction, hospital quality improvement, emergency room test
advisement
Oil and Gas

Exploration, smart sensors, reservoir modeling, well treatment
decisions, seismic interpretation
Robotics
Trajectory control, forklift robot, manipulator controllers, vi-
sion systems, autonomous vehicles
Speech
Speech recognition, speech compression, vowel classification,
text to speech synthesis
Securities
Market analysis, automatic bond rating, stock trading advisory
systems
Telecommunications
Image and data compression, automated information services,
real-time translation of spoken language, customer payment
processing systems
Transportation
Truck brake diagnosis systems, vehicle scheduling, routing
systems
Conclusion
The number of neural network applications, the money that has been in-
vested in neural network software and hardware, and the depth and
breadth of interest in these devices is enormous.
1-7
1 Introduction
Biological Inspiration
The artificial neural networks discussed in this text are only remotely re-
lated to their biological counterparts. In this section we will briefly describe
those characteristics of brain function that have inspired the development
of artificial neural networks.
The brain consists of a large number (approximately 1011) of highly con-
nected elements (approximately 104 connections per element) called neu-
rons. For our purposes these neurons have three principal components: the
dendrites, the cell body and the axon. The dendrites are tree-like receptive
networks of nerve fibers that carry electrical signals into the cell body. The
cell body effectively sums and thresholds these incoming signals. The axon
is a single long fiber that carries the signal from the cell body out to other
neurons. The point of contact between an axon of one cell and a dendrite of
another cell is called a synapse. It is the arrangement of neurons and the
strengths of the individual synapses, determined by a complex chemical
process, that establishes the function of the neural network. Figure 1.1 is
a simplified schematic diagram of two biological neurons.
Dendrites
Axon
Cell Body
Synapse
Figure 1.1 Schematic Drawing of Biological Neurons

Some of the neural structure is defined at birth. Other parts are developed
through learning, as new connections are made and others waste away.
This development is most noticeable in the early stages of life. For example,
1-8
Biological Inspiration
it has been shown that if a young cat is denied use of one eye during a crit-
1
ical window of time, it will never develop normal vision in that eye. Lin-
guists have discovered that infants over six months of age can no longer
discriminate certain speech sounds, unless they were exposed to them ear-
lier in life [WeTe84].
Neural structures continue to change throughout life. These later changes
tend to consist mainly of strengthening or weakening of synaptic junctions.
For instance, it is believed that new memories are formed by modification
of these synaptic strengths. Thus, the process of learning a new friends
face consists of altering various synapses. Neuroscientists have discovered
[MaGa2000], for example, that the hippocampi of London taxi drivers are
significantly larger than average. This is because they must memorize a
large amount of navigational informationa process that takes more than
two years.
Artificial neural networks do not approach the complexity of the brain.
There are, however, two key similarities between biological and artificial
neural networks. First, the building blocks of both networks are simple
computational devices (although artificial neurons are much simpler than
biological neurons) that are highly interconnected. Second, the connections
between neurons determine the function of the network. The primary ob-
jective of this book will be to determine the appropriate connections to solve
particular problems.
It is worth noting that even though biological neurons are very slow when
compared to electrical circuits (10-3 s compared to 10-10 s), the brain is
able to perform many tasks much faster than any conventional computer.
This is in part because of the massively parallel structure of biological neu-
ral networks; all of the neurons are operating at the same time. Artificial
neural networks share this parallel structure. Even though most artificial
neural networks are currently implemented on conventional digital com-
puters, their parallel structure makes them ideally suited to implementa-
tion using VLSI, optical devices and parallel processors.
In the following chapter we will introduce our basic artificial neuron and
will explain how we can combine such neurons to form networks. This will
provide a background for Chapter 3, where we take our first look at neural
networks in action.
1-9
1 Introduction
Further Reading
[Ande72] J. A. Anderson, A simple neural network generating an in-
teractive memory, Mathematical Biosciences, Vol. 14, pp.
197220, 1972.
Anderson proposed a linear associator model for associa-
tive memory. The model was trained, using a generaliza-
tion of the Hebb postulate, to learn an association between
input and output vectors. The physiological plausibility of
the network was emphasized. Kohonen published a closely
related paper at the same time [Koho72], although the two
researchers were working independently.
[AnRo88] J. A. Anderson and E. Rosenfeld, Neurocomputing: Foun-
dations of Research, Cambridge, MA: MIT Press, 1989.
Neurocomputing is a fundamental reference book. It con-
tains over forty of the most important neurocomputing
writings. Each paper is accompanied by an introduction
that summarizes its results and gives a perspective on the
position of the paper in the history of the field.
[DARP88] DARPA Neural Network Study, Lexington, MA: MIT Lin-
coln Laboratory, 1988.
This study is a compendium of knowledge of neural net-
works as they were known to 1988. It presents the theoret-
ical foundations of neural networks and discusses their
current applications. It contains sections on associative
memories, recurrent networks, vision, speech recognition,
and robotics. Finally, it discusses simulation tools and im-
plementation technology.
[Gros76] S. Grossberg, Adaptive pattern classification and univer-
sal recoding: I. Parallel development and coding of neural
feature detectors, Biological Cybernetics, Vol. 23, pp. 121
134, 1976.
Grossberg describes a self-organizing neural network
based on the visual system. The network, which consists of
short-term and long-term memory mechanisms, is a contin-
uous-time competitive network. It forms a basis for the
adaptive resonance theory (ART) networks.
1-10
Further Reading
[Gros80] S. Grossberg, How does the brain build a cognitive code?

1
Psychological Review, Vol. 88, pp. 375407, 1980.
Grossbergs 1980 paper proposes neural structures and
mechanisms that can explain many physiological behaviors
including spatial frequency adaptation, binocular rivalry,
etc. His systems perform error correction by themselves,
without outside help.
[Hebb 49] D. O. Hebb, The Organization of Behavior. New York:
Wiley, 1949.
The main premise of this seminal book is that behavior can
be explained by the action of neurons. In it, Hebb proposed
one of the first learning laws, which postulated a mecha-
nism for learning at the cellular level.
Hebb proposes that classical conditioning in biology is
present because of the properties of individual neurons.
[Hopf82] J. J. Hopfield, Neural networks and physical systems with
emergent collective computational abilities, Proceedings
of the National Academy of Sciences, Vol. 79, pp. 2554
2558, 1982.
Hopfield describes a content-addressable neural network.
He also presents a clear picture of how his neural network
operates, and of what it can do.
[Koho72] T. Kohonen, Correlation matrix memories, IEEE Trans-
actions on Computers, vol. 21, pp. 353359, 1972.
Kohonen proposed a correlation matrix model for associa-
tive memory. The model was trained, using the outer prod-
uct rule (also known as the Hebb rule), to learn an
association between input and output vectors. The mathe-
matical structure of the network was emphasized. Ander-
son published a closely related paper at the same time
[Ande72], although the two researchers were working inde-
pendently.
[MaGa00] E. A. Maguire, D. G. Gadian, I. S. Johnsrude, C. D. Good, J.
Ashburner, R. S. J. Frackowiak, and C. D. Frith, Naviga-
tion-related structural change in the hippocampi of taxi
drivers, Proceedings of the National Academy of Sciences,
Vol. 97, No. 8, pp. 4398-4403, 2000.
Taxi drivers in London must undergo extensive training,
learning how to navigate between thousands of places in
the city. This training is colloquially known as being on
The Knowledge and takes about 2 years to acquire on av-
1-11
1 Introduction
erage. This study demonstrated that the posterior hippoc-

ampi of London taxi drivers were significantly larger
relative to those of control subjects.
[McPi43] W. McCulloch and W. Pitts, A logical calculus of the ideas
immanent in nervous activity, Bulletin of Mathematical
Biophysics., Vol. 5, pp. 115133, 1943.
This article introduces the first mathematical model of a
neuron, in which a weighted sum of input signals is com-
pared to a threshold to determine whether or not the neu-
ron fires. This was the first attempt to describe what the
brain does, based on computing elements known at the
time. It shows that simple neural networks can compute
any arithmetic or logical function.
[MiPa69] M. Minsky and S. Papert, Perceptrons, Cambridge, MA:
MIT Press, 1969.
A landmark book that contains the first rigorous study de-
voted to determining what a perceptron network is capable
of learning. A formal treatment of the perceptron was need-
ed both to explain the perceptrons limitations and to indi-
cate directions for overcoming them. Unfortunately, the
book pessimistically predicted that the limitations of per-
ceptrons indicated that the field of neural networks was a
dead end. Although this was not true it temporarily cooled
research and funding for research for several years.
[Rose58] F. Rosenblatt, The perceptron: A probabilistic model for
information storage and organization in the brain, Psycho-
logical Review, Vol. 65, pp. 386408, 1958.
Rosenblatt presents the first practical artificial neural net-
work the perceptron.
[RuMc86] D. E. Rumelhart and J. L. McClelland, eds., Parallel Dis-
tributed Processing: Explorations in the Microstructure of
Cognition, Vol. 1, Cambridge, MA: MIT Press, 1986.
One of the two key influences in the resurgence of interest
in the neural network field during the 1980s. Among other
topics, it presents the backpropagation algorithm for train-
ing multilayer networks.
[WeTe84] J. F. Werker and R. C. Tees, Cross-language speech per-
ception: Evidence for perceptual reorganization during the
first year of life, Infant Behavior and Development, Vol. 7,
pp. 49-63, 1984.
1-12
Further Reading
This work describes an experiment in which infants from

1
the Interior Salish ethnic group in British Columbia, and
other infants outside that group, were tested on their abil-
ity to discriminate two different sounds from the Thompson
language, which is spoken by the Interior Salish. The re-
searchers discovered that infants less than 6 or 8 months of
age were generally able to distinguish the sounds, whether
or not they were Interior Salish. By 10 to 12 months of age,
only the Interior Salish children were able to distinguish
the two sounds.
[WiHo60] B. Widrow and M. E. Hoff, Adaptive switching cir-
cuits,1960 IRE WESCON Convention Record, New York:
IRE Part 4, pp. 96104, 1960.
This seminal paper describes an adaptive perceptron-like
network that can learn quickly and accurately. The authors
assume that the system has inputs and a desired output
classification for each input, and that the system can calcu-
late the error between the actual and desired output. The
weights are adjusted, using a gradient descent method, so
as to minimize the mean square error. (Least Mean Square
error or LMS algorithm.)
This paper is reprinted in [AnRo88].
1-13
Objectives
2 Neuron Model and Network

Architectures
Objectives 2-1
Notation 2-2
Neuron Model 2-2
Single-Input Neuron 2-2
Transfer Functions 2-3
Multiple-Input Neuron 2-7
Network Architectures 2-9
A Layer of Neurons 2-9
Multiple Layers of Neurons 2-10
Epilogue 2-22
Exercises 2-23
Objectives
In Chapter 1 we presented a simplified description of biological neurons
and neural networks. Now we will introduce our simplified mathematical
model of the neuron and will explain how these artificial neurons can be in-
terconnected to form a variety of network architectures. We will also illus-
trate the basic operation of these networks through some simple examples.
The concepts and notation introduced in this chapter will be used through-
out this book.
This chapter does not cover all of the architectures that will be used in this
book, but it does present the basic building blocks. More complex architec-
tures will be introduced and discussed as they are needed in later chapters.
Even so, a lot of detail is presented here. Please note that it is not necessary
for the reader to memorize all of the material in this chapter on a first read-
ing. Instead, treat it as a sample to get you started and a resource to which
you can return.
2-1
Theory and Examples
Notation
Unfortunately, there is no single neural network notation that is universal-
ly accepted. Papers and books on neural networks have come from many di-
verse fields, including engineering, physics, psychology and mathematics,
and many authors tend to use vocabulary peculiar to their specialty. As a
result, many books and papers in this field are difficult to read, and con-
cepts are made to seem more complex than they actually are. This is a
shame, as it has prevented the spread of important new ideas. It has also
led to more than one reinvention of the wheel.
In this book we have tried to use standard notation where possible, to be
clear and to keep matters simple without sacrificing rigor. In particular, we
have tried to define practical conventions and use them consistently.
Figures, mathematical equations and text discussing both figures and
mathematical equations will use the following notation:
Scalars small italic letters: a,b,c
Vectors small bold nonitalic letters: a,b,c
Matrices capital BOLD nonitalic letters: A,B,C
Additional notation concerning the network architectures will be intro-
duced as you read this chapter. A complete list of the notation that we use
throughout the book is given in Appendix B, so you can look there if you
have a question.
Neuron Model
Single-Input Neuron
A single-input neuron is shown in Figure 2.1. The scalar input p is multi-
Weight plied by the scalar weight w to form wp , one of the terms that is sent to the
Bias summer. The other input, 1 , is multiplied by a bias b and then passed to
Net Input the summer. The summer output n , often referred to as the net input, goes
Transfer Function into a transfer function f , which produces the scalar neuron output a .
(Some authors use the term activation function rather than transfer func-
tion and offset rather than bias.)
If we relate this simple model back to the biological neuron that we dis-
cussed in Chapter 1, the weight w corresponds to the strength of a synapse,
the cell body is represented by the summation and the transfer function,
and the neuron output a represents the signal on the axon.
2-2
Neuron Model
Inputs General Neuron
w n a
p f
b
1
a = f (wp + b)
Figure 2.1 Single-Input Neuron

The neuron output is calculated as
a = f wp + b .
If, for instance, w = 3 , p = 2 and b = 1.5 , then
a = f 3 2 1.5 = f 4.5
The actual output depends on the particular transfer function that is cho-
sen. We will discuss transfer functions in the next section.
The bias is much like a weight, except that it has a constant input of 1.
However, if you do not want to have a bias in a particular neuron, it can be
omitted. We will see examples of this in Chapters 3, 7 and 16.
Note that w and b are both adjustable scalar parameters of the neuron.
Typically the transfer function is chosen by the designer and then the pa-
rameters w and b will be adjusted by some learning rule so that the neu-
ron input/output relationship meets some specific goal (see Chapter 4 for
an introduction to learning rules). As described in the following section, we
have different transfer functions for different purposes.
Transfer Functions
The transfer function in Figure 2.1 may be a linear or a nonlinear function
of n . A particular transfer function is chosen to satisfy some specification
of the problem that the neuron is attempting to solve.
A variety of transfer functions have been included in this book. Three of the
most commonly used functions are discussed below.
Hard Limit The hard limit transfer function, shown on the left side of Figure 2.2, sets
Transfer Function the output of the neuron to 0 if the function argument is less than 0, or 1 if
its argument is greater than or equal to 0. We will use this function to cre-
ate neurons that classify inputs into two distinct categories. It will be used
extensively in Chapter 4.
2-3
a a
+1 +1
n p
0 -b/w 0
-1 -1
a = hardlim (n) a = hardlim (wp + b)
Hard Limit Transfer Function Single-Input hardlim Neuron
Figure 2.2 Hard Limit Transfer Function

The graph on the right side of Figure 2.2 illustrates the input/output char-
acteristic of a single-input neuron that uses a hard limit transfer function.
Here we can see the effect of the weight and the bias. Note that an icon for
the hard limit transfer function is shown between the two figures. Such
icons will replace the general f in network diagrams to show the particular
transfer function that is being used.
Linear The output of a linear transfer function is equal to its input:
Transfer Function
a = n, (2.1)
as illustrated in Figure 2.3.

Neurons with this transfer function are used in the ADALINE networks,
which are discussed in Chapter 10.
a a
+1 +b
-b/w
n p
0 0
-1
a = purelin (n) a = purelin (wp + b)
Linear Transfer Function Single-Input purelin Neuron
Figure 2.3 Linear Transfer Function

The output ( a ) versus input ( p ) characteristic of a single-input linear neu-
ron with a bias is shown on the right of Figure 2.3.
Log-Sigmoid The log-sigmoid transfer function is shown in Figure 2.4.
Transfer Function
2-4
Neuron Model
a a
+1 +1
n p
0 -b/w 0
-1 -1
a = logsig (n) a = logsig (wp + b)
Log-Sigmoid Transfer Function Single-Input logsig Neuron
Figure 2.4 Log-Sigmoid Transfer Function

This transfer function takes the input (which may have any value between
plus and minus infinity) and squashes the output into the range 0 to 1, ac-
cording to the expression:
1
-.
a = ---------------
n
(2.2)
1+e
The log-sigmoid transfer function is commonly used in multilayer networks

that are trained using the backpropagation algorithm, in part because this
function is differentiable (see Chapter 11).
Most of the transfer functions used in this book are summarized in Table
2.1. Of course, you can define other transfer functions in addition to those
shown in Table 2.1 if you wish.
To experiment with a single-input neuron, use the Neural Network Design
Demonstration One-Input Neuron nnd2n1.
2-5
MATLAB
Name Input/Output Relation Icon Function
a = 0 n0
Hard Limit hardlim
a = 1 n0
a = 1 n0
Symmetrical Hard Limit hardlims
a = +1 n0
Linear a = n purelin
a = 0 n0
Saturating Linear a = n 0n1 satlin
a = 1 n1
a = 1 n 1
Symmetric Saturating
a = n 1 n 1 satlins
Linear
a = 1 n1
1
Log-Sigmoid a = ---------------
n
- logsig
1+e
n n
Hyperbolic Tangent e e
a = -----------------
- tansig
Sigmoid n
e +e
n
a = 0 n0
Positive Linear poslin
a = n 0n
Competitive
a = 1
a = 0
neuron with max n
all other neurons
C compet
Table 2.1 Transfer Functions
2-6
Neuron Model
Multiple-Input Neuron
Typically, a neuron has more than one input. A neuron with R inputs is
shown in Figure 2.5. The individual inputs p 1 ,p 2 ,... ,p R are each weighted by
Weight Matrix corresponding elements w 1 1 ,w 1 2 ,... ,w 1 R of the weight matrix W .
Inputs Multiple-Input Neuron
p1 w1, 1
p2
p3 n a
f
w1, R b
pR
1
a = f (Wp + b)
Figure 2.5 Multiple-Input Neuron

The neuron has a bias b , which is summed with the weighted inputs to
form the net input n :
n = w 1 1 p 1 + w 1 2 p 2 + ... + w 1 R p R + b . (2.3)
This expression can be written in matrix form:
n = Wp + b , (2.4)
where the matrix W for the single neuron case has only one row.
Now the neuron output can be written as
a = f Wp + b . (2.5)
Fortunately, neural networks can often be described with matrices. This

kind of matrix expression will be used throughout the book. Dont be con-
cerned if you are rusty with matrix and vector operations. We will review
these topics in Chapters 5 and 6, and we will provide many examples and
solved problems that will spell out the procedures.
We have adopted a particular convention in assigning the indices of the el-
ements of the weight matrix. The first index indicates the particular neu-
Weight Indices ron destination for that weight. The second index indicates the source of
the signal fed to the neuron. Thus, the indices in w 1 2 say that this weight
represents the connection to the first (and only) neuron from the second
source. Of course, this convention is more useful if there is more than one
neuron, as will be the case later in this chapter.
2-7
We would like to draw networks with several neurons, each having several
inputs. Further, we would like to have more than one layer of neurons. You
can imagine how complex such a network might appear if all the lines were
drawn. It would take a lot of ink, could hardly be read, and the mass of de-
tail might obscure the main features. Thus, we will use an abbreviated no-
Abbreviated Notation tation. A multiple-input neuron using this notation is shown in Figure 2.6.
Input Multiple-Input Neuron
p a
Rx1 W 1x1
n
1xR
1x1 f
1 b
1x1
R 1
a = f (Wp + b)
Figure 2.6 Neuron with R Inputs, Abbreviated Notation
As shown in Figure 2.6, the input vector p is represented by the solid ver-
tical bar at the left. The dimensions of p are displayed below the variable
as R 1 , indicating that the input is a single vector of R elements. These
inputs go to the weight matrix W , which has R columns but only one row
in this single neuron case. A constant 1 enters the neuron as an input and
is multiplied by a scalar bias b . The net input to the transfer function f is
n , which is the sum of the bias b and the product Wp . The neurons output
a is a scalar in this case. If we had more than one neuron, the network out-
put would be a vector.
The dimensions of the variables in these abbreviated notation figures will
always be included, so that you can tell immediately if we are talking about
a scalar, a vector or a matrix. You will not have to guess the kind of variable
or its dimensions.
Note that the number of inputs to a network is set by the external specifi-
cations of the problem. If, for instance, you want to design a neural network
that is to predict kite-flying conditions and the inputs are air temperature,
wind velocity and humidity, then there would be three inputs to the net-
work.
To experiment with a two-input neuron, use the Neural Network Design
Demonstration Two-Input Neuron (nnd2n2).
2-8
Network Architectures
Commonly one neuron, even with many inputs, may not be sufficient. We
might need five or ten, operating in parallel, in what we will call a layer.
This concept of a layer is discussed below.
A Layer of Neurons
Layer A single-layer network of S neurons is shown in Figure 2.7. Note that each
of the R inputs is connected to each of the neurons and that the weight ma-
trix now has S rows.
Inputs Layer of S Neurons
n1 a1
w1,1
p1 f
b1
p2 1
n2 a2
p3
f
b2
1
pR nS aS
wS, R f
bS
1
a = f(Wp + b)
Figure 2.7 Layer of S Neurons

The layer includes the weight matrix, the summers, the bias vector b , the
transfer function boxes and the output vector a . Some authors refer to the
inputs as another layer, but we will not do that here.
Each element of the input vector p is connected to each neuron through the
weight matrix W . Each neuron has a bias b i , a summer, a transfer func-
tion f and an output a i . Taken together, the outputs form the output vector
a.
It is common for the number of inputs to a layer to be different from the
number of neurons (i.e., R S ).
You might ask if all the neurons in a layer must have the same transfer
function. The answer is no; you can define a single (composite) layer of neu-
rons having different transfer functions by combining two of the networks
2-9
shown above in parallel. Both networks would have the same inputs, and
each network would create some of the outputs.
The input vector elements enter the network through the weight matrix
W:
w 1 1 w 1 2 w 1 R
w 2 1 w 2 2 w 2 R
W =

w S 1 w S 2 w S R
. (2.6)
As noted previously, the row indices of the elements of matrix W indicate
the destination neuron associated with that weight, while the column indi-
ces indicate the source of the input for that weight. Thus, the indices in
w 3 2 say that this weight represents the connection to the third neuron
from the second source.
Fortunately, the S-neuron, R-input, one-layer network also can be drawn in
abbreviated notation, as shown in Figure 2.8.
Input Layer of S Neurons
p a
Rx1 W Sx1
n
SxR
Sx1 f
1 b
Sx1
R S
a = f(Wp + b)
Figure 2.8 Layer of S Neurons, Abbreviated Notation
Here again, the symbols below the variables tell you that for this layer, p
is a vector of length R , W is an S R matrix, and a and b are vectors of
length S . As defined previously, the layer includes the weight matrix, the
summation and multiplication operations, the bias vector b , the transfer
function boxes and the output vector.
Multiple Layers of Neurons

Now consider a network with several layers. Each layer has its own weight
matrix W , its own bias vector b , a net input vector n and an output vector
a . We need to introduce some additional notation to distinguish between
2-10
these layers. We will use superscripts to identify the layers. Specifically, we

Layer Superscript append the number of the layer as a superscript to the names for each of
1
these variables. Thus, the weight matrix for the first layer is written as W ,
2
and the weight matrix for the second layer is written as W . This notation
is used in the three-layer network shown in Figure 2.9.
Inputs First Layer Second Layer Third Layer
w 11,1 n11 a11 w 21,1 n21 a21 w 31,1 n31 a31

p1
f1 f2 f3
b11 b21 b31
p2 1 1 1
n12 a12 n22 a22 n32 a32
f 1
f2 f 3
p3 b12 b22 b32
1 1 1
pR n1S 1 a1S 1 n2S 2 a2S 2 n3S 3 a3S 3
w 1S 1, R f1
w 2S 2, S 1
f2
w 3S 3, S 2
f3
b1S 1 b2S 2 b3S 3
1 1 1
a1 = f 1 (W1p + b1) a2 = f 2 (W2a1 + b2) a3 = f 3 (W3a2 + b3)

a3 = f 3 (W3f 2 (W2f 1 (W1p + b1) + b2) + b3)
Figure 2.9 Three-Layer Network

1 2
As shown, there are R inputs, S neurons in the first layer, S neurons in
the second layer, etc. As noted, different layers can have different numbers
of neurons.
The outputs of layers one and two are the inputs for layers two and three.
1
Thus layer 2 can be viewed as a one-layer network with R = S inputs,
2 2 1 2
S = S neurons, and an S S weight matrix W . The input to layer 2 is
1 2
a , and the output is a .
Output Layer A layer whose output is the network output is called an output layer. The
Hidden Layers other layers are called hidden layers. The network shown above has an out-
put layer (layer 3) and two hidden layers (layers 1 and 2).
The same three-layer network discussed previously also can be drawn us-
ing our abbreviated notation, as shown in Figure 2.10.
2-11
Input First Layer Second Layer Third Layer
p a1 a2 a3
Rx1
W1 S1 x 1
W2 S2 x 1
W3 S3 x 1
n1 n2 n3
S1 x R
S1 x 1
f1 S2 x S1
S2 x 1
f2 S3 x S2
S3 x 1
f3
1 b1 1 b2 1 b3
S1 x 1 S2 x 1 S3 x 1
R S1 S2 S3

a3 = f 3 (W3 f 2 (W2f 1 (W1p + b1) + b2) + b3)
Figure 2.10 Three-Layer Network, Abbreviated Notation

Multilayer networks are more powerful than single-layer networks. For in-
stance, a two-layer network having a sigmoid first layer and a linear second
layer can be trained to approximate most functions arbitrarily well. Single-
layer networks cannot do this.
At this point the number of choices to be made in specifying a network may
look overwhelming, so let us consider this topic. The problem is not as bad
as it looks. First, recall that the number of inputs to the network and the
number of outputs from the network are defined by external problem spec-
ifications. So if there are four external variables to be used as inputs, there
are four inputs to the network. Similarly, if there are to be seven outputs
from the network, there must be seven neurons in the output layer. Finally,
the desired characteristics of the output signal also help to select the trans-
fer function for the output layer. If an output is to be either 1 or 1 , then
a symmetrical hard limit transfer function should be used. Thus, the archi-
tecture of a single-layer network is almost completely determined by prob-
lem specifications, including the specific number of inputs and outputs and
the particular output signal characteristic.
Now, what if we have more than two layers? Here the external problem
does not tell you directly the number of neurons required in the hidden lay-
ers. In fact, there are few problems for which one can predict the optimal
number of neurons needed in a hidden layer. This problem is an active area
of research. We will develop some feeling on this matter as we proceed to
Chapter 11, Backpropagation.
As for the number of layers, most practical neural networks have just two
or three layers. Four or more layers are used rarely.
We should say something about the use of biases. One can choose neurons
with or without biases. The bias gives the network an extra variable, and
so you might expect that networks with biases would be more powerful
2-12
than those without, and that is true. Note, for instance, that a neuron with-
out a bias will always have a net input n of zero when the network inputs
p are zero. This may not be desirable and can be avoided by the use of a
bias. The effect of the bias is discussed more fully in Chapters 3, 4 and 5.
In later chapters we will omit a bias in some examples or demonstrations.
In some cases this is done simply to reduce the number of network param-
eters. With just two variables, we can plot system convergence in a two-di-
mensional plane. Three or more variables are difficult to display.
Recurrent Networks
Before we discuss recurrent networks, we need to introduce some simple
Delay building blocks. The first is the delay block, which is illustrated in Figure
2.11.
Delay
u(t) a(t)
D
a(0)
a(t) = u(t - 1)
Figure 2.11 Delay Block

The delay output a t is computed from its input u t according to
at = u t 1 . (2.7)
Thus the output is the input delayed by one time step. (This assumes that
time is updated in discrete steps and takes on only integer values.) Eq. (2.7)
requires that the output be initialized at time t = 0 . This initial condition
is indicated in Figure 2.11 by the arrow coming into the bottom of the delay
block.
Another related building block, which we will use for the continuous-time
Integrator recurrent networks in Chapters 1821, is the integrator, which is shown in
Figure 2.12.
2-13
Integrator
u(t) a(t)
a(0)
t
a(t) = u() d + a(0)
0
Figure 2.12 Integrator Block

The integrator output a t is computed from its input u t according to
t
at = 0 u d + a 0 . (2.8)
The initial condition a 0 is indicated by the arrow coming into the bottom
of the integrator block.
Recurrent Network We are now ready to introduce recurrent networks. A recurrent network is
a network with feedback; some of its outputs are connected to its inputs.
This is quite different from the networks that we have studied thus far,
which were strictly feedforward with no backward connections. One type of
discrete-time recurrent network is shown in Figure 2.13.
Initial
Condition Recurrent Layer
p
Sx1 W n(t + 1) a(t + 1) a(t)
SxS
Sx1 Sx1 D Sx1
1 b
Sx1
S S
a(0) = p a(t + 1) = satlins (Wa(t) + b)
Figure 2.13 Recurrent Network
2-14
In this particular network the vector p supplies the initial conditions (i.e.,
a 0 = p ). Then future outputs of the network are computed from previous
outputs:
a 1 = satlins Wa 0 + b , a 2 = satlins Wa 1 + b , . . . 2
Recurrent networks are potentially more powerful than feedforward net-
works and can exhibit temporal behavior. These types of networks are dis-
cussed in Chapters 3, 14 and 1821.
2-15
Summary of Results
Single-Input Neuron
Inputs General Neuron
w n a
p f
b
1
a = f (wp + b)
Multiple-Input Neuron
Inputs Multiple-Input Neuron
p1 w1, 1
p2
p3 n a
f
w1, R b
pR
1
a = f (Wp + b)
Input Multiple-Input Neuron
p a
Rx1 W 1x1
n
1xR
1x1 f
1 b
1x1
R 1
a = f (Wp + b)
2-16
Summary of Results
Transfer Functions
MATLAB
Name Input/Output Relation Icon Function 2
a = 0 n0
Hard Limit hardlim
a = 1 n0
a = 1 n0
Symmetrical Hard Limit hardlims
a = +1 n0
Linear a = n purelin
a = 0 n0
Saturating Linear a = n 0n1 satlin
a = 1 n1
a = 1 n 1
a = n 1 n 1 satlins
Linear
a = 1 n1
1
Log-Sigmoid a = ---------------
n
- logsig
1+e
n n
Hyperbolic Tangent e e
a = -----------------
- tansig
Sigmoid n
e +e
n
a = 0 n0
Positive Linear poslin
a = n 0n
Competitive
a = 1
a = 0
neuron with max n
all other neurons
C compet
2-17
Layer of Neurons
Input Layer of S Neurons
p a
Rx1 W Sx1
n
SxR
Sx1 f
1 b
Sx1
R S
a = f(Wp + b)
Three Layers of Neurons

p a1 a2 a3
Rx1
W1 S1 x 1
W2 S2 x 1
W3 S3 x 1
n1 n2 n3
S1 x R
S1 x 1
f1 S2 x S1
S2 x 1
f2 S3 x S2
S3 x 1
f3
1 b1 1 b2 1 b3
S1 x 1 S2 x 1 S3 x 1
R S1 S2 S3

a3 = f 3 (W3 f 2 (W2f 1 (W1p + b1) + b2) + b3)
Delay
Delay
u(t) a(t)
D
a(0)
a(t) = u(t - 1)
2-18
Summary of Results
Integrator
Integrator
u(t) a(t)
2
a(0)
t
a(t) = u() d + a(0)
0
Recurrent Network
Initial
p
Sx1 W n(t + 1) a(t + 1) a(t)
SxS
Sx1 Sx1 D Sx1
1 b
Sx1
S S
How to Pick an Architecture

Problem specifications help define the network in the following ways:
1. Number of network inputs = number of problem inputs
2. Number of neurons in output layer = number of problem outputs
3. Output layer transfer function choice at least partly determined by
problem specification of the outputs
2-19
Solved Problems
P2.1 The input to a single-input neuron is 2.0, its weight is 2.3 and its
bias is -3.
i. What is the net input to the transfer function?
ii. What is the neuron output?
i. The net input is given by:
n = wp + b = 2.3 2 + 3 = 1.6
ii. The output cannot be determined because the transfer function is not
specified.
P2.2 What is the output of the neuron of P2.1 if it has the following
transfer functions?
i. Hard limit
ii. Linear
iii. Log-sigmoid
i. For the hard limit transfer function:
a = hardlim 1.6 = 1.0
ii. For the linear transfer function:
a = purelin 1.6 = 1.6
iii. For the log-sigmoid transfer function:
1
a = logsig 1.6 = ------------------
1.6
- = 0.8320
1+e
2+2
ans = Verify this result using MATLAB and the function logsig, which is in the
4 MININNET directory (see Appendix B).
P2.3 Given a two-input neuron with the following parameters: b = 1.2 ,

T
W = 3 2 and p = 5 6 , calculate the neuron output for the fol-
lowing transfer functions:
i. A symmetrical hard limit transfer function
ii. A saturating linear transfer function
2-20
Solved Problems
iii. A hyperbolic tangent sigmoid (tansig) transfer function

First calculate the net input n :
n = Wp + b = 3 2 5 + 1.2 = 1.8 .
2
6
Now find the outputs for each of the transfer functions.

i. a = hardlims 1.8 = 1
ii. a = satlin 1.8 = 0
iii. a = tansig 1.8 = 0.9468
P2.4 A single-layer neural network is to have six inputs and two out-
puts. The outputs are to be limited to and continuous over the
range 0 to 1. What can you tell about the network architecture?
Specifically:
i. How many neurons are required?
ii. What are the dimensions of the weight matrix?
iii. What kind of transfer functions could be used?
iv. Is a bias required?
The problem specifications allow you to say the following about the net-
work.
i. Two neurons, one for each output, are required.
ii. The weight matrix has two rows corresponding to the two neurons and
six columns corresponding to the six inputs. (The product Wp is a two-el-
ement vector.)
iii. Of the transfer functions we have discussed, the logsig transfer func-
tion would be most appropriate.
iv. Not enough information is given to determine if a bias is required.
2-21
Epilogue
This chapter has introduced a simple artificial neuron and has illustrated
how different neural networks can be created by connecting groups of neu-
rons in various ways. One of the main objectives of this chapter has been to
introduce our basic notation. As the networks are discussed in more detail
in later chapters, you may wish to return to Chapter 2 to refresh your mem-
ory of the appropriate notation.
This chapter was not meant to be a complete presentation of the networks
we have discussed here. That will be done in the chapters that follow. We
will begin in Chapter 3, which will present a simple example that uses
some of the networks described in this chapter, and will give you an oppor-
tunity to see these networks in action. The networks demonstrated in
Chapter 3 are representative of the types of networks that are covered in
the remainder of this text.
2-22
Exercises
Exercises
E2.1 A single input neuron has a weight of 1.3 and a bias of 3.0. What possible
kinds of transfer functions, from Table 2.1, could this neuron have, if its
output is given below. In each case, give the value of the input that would
produce these outputs.
i. 1.6
ii. 1.0
iii. 0.9963
iv. -1.0
E2.2 Consider a single-input neuron with a bias. We would like the output to be
-1 for inputs less than 3 and +1 for inputs greater than or equal to 3.
i. What kind of a transfer function is required?
ii. What bias would you suggest? Is your bias in any way related to the
input weight? If yes, how?
2+2
ans = iii. Summarize your network by naming the transfer function and stat-
4 ing the bias and the weight. Draw a diagram of the network. Verify
the network performance using MATLAB.
E2.3 Given a two-input neuron with the following weight matrix and input vec-
T
tor: W = 3 2 and p = 5 7 , we would like to have an output of 0.5. Do
you suppose that there is a combination of bias and transfer function that
might allow this?
i. Is there a transfer function from Table 2.1 that will do the job if the
bias is zero?
ii. Is there a bias that will do the job if the linear transfer function is
used? If yes, what is it?
iii. Is there a bias that will do the job if a log-sigmoid transfer function
is used? Again, if yes, what is it?
iv. Is there a bias that will do the job if a symmetrical hard limit trans-
fer function is used? Again, if yes, what is it?
E2.4 A two-layer neural network is to have four inputs and six outputs. The
range of the outputs is to be continuous between 0 and 1. What can you tell
about the network architecture? Specifically:
2-23
i. How many neurons are required in each layer?

ii. What are the dimensions of the first-layer and second-layer weight
matrices?
iii. What kinds of transfer functions can be used in each layer?
iv. Are biases required in either layer?
E2.5 Consider the following neuron.
Input General Neuron
p w n a
S f
b
1
a = f (wp + b)
Figure P15.1 General Neuron

Sketch the neuron response (plot a versus p for -2<p<2) for the following
cases.
i. w = 1 , b = 1 , f = hardlims .
ii. w = 1 , b = 1 , f = hardlims .
iii. w = 2 , b = 3 , f = purelin .
iv. w = 2 , b = 3 , f = satlins .
v. w = 2 , b = 1 , f = poslin .
2-24
Exercises
E2.6 Consider the following neural network.
Input Sat. Linear Layer Linear Layer
w11,1 n11 a11 2

S w21,1
b11 n21 a21
p
1 1
S
n2 a12 b21
S
2
w 1,2
w12,1 1
1
b2
1
a = satlin(W p + b )
1 1 1 2 2 1 2
a = purelin(W a + b )
1 1 1 1 2 2 2
w 1 1 = 2 , w 2 1 = 1 , b 1 = 2 , b 2 = 1 , w 1 1 = 1 , w 1 2 = 1 , b 1 = 0
Sketch the following responses (plot the indicated variable versus p for
3 p 3 ).
1
i. n 1 .
1
ii. a 1 .
1
iii. n 2
1
iv. a 2 .
2
v. n 1 .
2
vi. a 1 .
2-25
Objectives
3 An Illustrative Example
Objectives 3-1
Perceptron 3-3
Two-Input Case 3-4
Pattern Recognition Example 3-5
Hamming Network 3-8
Feedforward Layer 3-9
Recurrent Layer 3-10
Hopfield Network 3-12
Epilogue 3-15
Exercises 3-16
Objectives
Think of this chapter as a preview of coming attractions. We will take a
simple pattern recognition problem and show how it can be solved using
three different neural network architectures. It will be an opportunity to
see how the architectures described in the previous chapter can be used to
solve a practical (although extremely oversimplified) problem. Do not ex-
pect to completely understand these three networks after reading this
chapter. We present them simply to give you a taste of what can be done
with neural networks, and to demonstrate that there are many different
types of networks that can be used to solve a given problem.
The three networks presented in this chapter are representative of the
types of networks discussed in the remaining chapters: feedforward net-
works (represented here by the perceptron), competitive networks (repre-
sented here by the Hamming network) and recurrent associative memory
networks (represented here by the Hopfield network).
3-1
Theory and Examples
Problem Statement
A produce dealer has a warehouse that stores a variety of fruits and vege-
tables. When fruit is brought to the warehouse, various types of fruit may
be mixed together. The dealer wants a machine that will sort the fruit ac-
cording to type. There is a conveyer belt on which the fruit is loaded. This
conveyer passes through a set of sensors, which measure three properties
of the fruit: shape, texture and weight. These sensors are somewhat primi-
tive. The shape sensor will output a 1 if the fruit is approximately round
and a 1 if it is more elliptical. The texture sensor will output a 1 if the sur-
face of the fruit is smooth and a 1 if it is rough. The weight sensor will
output a 1 if the fruit is more than one pound and a 1 if it is less than one
pound.
The three sensor outputs will then be input to a neural network. The pur-
pose of the network is to decide which kind of fruit is on the conveyor, so
that the fruit can be directed to the correct storage bin. To make the prob-
lem even simpler, lets assume that there are only two kinds of fruit on the
conveyor: apples and oranges.
Neural
Network
Sensors
Sorter
Apples Oranges
As each fruit passes through the sensors it can be represented by a three-

dimensional vector. The first element of the vector will represent shape,
the second element will represent texture and the third element will repre-
sent weight:
3-2
Perceptron
shape
p = texture . (3.1)
weight
Therefore, a prototype orange would be represented by
1
p1 = 1 , (3.2)
1
and a prototype apple would be represented by
1
p2 = 1 . (3.3)
1
The neural network will receive one three-dimensional input vector for
each fruit on the conveyer and must make a decision as to whether the fruit
is an orange p 1 or an apple p 2 .
Now that we have defined this simple (trivial?) pattern recognition prob-
lem, lets look briefly at three different neural networks that could be used
to solve it. The simplicity of our problem will facilitate our understanding
of the operation of the networks.
Perceptron
The first network we will discuss is the perceptron. Figure 3.1 illustrates a
single-layer perceptron with a symmetric hard limit transfer function hard-
lims.
Inputs Sym. Hard Limit

- Title - Layer
p a
Rx1
W Sx1
SxR
n
Sx1
1 b
R Sx1 S
- Exp (Wp
a = hardlims - + b)
Figure 3.1 Single-Layer Perceptron
3-3
Two-Input Case
Before we use the perceptron to solve the orange and apple recognition
problem (which will require a three-input perceptron, i.e., R = 3 ), it is use-
ful to investigate the capabilities of a two-input/single-neuron perceptron
( R = 2 ), which can be easily analyzed graphically. The two-input percep-
tron is shown in Figure 3.2.
Inputs Two-Input Neuron
p1 w1,1
n a
p2

w1,2 b
1
a = hardlims (Wp + b)
Figure 3.2 Two-Input/Single-Neuron Perceptron

Single-neuron perceptrons can classify input vectors into two categories.
For example, for a two-input perceptron, if w 1 1 = 1 and w 1 2 = 1 then
a = hardlims n = hardlims 1 1 p + b . (3.4)
Therefore, if the inner product of the weight matrix (a single row vector in
this case) with the input vector is greater than or equal to b , the output
will be 1. If the inner product of the weight vector and the input is less than
b , the output will be 1 . This divides the input space into two parts. Fig-
ure 3.3 illustrates this for the case where b = 1 . The blue line in the fig-
ure represents all points for which the net input n is equal to 0:
n = 1 1 p 1 = 0 . (3.5)
Notice that this decision boundary will always be orthogonal to the weight
matrix, and the position of the boundary can be shifted by changing b . (In
the general case, W is a matrix consisting of a number of row vectors, each
of which will be used in an equation like Eq. (3.5). There will be one bound-
ary for each row of W . See Chapter 4 for more on this topic.) The shaded
region contains all input vectors for which the output of the network will
be 1. The output will be 1 for all other input vectors.
3-4
Perceptron
p2
W
1
n>0 n<0
p1
-1 1
Figure 3.3 Perceptron Decision Boundary

The key property of the single-neuron perceptron, therefore, is that it can
separate input vectors into two categories. The decision boundary between
the categories is determined by the equation
Wp + b = 0 . (3.6)
Because the boundary must be linear, the single-layer perceptron can only
be used to recognize patterns that are linearly separable (can be separated
by a linear boundary). These concepts will be discussed in more detail in
Chapter 4.
Pattern Recognition Example

Now consider the apple and orange pattern recognition problem. Because
there are only two categories, we can use a single-neuron perceptron. The
vector inputs are three-dimensional ( R = 3 ), therefore the perceptron
equation will be
p1

a = hardlims w 1 1 w 1 2 w 1 3 p 2 + b . (3.7)

p3
We want to choose the bias b and the elements of the weight matrix so that
the perceptron will be able to distinguish between apples and oranges. For
example, we may want the output of the perceptron to be 1 when an apple
is input and 1 when an orange is input. Using the concept illustrated in
Figure 3.3, lets find a linear boundary that can separate oranges and ap-
3-5
ples. The two prototype vectors (recall Eq. (3.2) and Eq. (3.3)) are shown in
Figure 3.4. From this figure we can see that the linear boundary that di-
vides these two vectors symmetrically is the p 1 p 3 plane.
p3
p2
p1
p1 (orange) p2 (apple)
Figure 3.4 Prototype Vectors

The p 1 p 3 plane, which will be our decision boundary, can be described by
the equation
p2 = 0 , (3.8)
or
p1
0 1 0 p2 + 0 = 0 . (3.9)
p3
Therefore the weight matrix and bias will be
W = 0 1 0 , b = 0. (3.10)
The weight matrix is orthogonal to the decision boundary and points to-
ward the region that contains the prototype pattern p 2 (apple) for which we
want the perceptron to produce an output of 1. The bias is 0 because the
decision boundary passes through the origin.
Now lets test the operation of our perceptron pattern classifier. It classifies
perfect apples and oranges correctly since
3-6
Perceptron
Orange:
1

a = hardlims 0 1 0 1 + 0 = 1 orange , (3.11)

1
Apple:

3
1
a = hardlims 0 1 0 1 + 0 = 1 apple . (3.12)

1
But what happens if we put a not-so-perfect orange into the classifier? Lets
say that an orange with an elliptical shape is passed through the sensors.
The input vector would then be
1
p = 1 . (3.13)
1
The response of the network would be
1

a = hardlims 0 1 0 1 + 0 = 1 orange . (3.14)

1
In fact, any input vector that is closer to the orange prototype vector than
to the apple prototype vector (in Euclidean distance) will be classified as an
orange (and vice versa).
To experiment with the perceptron network and the apple/orange classifica-
tion problem, use the Neural Network Design Demonstration Perceptron
Classification (nnd3pc).
This example has demonstrated some of the features of the perceptron net-
work, but by no means have we exhausted our investigation of perceptrons.
This network, and variations on it, will be examined in Chapters 4 through
13. Lets consider some of these future topics.
In the apple/orange example we were able to design a network graphically,
by choosing a decision boundary that clearly separated the patterns. What
about practical problems, with high dimensional input spaces? In Chapters
4, 7, 10 and 11 we will introduce learning algorithms that can be used to
3-7
train networks to solve complex problems by using a set of examples of

proper network behavior.
The key characteristic of the single-layer perceptron is that it creates linear
decision boundaries to separate categories of input vector. What if we have
categories that cannot be separated by linear boundaries? This question
will be addressed in Chapter 11, where we will introduce the multilayer
perceptron. The multilayer networks are able to solve classification prob-
lems of arbitrary complexity.
Hamming Network
The next network we will consider is the Hamming network [Lipp87]. It
was designed explicitly to solve binary pattern recognition problems
(where each element of the input vector has only two possible values in
our example 1 or 1 ). This is an interesting network, because it uses both
feedforward and recurrent (feedback) layers, which were both described in
Chapter 2. Figure 3.5 shows the standard Hamming network. Note that
the number of neurons in the first layer is the same as the number of neu-
rons in the second layer.
The objective of the Hamming network is to decide which prototype vector
is closest to the input vector. This decision is indicated by the output of the
recurrent layer. There is one neuron in the recurrent layer for each proto-
type pattern. When the recurrent layer converges, there will be only one
neuron with nonzero output. This neuron indicates the prototype pattern
that is closest to the input vector. Now lets investigate the two layers of the
Hamming network in detail.
Feedforward Layer Recurrent Layer
p
Rx1
W1 n2(t + 1) a2(t + 1) a2(t)
n1 a1
SxR
Sx1 Sx1
W2 Sx1 Sx1 D Sx1
SxS
1 b 1
Sx1
R S S
- Exp 1(W
a1 = purelin - 1 p + b 1) a2(0) = a1 a2(t + 1) = poslin (W2a2(t))
Figure 3.5 Hamming Network
3-8
Hamming Network
Feedforward Layer
The feedforward layer performs a correlation, or inner product, between
each of the prototype patterns and the input pattern (as we will see in Eq.
(3.17)). In order for the feedforward layer to perform this correlation, the
rows of the weight matrix in the feedforward layer, represented by the con-
1
nection matrix W , are set to the prototype patterns. For our apple and or-
ange example this would mean
p 1T
W1 = = 1 1 1 . (3.15)
p 2T 1 1 1
The feedforward layer uses a linear transfer function, and each element of
the bias vector is equal to R , where R is the number of elements in the in-
put vector. For our example the bias vector would be
b1 = 3 . (3.16)
3
With these choices for the weight matrix and bias vector, the output of the
feedforward layer is
T T
p1 p1 p + 3
p+ 3 =
1 1 1
a = W p+b = . (3.17)
T 3 T
p2 p2 p + 3
Note that the outputs of the feedforward layer are equal to the inner prod-
ucts of each prototype pattern with the input, plus R . For two vectors of
equal length (norm), their inner product will be largest when the vectors
point in the same direction, and will be smallest when they point in oppo-
site directions. (We will discuss this concept in more depth in Chapters 5,
8 and 9.) By adding R to the inner product we guarantee that the outputs
of the feedforward layer can never be negative. This is required for proper
operation of the recurrent layer.
This network is called the Hamming network because the neuron in the
feedforward layer with the largest output will correspond to the prototype
pattern that is closest in Hamming distance to the input pattern. (The
Hamming distance between two vectors is equal to the number of elements
that are different. It is defined only for binary vectors.) We leave it to the
reader to show that the outputs of the feedforward layer are equal to 2R
minus twice the Hamming distances from the prototype patterns to the in-
put pattern.
3-9
Recurrent Layer
The recurrent layer of the Hamming network is what is known as a com-
petitive layer. The neurons in this layer are initialized with the outputs of
the feedforward layer, which indicate the correlation between the proto-
type patterns and the input vector. Then the neurons compete with each
other to determine a winner. After the competition, only one neuron will
have a nonzero output. The winning neuron indicates which category of in-
put was presented to the network (for our example the two categories are
apples and oranges). The equations that describe the competition are:
1
a2 0 = a (Initial Condition), (3.18)
and
a 2 t + 1 = poslin W 2 a 2 t . (3.19)
(Dont forget that the superscripts here indicate the layer number, not a
power of 2.) The poslin transfer function is linear for positive values and
zero for negative values. The weight matrix W2 has the form
W2 = 1 , (3.20)
1
where is some number less than 1 S 1 , and S is the number of neu-

rons in the recurrent layer. (Can you show why must be less than
1 S 1 ?)
An iteration of the recurrent layer proceeds as follows:
2 2
a t a 2 t
a 2 t + 1 = poslin 1 a 2 t = poslin 1 . (3.21)
1 2 2
a 2 t a 1 t
Each element is reduced by the same fraction of the other. The larger ele-
ment will be reduced by less, and the smaller element will be reduced by
more, therefore the difference between large and small will be increased.
The effect of the recurrent layer is to zero out all neuron outputs, except the
one with the largest initial value (which corresponds to the prototype pat-
tern that is closest in Hamming distance to the input).
To illustrate the operation of the Hamming network, consider again the ob-
long orange that we used to test the perceptron:
3-10
Hamming Network
1
p = 1 . (3.22)
1
The output of the feedforward layer will be
1
a 1 = 1 1 1 1 + 3 = 1 + 3 = 4 , (3.23)
1 1 1 3 1 + 3 2
1
which will then become the initial condition for the recurrent layer.
The weight matrix for the recurrent layer will be given by Eq. (3.20) with
= 1 2 (any number less than 1 would work). The first iteration of the re-
current layer produces

poslin 1 0.5 4

0.5 1 2
a 2 1 = poslin W 2 a 2 0 = . (3.24)

poslin 3 = 3
0 0

The second iteration produces

poslin 1 0.5 3

0.5 1 0
a 2 2 = poslin W 2 a 2 1 = . (3.25)

poslin 3 = 3
1.5 0

Since the outputs of successive iterations produce the same result, the net-
work has converged. Prototype pattern number one, the orange, is chosen
as the correct match, since neuron number one has the only nonzero out-
1 T
put. (Recall that the first element of a was p 1 p + 3 .) This is the correct
choice, since the Hamming distance from the orange prototype to the input
pattern is 1, and the Hamming distance from the apple prototype to the in-
put pattern is 2.
To experiment with the Hamming network and the apple/orange classifica-
tion problem, use the Neural Network Design Demonstration Hamming
Classification (nnd3hamc).
3-11
There are a number of networks whose operation is based on the same prin-
ciples as the Hamming network; that is, where an inner product operation
(feedforward layer) is followed by a competitive dynamic layer. These com-
petitive networks will be discussed in Chapters 15, 16, 18 and 19. They are
self-organizing networks, which can learn to adjust their prototype vectors
based on the inputs that have been presented.
Hopfield Network
The final network we will discuss in this brief preview is the Hopfield net-
work. This is a recurrent network that is similar in some respects to the re-
current layer of the Hamming network, but which can effectively perform
the operations of both layers of the Hamming network. A diagram of the
Hopfield network is shown in Figure 3.6. (This figure is actually a slight
variation of the standard Hopfield network. We use this variation because
it is somewhat simpler to describe and yet demonstrates the basic con-
cepts.)
The neurons in this network are initialized with the input vector, then the
network iterates until the output converges. When the network is operat-
ing correctly, the resulting output should be one of the prototype vectors.
Therefore, whereas in the Hamming network the nonzero neuron indicates
which prototype pattern is chosen, the Hopfield network actually produces
the selected prototype pattern at its output.
Initial
p
Sx1 W n(t + 1) a(t + 1) a(t)
SxS
Sx1 Sx1 D Sx1
1 b
Sx1
S S
Figure 3.6 Hopfield Network

The equations that describe the network operation are
a0 = p (3.26)
and
3-12
Hopfield Network
a t + 1 = satlins Wa t + b , (3.27)
where satlins is the transfer function that is linear in the range [-1, 1] and
saturates at 1 for inputs greater than 1 and at -1 for inputs less than -1.
The design of the weight matrix and the bias vector for the Hopfield net-
work is a more complex procedure than it is for the Hamming network,
where the weights in the feedforward layer are the prototype patterns.
Hopfield design procedures will be discussed in detail in Chapter 21.
To illustrate the operation of the network, we have determined a weight
matrix and a bias vector that can solve our orange and apple pattern rec-
ognition problem. They are given in Eq. (3.28).
0.2 0 0 0.9
W = 0 1.2 0 b = 0 (3.28)
0 0 0.2 0.9
Although the procedure for computing the weights and biases for the
Hopfield network is beyond the scope of this chapter, we can say a few
things about why the parameters in Eq. (3.28) work for the apple and or-
ange example.
We want the network output to converge to either the orange pattern, p 1 ,
or the apple pattern, p 2 . In both patterns, the first element is 1 , and the
third element is 1 . The difference between the patterns occurs in the sec-
ond element. Therefore, no matter what pattern is input to the network, we
want the first element of the output pattern to converge to 1 , the third el-
ement to converge to 1 , and the second element to go to either 1 or 1 ,
whichever is closer to the second element of the input vector.
The equations of operation of the Hopfield network, using the parameters

given in Eq. (3.28), are
a 1 t + 1 = satlins 0.2a 1 t + 0.9
a 2 t + 1 = satlins 1.2a 2 t
a 3 t + 1 = satlins 0.2a 3 t 0.9

(3.29)
Regardless of the initial values, a i 0 , the first element will be increased
until it saturates at 1 , and the third element will be decreased until it sat-
urates at 1 . The second element is multiplied by a number larger than 1 .
Therefore, if it is initially negative, it will eventually saturate at 1 ; if it is
initially positive it will saturate at 1 .
3-13
(It should be noted that this is not the only (W, b) pair that could be used.
You might want to try some others. See if you can discover what makes
these work.)
Lets again take our oblong orange to test the Hopfield network. The out-
puts of the Hopfield network for the first three iterations would be
1 0.7 1 1
a 0 = 1 , a 1 = 1 , a 2 = 1 , a 3 = 1 (3.30)
1 1 1 1
The network has converged to the orange pattern, as did both the Hamming
network and the perceptron, although each network operated in a different
way. The perceptron had a single output, which could take on values of -1
(orange) or 1 (apple). In the Hamming network the single nonzero neuron in-
dicated which prototype pattern had the closest match. If the first neuron
was nonzero, that indicated orange, and if the second neuron was nonzero,
that indicated apple. In the Hopfield network the prototype pattern itself
appears at the output of the network.
To experiment with the Hopfield network and the apple/orange classifica-
tion problem, use the Neural Network Design Demonstration Hopfield Clas-
sification (nnd3hopc).
As with the other networks demonstrated in this chapter, do not expect to
feel completely comfortable with the Hopfield network at this point. There
are a number of questions that we have not discussed. For example, How
do we know that the network will eventually converge? It is possible for
recurrent networks to oscillate or to have chaotic behavior. In addition, we
have not discussed general procedures for designing the weight matrix and
the bias vector. These topics will be discussed in detail in Chapters 20 and
21.
3-14
Epilogue
Epilogue
The three networks that we have introduced in this chapter demonstrate
many of the characteristics that are found in the architectures which are
discussed throughout this book.
Feedforward networks, of which the perceptron is one example, are pre-
sented in Chapters 4, 7, 11, 12, 13 and 17. In these networks, the output is
computed directly from the input in one pass; no feedback is involved.
Feedforward networks are used for pattern recognition, as in the apple and
3
orange example, and also for function approximation (see Chapter 11).
Function approximation applications are found in such areas as adaptive
filtering (see Chapter 10) and automatic control.
Competitive networks, represented here by the Hamming network, are
characterized by two properties. First, they compute some measure of dis-
tance between stored prototype patterns and the input pattern. Second,
they perform a competition to determine which neuron represents the pro-
totype pattern closest to the input. In the competitive networks that are
discussed in Chapters 16, 18 and 19, the prototype patterns are adjusted
as new inputs are applied to the network. These adaptive networks learn
to cluster the inputs into different categories.
Recurrent networks, like the Hopfield network, were originally inspired by
statistical mechanics. They have been used as associative memories, in
which stored data is recalled by association with input data, rather than by
an address. They have also been used to solve a variety of optimization
problems. We will discuss these recurrent networks in Chapters 20 and 21.
We hope this chapter has piqued your curiosity about the capabilities of
neural networks and has raised some questions. A few of the questions we
will answer in later chapters are:
1. How do we determine the weight matrix and bias for perceptron net-
works with many inputs, where it is impossible to visualize the decision
boundary? (Chapters 4 and 10)
2. If the categories to be recognized are not linearly separable, can we ex-
tend the standard perceptron to solve the problem? (Chapters 11, 12
and 13)
3. Can we learn the weights and biases of the Hamming network when we
dont know the prototype patterns? (Chapters 16, 18 and 19)
4. How do we determine the weight matrix and bias vector for the
Hopfield network? (Chapter 21)
5. How do we know that the Hopfield network will eventually converge?
(Chapters 20 and 21)
3-15
Exercises
E3.1 In this chapter we have designed three different neural networks to distin-
guish between apples and oranges, based on three sensor measurements
(shape, texture and weight). Suppose that we want to distinguish between
bananas and pineapples:
1
p1 = 1 (Banana)
1
1
p 2 = 1 (Pineapple)
1
i. Design a perceptron to recognize these patterns.

ii. Design a Hamming network to recognize these patterns.
iii. Design a Hopfield network to recognize these patterns.
iv. Test the operation of your networks by applying several different in-
put patterns. Discuss the advantages and disadvantages of each
network.
E3.2 Consider the following prototype patterns.
p1 = 1 ,p = 2
2
0.5 1
i. Find and sketch a decision boundary for a perceptron network that

will recognize these two vectors.
ii. Find weights and bias which will produce the decision boundary you
found in part i, and sketch the network diagram.
iii. Calculate the network output for the following input. Is the network
response (decision) reasonable? Explain.
p = 1
0
iv. Design a Hamming network to recognize the two prototype vectors

above.
3-16
Exercises
v. Calculate the network output for the Hamming network with the in-
put vector given in part iii, showing all steps. Does the Hamming
network produce the same decision as the perceptron? Explain why
or why not. Which network is better suited to this problem? Explain.
E3.3 Consider a Hopfield network, with the following weight and bias.
1 1 , b = 0
W =
1 1 0 3
i. The following input (initial condition) is applied to the network.
Find the network response (show the network output at each itera-
tion until the network converges).
p = 0.9
1
ii. Draw a sketch indicating what region of the input space will con-
verge to the same final output that you found in part i. (In other
words, for what other p vectors will the network converge to the
same final output?) Explain how you obtained your answer.
iii. What other prototypes will this network converge to, and what re-
gions of the input space correspond to each prototype (sketch the re-
gions). Explain how you obtained your answer.
E3.4 Consider the following perceptron network.

- Title - Layer
p a W = 1 1
Rx1
W Sx1 1 1
SxR
n
Sx1
b = 2
1 b 0
R Sx1 S
- Exp (Wp
a = hardlims - + b)
i. How many different classes can this network classify?

ii. Draw a diagram illustrating the regions corresponding to each
class. Label each region with the corresponding network output.
iii. Calculate the network output for the following input.
3-17
p = 1
1
iv. Plot the input from part iii in your diagram from part ii, and verify
that it falls in the correctly labeled region.
E3.5 We want to design a perceptron network to output a 1 when either of these

two vectors are input to the network:
1 1
,
0 2
and to output a -1 when either of the following vectors are input to the net-
work:
1 0
.
1 2
i. Find and sketch a decision boundary for a network that will solve
this problem.
ii. Find weights and biases that will produce the decision boundary
you found in part i. Show all work.
iii. Draw the network diagram using abreviated notation.
iv. For each of the four vectors given above, calculate the net input, n,
and the network output, a, for the network you have designed. Ver-
ify that your network solves the problem.
v. Are there other weights and biases that would solve the problem? If
so, would you consider your weights best? Explain.
E3.6 We have the folowing two prototype vectors:

. 1 1
1 1
i. Find and sketch a decision boundary for a perceptron network that

will recognize these two vectors.
ii. Find weights and bias that will produce the decision boundary you
found in part i.
iii. Draw the network diagram using abreviated notation.
3-18
Exercises
iv. For the vector given below, calculate the net input, n, and the net-
work output, a, for the network you have designed. Does the net-
work produce a good output? Explain.
0.5
0.5
v. Design a Hamming network to recognize the two vectors used in

part i. 3
vi. Calculate the network output for the Hamming network for the in-
put vector given in part iv. Does the network produce a good output?
Explain.
vii. Design a Hopfield network to recognize the two vectors used in part
i.
viii. Calculate the network output for the Hopfield network for the input
vector given in part iv. Does the network produce a good output? Ex-
plain.
E3.7 We want to design a Hamming network to recognize the following proto-

type vectors:
1 1 1
.
1 1 1
i. Find the weight matrices and bias vectors for the Hamming net-
work.
ii. Draw the network diagram.
iii. Apply the following input vector and calculate the total network re-
sponse (iterating the second layer to convergence). Explain the
meaning of the final network output.
p = 1
0
iv. Sketch the decision boundaries for this network. Explain how you
determined the boundaries.
3-19
Objectives

Objectives 4-1
Learning Rules 4-2
Perceptron Architecture 4-3
Single-Neuron Perceptron 4-5
Multiple-Neuron Perceptron 4-8
Perceptron Learning Rule 4-8
Test Problem 4-9
Constructing Learning Rules 4-10
Unified Learning Rule 4-12
Training Multiple-Neuron Perceptrons 4-13
Proof of Convergence 4-15
Notation 4-15
Proof 4-16
Limitations 4-18
Epilogue 4-34
Exercises 4-37
Objectives
One of the questions we raised in Chapter 3 was: How do we determine the
weight matrix and bias for perceptron networks with many inputs, where
it is impossible to visualize the decision boundaries? In this chapter we
will describe an algorithm for training perceptron networks, so that they
can learn to solve classification problems. We will begin by explaining what
a learning rule is and will then develop the perceptron learning rule. We
will conclude by discussing the advantages and limitations of the single-
layer perceptron network. This discussion will lead us into future chapters.
4-1
Theory and Examples

In 1943, Warren McCulloch and Walter Pitts introduced one of the first ar-
tificial neurons [McPi43]. The main feature of their neuron model is that a
weighted sum of input signals is compared to a threshold to determine the
neuron output. When the sum is greater than or equal to the threshold, the
output is 1. When the sum is less than the threshold, the output is 0. They
went on to show that networks of these neurons could, in principle, com-
pute any arithmetic or logical function. Unlike biological networks, the pa-
rameters of their networks had to be designed, as no training method was
available. However, the perceived connection between biology and digital
computers generated a great deal of interest.
In the late 1950s, Frank Rosenblatt and several other researchers devel-
oped a class of neural networks called perceptrons. The neurons in these
networks were similar to those of McCulloch and Pitts. Rosenblatts key
contribution was the introduction of a learning rule for training perceptron
networks to solve pattern recognition problems [Rose58]. He proved that
his learning rule will always converge to the correct network weights, if
weights exist that solve the problem. Learning was simple and automatic.
Examples of proper behavior were presented to the network, which learned
from its mistakes. The perceptron could even learn when initialized with
random values for its weights and biases.
Unfortunately, the perceptron network is inherently limited. These limita-
tions were widely publicized in the book Perceptrons [MiPa69] by Marvin
Minsky and Seymour Papert. They demonstrated that the perceptron net-
works were incapable of implementing certain elementary functions. It
was not until the 1980s that these limitations were overcome with im-
proved (multilayer) perceptron networks and associated learning rules. We
will discuss these improvements in Chapters 11 and 12.
Today the perceptron is still viewed as an important network. It remains a
fast and reliable network for the class of problems that it can solve. In ad-
dition, an understanding of the operations of the perceptron provides a
good basis for understanding more complex networks. Thus, the perceptron
network, and its associated learning rule, are well worth discussing here.
In the remainder of this chapter we will define what we mean by a learning
rule, explain the perceptron network and learning rule, and discuss the
limitations of the perceptron network.
Learning Rules
As we begin our presentation of the perceptron learning rule, we want to
Learning Rule discuss learning rules in general. By learning rule we mean a procedure for
modifying the weights and biases of a network. (This procedure may also
4-2
Perceptron Architecture
be referred to as a training algorithm.) The purpose of the learning rule is

to train the network to perform some task. There are many types of neural
network learning rules. They fall into three broad categories: supervised
learning, unsupervised learning and reinforcement (or graded) learning.
Supervised Learning In supervised learning, the learning rule is provided with a set of examples
Training Set (the training set) of proper network behavior:
{p1, t 1} {p2, t 2} {pQ, tQ} , (4.1)
where p q is an input to the network and t q is the corresponding correct

Target (target) output. As the inputs are applied to the network, the network out-
puts are compared to the targets. The learning rule is then used to adjust
the weights and biases of the network in order to move the network outputs
closer to the targets. The perceptron learning rule falls in this supervised
learning category. We will also investigate supervised learning algorithms
in Chapters 714.
Reinforcement Learning Reinforcement learning is similar to supervised learning, except that, in-
stead of being provided with the correct output for each network input, the
algorithm is only given a grade. The grade (or score) is a measure of the net-
work performance over some sequence of inputs. This type of learning is
currently much less common than supervised learning. It appears to be
most suited to control system applications (see [BaSu83], [WhSo92]).
Unsupervised Learning In unsupervised learning, the weights and biases are modified in response
to network inputs only. There are no target outputs available. At first
glance this might seem to be impractical. How can you train a network if
you dont know what it is supposed to do? Most of these algorithms perform
some kind of clustering operation. They learn to categorize the input pat-
terns into a finite number of classes. This is especially useful in such appli-
cations as vector quantization. We will see in Chapters 1519 that there
are a number of unsupervised learning algorithms.
Before we present the perceptron learning rule, lets expand our investiga-
tion of the perceptron network, which we began in Chapter 3. The general
perceptron network is shown in Figure 4.1.
The output of the network is given by
a = hardlim Wp + b . (4.2)
(Note that in Chapter 3 we used the hardlims transfer function, instead of

hardlim. This does not affect the capabilities of the network. See Exercise
E4.10.)
4-3
Input Hard Limit Layer
p a
Rx1
W Sx1
SxR
n
Sx1
1 b
Sx1
R S
a = hardlim (Wp + b)
Figure 4.1 Perceptron Network

It will be useful in our development of the perceptron learning rule to be
able to conveniently reference individual elements of the network output.
Lets see how this can be done. First, consider the network weight matrix:
w 1 1 w 1 2 w 1 R
w 2 1 w 2 2 w 2 R
W = . (4.3)

w S 1 w S 2 w S R
We will define a vector composed of the elements of the ith row of W :
w i 1
w i 2
iw = . (4.4)

w i R
Now we can partition the weight matrix:
T
1w
T
W = 2w . (4.5)

T
Sw
This allows us to write the ith element of the network output vector as
T
a i = hardlim n i = hardlim iw p + b i . (4.6)
4-4
Recall that the hardlim transfer function (shown at left) is defined as:
a = hardlim (n)

a = hardlim n = 1 if n 0 (4.7)
0 otherwise.
Therefore, if the inner product of the ith row of the weight matrix with the
n = Wp + b input vector is greater than or equal to b i , the output will be 1, otherwise
the output will be 0. Thus each neuron in the network divides the input
space into two regions. It is useful to investigate the boundaries between
these regions. We will begin with the simple case of a single-neuron percep-
tron with two inputs.
Single-Neuron Perceptron
Lets consider a two-input perceptron with one neuron, as shown in Figure
4.2.
p1 w1,1
n a
p2

w1,2 b
1
Figure 4.2 Two-Input/Single-Output Perceptron

The output of this network is determined by
a = hardlim n = hardlim Wp + b
T
= hardlim 1w p + b = hardlim w 1 1 p 1 + w 1 2 p 2 + b
(4.8)
Decision Boundary The decision boundary is determined by the input vectors for which the net
input n is zero:
T
n = 1w p + b = w 1 1 p 1 + w 1 2 p 2 + b = 0 . (4.9)
To make the example more concrete, lets assign the following values for
the weights and bias:
w 1 1 = 1 , w 1 2 = 1 , b = 1 . (4.10)
4-5
The decision boundary is then

T
n = 1w p + b = w 1 1 p 1 + w 1 2 p 2 + b = p 1 + p 2 1 = 0 . (4.11)
This defines a line in the input space. On one side of the line the network
output will be 0; on the line and on the other side of the line the output will
be 1. To draw the line, we can find the points where it intersects the p 1 and
p 2 axes. To find the p 2 intercept set p 1 = 0 :
b 1
p 2 = ---------- = ------ = 1 if p 1 = 0
w 1 2 1
. (4.12)
To find the p 1 intercept, set p 2 = 0 :
b 1
p 1 = ---------- = ------ = 1 if p 2 = 0
w 1 1 1
. (4.13)
The resulting decision boundary is illustrated in Figure 4.3.
To find out which side of the boundary corresponds to an output of 1, we
T
just need to test one point. For the input p = 2 0 , the network output
will be

a = hardlim 1w p + b = hardlim 1 1 2 1 = 1 .
T
(4.14)
0
Therefore, the network output will be 1 for the region above and to the right
of the decision boundary. This region is indicated by the shaded area in Fig-
ure 4.3.
p2
a=1
w
1w Tp + b = 0 1 1
p1
a=0 1
Figure 4.3 Decision Boundary for Two-Input Perceptron
4-6
We can also find the decision boundary graphically. The first step is to note
1 w that the boundary is always orthogonal to 1w , as illustrated in the adjacent
figures. The boundary is defined by
T
1w p+b = 0. (4.15)
w
1 For all points on the boundary, the inner product of the input vector with
the weight vector is the same. This implies that these input vectors will all
have the same projection onto the weight vector, so they must lie on a line
orthogonal to the weight vector. (These concepts will be covered in more de-
tail in Chapter 5.) In addition, any vector in the shaded region of Figure 4.3
will have an inner product greater than b , and vectors in the unshaded
region will have inner products less than b . Therefore the weight vector
1w will always point toward the region where the neuron output is 1.
After we have selected a weight vector with the correct angular orientation,
the bias value can be computed by selecting a point on the boundary and
satisfying Eq. (4.15).
2 Lets apply some of these concepts to the design of a perceptron network to
+2
implement a simple logic function: the AND gate. The input/target pairs for
the AND gate are
0 t = 0 p = 0 t = 0 p = 1 t = 0 p = 1 t = 1 .
p1 = 1 2 2 3 3 4 4
0 1 0 1
The figure to the left illustrates the problem graphically. It displays the in-
put space, with each input vector labeled according to its target. The dark
circles indicate that the target is 1, and the light circles indicate that
the target is 0.
The first step of the design is to select a decision boundary. We want to
have a line that separates the dark circles and the light circles. There are
an infinite number of solutions to this problem. It seems reasonable to
choose the line that falls halfway between the two categories of inputs, as
shown in the adjacent figure.
Next we want to choose a weight vector that is orthogonal to the decision
boundary. The weight vector can be any length, so there are infinite possi-
bilities. One choice is
1w
1w = 2 , (4.16)
AND 2
as displayed in the figure to the left.
4-7
Finally, we need to find the bias, b . We can do this by picking a point on

T
the decision boundary and satisfying Eq. (4.15). If we use p = 1.5 0 we
find
p + b = 2 2 1.5 + b = 3 + b = 0
T
1w b = 3 . (4.17)
0
We can now test the network on one of the input/target pairs. If we apply
p 2 to the network, the output will be

a = hardlim 1w p 2 + b = hardlim 2 2 0 3
T
1
a = hardlim 1 = 0, (4.18)
which is equal to the target output t 2 . Verify for yourself that all inputs are
correctly classified.
To experiment with decision boundaries, use the Neural Network Design
Demonstration Decision Boundaries (nnd4db).
Multiple-Neuron Perceptron
Note that for perceptrons with multiple neurons, as in Figure 4.1, there
will be one decision boundary for each neuron. The decision boundary for
neuron i will be defined by
T
iw p + bi = 0 . (4.19)
A single-neuron perceptron can classify input vectors into two categories,

since its output can be either 0 or 1. A multiple-neuron perceptron can clas-
sify inputs into many categories. Each category is represented by a differ-
ent output vector. Since each element of the output vector can be either 0
S
or 1, there are a total of 2 possible categories, where S is the number of
neurons.
Perceptron Learning Rule

Now that we have examined the performance of perceptron networks, we
are in a position to introduce the perceptron learning rule. This learning
rule is an example of supervised training, in which the learning rule is pro-
vided with a set of examples of proper network behavior:
{p 1, t 1} { p 2, t 2} {p Q, tQ} , (4.20)
4-8
where p q is an input to the network and t q is the corresponding target out-

put. As each input is applied to the network, the network output is com-
pared to the target. The learning rule then adjusts the weights and biases
of the network in order to move the network output closer to the target.
Test Problem
In our presentation of the perceptron learning rule we will begin with a
simple test problem and will experiment with possible rules to develop
some intuition about how the rule should work. The input/target pairs for
our test problem are
1 t = 1 p = 1 t = 0 p = 0 t = 0 .
p1 = 1 2 2 3 3
2 2 1
The problem is displayed graphically in the adjacent figure, where the two
input vectors whose target is 0 are represented with a light circle , and
1
2 the vector whose target is 1 is represented with a dark circle . This is a
very simple problem, and we could almost obtain a solution by inspection.
This simplicity will help us gain some intuitive understanding of the basic
concepts of the perceptron learning rule.
3 The network for this problem should have two-inputs and one output. To
simplify our development of the learning rule, we will begin with a network
without a bias. The network will then have just two parameters, w 1 1 and
w 1 2 , as shown in Figure 4.4.
Inputs No-Bias Neuron
p1 w1,1
n a
p2

w1,2
a = hardlim(Wp)
Figure 4.4 Test Problem Network

1 By removing the bias we are left with a network whose decision boundary
2
must pass through the origin. We need to be sure that this network is still
able to solve the test problem. There must be an allowable decision bound-
ary that can separate the vectors p 2 and p 3 from the vector p 1 . The figure
to the left illustrates that there are indeed an infinite number of such
3
boundaries.
4-9
The adjacent figure shows the weight vectors that correspond to the allow-
able decision boundaries. (Recall that the weight vector is orthogonal to the
1
decision boundary.) We would like a learning rule that will find a weight
2 vector that points in one of these directions. Remember that the length of
the weight vector does not matter; only its direction is important.
Constructing Learning Rules

3
Training begins by assigning some initial values for the network parame-
ters. In this case we are training a two-input/single-output network with-
out a bias, so we only have to initialize its two weights. Here we set the
elements of the weight vector, 1w , to the following randomly generated val-
ues:
T
1w = 1.0 0.8 . (4.21)
We will now begin presenting the input vectors to the network. We begin
with p 1 :

a = hardlim 1w p 1 = hardlim 1.0 0.8 1
T
2
a = hardlim 0.6 = 0 . (4.22)
The network has not returned the correct value. The network output is 0,
while the target response, t 1 , is 1.
1
2 We can see what happened by looking at the adjacent diagram. The initial
weight vector results in a decision boundary that incorrectly classifies the
vector p 1 . We need to alter the weight vector so that it points more toward
p 1 , so that in the future it has a better chance of classifying it correctly.
1w
3
One approach would be to set 1w equal to p 1 . This is simple and would en-
sure that p 1 was classified properly in the future. Unfortunately, it is easy
to construct a problem for which this rule cannot find a solution. The dia-
gram to the lower left shows a problem that cannot be solved with the
weight vector pointing directly at either of the two class 1 vectors. If we ap-
ply the rule 1w = p every time one of these vectors is misclassified, the net-
works weights will simply oscillate back and forth and will never find a
solution.
Another possibility would be to add p 1 to 1w . Adding p 1 to 1w would make
1w point more in the direction of p 1 . Repeated presentations of p 1 would
cause the direction of 1w to asymptotically approach the direction of p 1 .
This rule can be stated:
new old
If t = 1 and a = 0, then 1w = 1w +p. (4.23)
4-10
Applying this rule to our test problem results in new values for 1w :
new
= 1w
old
+ p1 = 1.0 + 1 = 2.0 . (4.24)
1w
0.8 2 1.2
1
2 w This operation is illustrated in the adjacent figure.
1
We now move on to the next input vector and will continue making changes
to the weights and cycling through the inputs until they are all classified
correctly.
3
The next input vector is p 2 . When it is presented to the network we find:

T
2
= hardlim 0.4 = 1 . (4.25)
The target t 2 associated with p 2 is 0 and the output a is 1. A class 0 vector
was misclassified as a 1.
Since we would now like to move the weight vector 1w away from the input,
we can simply change the addition in Eq. (4.23) to subtraction:
new old
If t = 0 and a = 1, then 1w = 1w p. (4.26)
If we apply this to the test problem we find:
p 2 = 2.0 1 = 3.0 ,
new old
1w = 1w (4.27)
1 1.2 2 0.8
2
which is illustrated in the adjacent figure.
w Now we present the third vector p 3 :

1
3

T
1
= hardlim 0.8 = 1 . (4.28)
The current 1w results in a decision boundary that misclassifies p 3 . This
is a situation for which we already have a rule, so 1w will be updated again,
according to Eq. (4.26):
new
= 1w
old
p3 = 3.0 0 = 3.0 . (4.29)
1w
0.8 1 0.2
4-11
The diagram to the left shows that the perceptron has finally learned to
classify the three vectors properly. If we present any of the input vectors to
1
the neuron, it will output the correct class for that input vector.
2
This brings us to our third and final rule: if it works, dont fix it.
1w
new old
If t = a, then 1w = 1w . (4.30)
3
Here are the three rules, which cover all possible combinations of output
and target values:
new old
If t = 1 and a = 0, then 1w = 1w + p.
If t = 0 and a = 1, then 1w
new
= 1w
old
p. (4.31)
new old
If t = a, then 1w = 1w .
Unified Learning Rule

The three rules in Eq. (4.31) can be rewritten as a single expression. First
we will define a new variable, the perceptron error e:
e = t a. (4.32)
We can now rewrite the three rules of Eq. (4.31) as:
new old
If e = 1, then 1w = 1w +p.
If e = 1, then 1w
new
= 1w
old
p. (4.33)
new old
If e = 0, then 1w = 1w .
Looking carefully at the first two rules in Eq. (4.33) we can see that the sign
of p is the same as the sign on the error, e. Furthermore, the absence of p
in the third rule corresponds to an e of 0. Thus, we can unify the three rules
into a single expression:
new old old
1w = 1w + ep = 1w + t a p . (4.34)
This rule can be extended to train the bias by noting that a bias is simply
a weight whose input is always 1. We can thus replace the input p in Eq.
(4.34) with the input to the bias, which is 1. The result is the perceptron
rule for a bias:
new old
b = b +e. (4.35)
4-12
Training Multiple-Neuron Perceptrons

The perceptron rule, as given by Eq. (4.34) and Eq. (4.35), updates the
weight vector of a single neuron perceptron. We can generalize this rule for
the multiple-neuron perceptron of Figure 4.1 as follows. To update the ith
row of the weight matrix use:
new old
iw = iw + ei p . (4.36)
To update the ith element of the bias vector use:

new old
bi = bi + ei . (4.37)
Perceptron Rule The perceptron rule can be written conveniently in matrix notation:
new old T
W = W + ep , (4.38)
and
new old
b = b +e. (4.39)
2 To test the perceptron learning rule, consider again the apple/orange rec-
+2
ognition problem of Chapter 3. The input/output prototype vectors will be
1 1

p1 = 1 t1 = 0 p2 = 1 t2 = 1 . (4.40)

1 1
(Note that we are using 0 as the target output for the orange pattern, p 1 ,
instead of -1, as was used in Chapter 3. This is because we are using the
hardlim transfer function, instead of hardlims .)
Typically the weights and biases are initialized to small random numbers.
Suppose that here we start with the initial weight matrix and bias:
W = 0.5 1 0.5 , b = 0.5 . (4.41)
The first step is to apply the first input vector, p 1 , to the network:
1

a = hardlim Wp 1 + b = hardlim 0.5 1 0.5 1 + 0.5

1
= hardlim 2.5 = 1 (4.42)
4-13
Then we calculate the error:
e = t1 a = 0 1 = 1 . (4.43)
The weight update is

new old T
W = W + ep = 0.5 1 0.5 + 1 1 1 1 (4.44)
= 0.5 0 0.5 .
The bias update is

new old
b = b + e = 0.5 + 1 = 0.5 . (4.45)
This completes the first iteration.

The second iteration of the perceptron rule is:
1
a = hardlim (Wp 2 + b) = hardlim ( 0.5 0 0.5 1 + 0.5 )
1
= hardlim ( 0.5) = 0 (4.46)
e = t2 a = 1 0 = 1 (4.47)
new old T
W = W + ep = 0.5 0 0.5 + 1 1 1 1 = 0.5 1 0.5 (4.48)
new old
b = b + e = 0.5 + 1 = 0.5 (4.49)
The third iteration begins again with the first input vector:
1
a = hardlim (Wp 1 + b) = hardlim ( 0.5 1 0.5 1 + 0.5)
1
= hardlim (0.5) = 1 (4.50)
e = t1 a = 0 1 = 1 (4.51)
new old T
W = W + ep = 0.5 1 0.5 + 1 1 1 1 (4.52)
= 0.5 2 0.5
4-14
Proof of Convergence
new old
b = b + e = 0.5 + 1 = 0.5 . (4.53)
If you continue with the iterations you will find that both input vectors will
now be correctly classified. The algorithm has converged to a solution. Note
that the final decision boundary is not the same as the one we developed in
Chapter 3, although both boundaries correctly classify the two input vec-
tors.
To experiment with the perceptron learning rule, use the Neural Network
Design Demonstration Perceptron Rule (nnd4pr).
Although the perceptron learning rule is simple, it is quite powerful. In
fact, it can be shown that the rule will always converge to weights that ac-
complish the desired classification (assuming that such weights exist). In
this section we will present a proof of convergence for the perceptron learn-
ing rule for the single-neuron perceptron shown in Figure 4.5.
Inputs Hard Limit Neuron
p1 w1,1
p2
p3 n a

w1,R b
pR
1
a = hardlim (1wTp + b)
Figure 4.5 Single-Neuron Perceptron

The output of this perceptron is obtained from
T
a = hardlim 1w p + b . (4.54)
The network is provided with the following examples of proper network be-
havior:
{p 1, t 1} {p 2, t 2} {p Q, t Q} . (4.55)
where each target output, t q , is either 0 or 1 .
Notation
To conveniently present the proof we will first introduce some new nota-
tion. We will combine the weight matrix and the bias into a single vector:
4-15
x = 1w . (4.56)
b
We will also augment the input vectors with a 1, corresponding to the bias
input:
pq
zq = . (4.57)
1
Now we can express the net input to the neuron as follows:

T T
n = 1w p + b = x z . (4.58)
The perceptron learning rule for a single-neuron perceptron (Eq. (4.34) and
Eq. (4.35)) can now be written
new old
x = x + ez . (4.59)
The error e can be either 1 , 1 or 0 . If e = 0 , then no change is made to

the weights. If e = 1 , then the input vector is added to the weight vector.
If e = 1 , then the negative of the input vector is added to the weight vec-
tor. If we count only those iterations for which the weight vector is changed,
the learning rule becomes
x k = x k 1 + z' k 1 , (4.60)
where z' k 1 is the appropriate member of the set
z 1 z 2 z Q z 1 z 2 z Q . (4.61)
We will assume that a weight vector exists that can correctly categorize all
Q input vectors. This solution will be denoted x . For this weight vector
we will assume that
x z q 0 if t q = 1 ,
T
(4.62)
and
x z q 0 if t q = 0 .
T
(4.63)
Proof
We are now ready to begin the proof of the perceptron convergence theo-
rem. The objective of the proof is to find upper and lower bounds on the
length of the weight vector at each stage of the algorithm.
4-16
Assume that the algorithm is initialized with the zero weight vector:
x 0 = 0 . (This does not affect the generality of our argument.) Then, after
k iterations (changes to the weight vector), we find from Eq. (4.60):
x k = z' 0 + z' 1 + + z' k 1 . (4.64)
If we take the inner product of the solution weight vector with the weight
vector at iteration k we obtain
x x k = x z' 0 + x z' 1 + + x z' k 1 .

T T T T
(4.65)
From Eq. (4.61)Eq. (4.63) we can show that
x z' i .
T
(4.66)
Therefore
x x k k .
T
(4.67)
From the Cauchy-Schwartz inequality (see [Brog91])

2
x x k x
T 2 2
xk , (4.68)
where
2 T
x = x x. (4.69)
If we combine Eq. (4.67) and Eq. (4.68) we can put a lower bound on the
squared length of the weight vector at iteration k :
2
x x k k
T 2
2
xk -------------------------- ------------2- . (4.70)
x x
2
Next we want to find an upper bound for the length of the weight vector.
We begin by finding the change in the length at iteration k :
2 T
xk = x k x k
T
= x k 1 + z' k 1 x k 1 + z' k 1
(4.71)
T T
= x k 1 x k 1 + 2x k 1 z' k 1
T
+ z' k 1 z' k 1
Note that
T
x k 1 z' k 1 0 , (4.72)
4-17
since the weights would not be updated unless the previous input vector
had been misclassified. Now Eq. (4.71) can be simplified to
2 2 2
xk xk 1 + z' k 1 . (4.73)
2 2
We can repeat this process for x k 1 , x k 2 , etc., to obtain
+ + z' k 1 .
2 2 2
xk z' 0 (4.74)
2
If = max z' i , this upper bound can be simplified to
2
xk k . (4.75)
We now have an upper bound (Eq. (4.75)) and a lower bound (Eq. (4.70)) on
the squared length of the weight vector at iteration k . If we combine the
two inequalities we find
x
2 2
2 k
k x k ------------2- or k ----------------- . (4.76)
x
2

Because k has an upper bound, this means that the weights will only be
changed a finite number of times. Therefore, the perceptron learning rule
will converge in a finite number of iterations.
The maximum number of iterations (changes to the weight vector) is in-
versely related to the square of . This parameter is a measure of how close
the solution decision boundary is to the input patterns. This means that if
the input classes are difficult to separate (are close to the decision bound-
ary) it will take many iterations for the algorithm to converge.
Note that there are only three key assumptions required for the proof:
1. A solution to the problem exists, so that Eq. (4.66) is satisfied.
2. The weights are only updated when the input vector is misclassified,
therefore Eq. (4.72) is satisfied.
3. An upper bound, , exists for the length of the input vectors.
Because of the generality of the proof, there are many variations of the per-
ceptron learning rule that can also be shown to converge. (See Exercise
E4.13.)
Limitations
The perceptron learning rule is guaranteed to converge to a solution in a
finite number of steps, so long as a solution exists. This brings us to an im-
portant question. What problems can a perceptron solve? Recall that a sin-
4-18
gle-neuron perceptron is able to divide the input space into two regions.
The boundary between the regions is defined by the equation
T
1w p+b = 0. (4.77)
This is a linear boundary (hyperplane). The perceptron can be used to clas-

sify input vectors that can be separated by a linear boundary. We call such
Linear Separability vectors linearly separable. The logical AND gate example on page 4-7 illus-
trates a two-dimensional example of a linearly separable problem. The ap-
ple/orange recognition problem of Chapter 3 was a three-dimensional
example.
Unfortunately, many problems are not linearly separable. The classic ex-
ample is the XOR gate. The input/target pairs for the XOR gate are
0 t = 0 p = 0 t = 1 p = 1 t = 1 p = 1 t = 0 .
p1 = 1 2 2 3 3 4 4
0 1 0 1
This problem is illustrated graphically on the left side of Figure 4.6, which
also shows two other linearly inseparable problems. Try drawing a straight
line between the vectors with targets of 1 and those with targets of 0 in any
of the diagrams of Figure 4.6.
Figure 4.6 Linearly Inseparable Problems

It was the inability of the basic perceptron to solve such simple problems
that led, in part, to a reduction in interest in neural network research dur-
ing the 1970s. Rosenblatt had investigated more complex networks, which
he felt would overcome the limitations of the basic perceptron, but he was
never able to effectively extend the perceptron rule to such networks. In
Chapter 11 we will introduce multilayer perceptrons, which can solve arbi-
trary classification problems, and will describe the backpropagation algo-
rithm, which can be used to train them.
4-19
Summary of Results
p a
Rx1
W Sx1
SxR
n
Sx1
1 b
Sx1
R S
T
1w
T
W = 2w

T
Sw
a = hardlim Wp + b
T
a i = hardlim n i = hardlim iw p + b i
Decision Boundary
T
iw p + bi = 0 .
The decision boundary is always orthogonal to the weight vector.

Single-layer perceptrons can only classify linearly separable vectors.

T
W new = W old + ep
new old
b = b +e
where e = t a .
4-20
Solved Problems
Solved Problems
P4.1 Solve the three simple classification problems shown in Figure
P4.1 by drawing a decision boundary. Find weight and bias values
that result in single-neuron perceptrons with the chosen decision
boundaries.
(a) (b) (c)
Figure P4.1 Simple Classification Problems

First we draw a line between each set of dark and light data points.
(a) (b) (c)
The next step is to find the weights and biases. The weight vectors must be
orthogonal to the decision boundaries, and pointing in the direction of
points to be classified as 1 (the dark points). The weight vectors can have
any length we like.
1w
1w
1w
(a) (b) (c)
Here is one set of choices for the weight vectors:

T T T
(a) 1w = 2 1 , (b) 1w = 0 2 , (c) 1w = 2 2 .
4-21
Now we find the bias values for each perceptron by picking a point on the
decision boundary and satisfying Eq. (4.15).
T
1w p+b = 0
T
b = 1w p
This gives us the following three biases:
(a) b = 2 1 0 = 0 , (b) b = 0 2 0 = 2 , (c) b = 2 2 2 = 6

0 1 1
We can now check our solution against the original points. Here we test the
T
first network on the input vector p = 2 2 .
T
a = hardlim 1w p + b

= hardlim 2 1 2 + 0
2
= hardlim 6
= 1
2+2
ans = We can use MATLAB to automate the testing process and to try new points.
4 Here the first network is used to classify a point that was not in the original
problem.
w=[-2 1]; b = 0;
a = hardlim(w*[1;1]+b)
a =
0
P4.2 Convert the classification problem defined below into an equiva-

lent problem definition consisting of inequalities constraining
weight and bias values.
0 t = 1 p = 1 t = 1 p = 0 t = 0 p = 2 t = 0
p1 = 1 2 2 3 3 4 4
2 0 2 0
4-22
Solved Problems
Each target t i indicates whether or not the net input in response to p i must
be less than 0, or greater than or equal to 0. For example, since t 1 is 1, we
know that the net input corresponding to p 1 must be greater than or equal
to 0. Thus we get the following inequality:
Wp 1 + b 0
0w 1 1 + 2w 1 2 + b 0
2w 1 2 + b 0.
Applying the same procedure to the input/target pairs for p 2 t 2 , p 3 t 3

and p 4 t 4 results in the following set of inequalities.
2w 1 2 + b 0 i
w 1 1 + b 0 ii 4
2w 1 2 + b 0 iii
2w 1 1 + b 0 iv
Solving a set of inequalities is more difficult than solving a set of equalities.

One added complexity is that there are often an infinite number of solu-
tions (just as there are often an infinite number of linear decision bound-
aries that can solve a linearly separable classification problem).
However, because of the simplicity of this problem, we can solve it by
graphing the solution spaces defined by the inequalities. Note that w 1 1
only appears in inequalities (ii) and (iv), and w 1 2 only appears in inequal-
ities (i) and (iii). We can plot each pair of inequalities with two graphs.
w1,1 w1,2
ii
iv
b b
iii i
Any weight and bias values that fall in both dark gray regions will solve the
classification problem.
Here is one such solution:
W = 2 3 b = 3.
4-23
P4.3 We have a classification problem with four classes of input vector.

The four classes are
2 p = 2 ,
class 1: p 1 = 1 p 2 = 1 , class 2: p3 = 4
1 2 1 0

class 3: p 5 = 1 p 6 = 2 , class 4: p 7 = 1 p 8 = 2 .
2 1 1 2
Design a perceptron network to solve this problem.

To solve a problem with four classes of input vector we will need a percep-
tron with at least two neurons, since an S -neuron perceptron can catego-
S
rize 2 classes. The two-neuron perceptron is shown in Figure P4.2.
p a
2x1
W 2x1
2x2
n
2x1
1 b
2x1
2 2
Figure P4.2 Two-Neuron Perceptron

Lets begin by displaying the input vectors, as in Figure P4.3. The light cir-
cles indicate class 1 vectors, the light squares indicate class 2 vectors,
the dark circles indicate class 3 vectors, and the dark squares indicate
class 4 vectors.
A two-neuron perceptron creates two decision boundaries. Therefore, to di-
vide the input space into the four categories, we need to have one decision
boundary divide the four classes into two sets of two. The remaining bound-
ary must then isolate each class. Two such boundaries are illustrated in
Figure P4.4. We now know that our patterns are linearly separable.
4-24
Solved Problems
1
3
Figure P4.3 Input Vectors for Problem P4.3
1
3
Figure P4.4 Tentative Decision Boundaries for Problem P4.3

The weight vectors should be orthogonal to the decision boundaries and
should point toward the regions where the neuron outputs are 1. The next
step is to decide which side of each boundary should produce a 1. One choice
is illustrated in Figure P4.5, where the shaded areas represent outputs of
1. The darkest shading indicates that both neuron outputs are 1. Note that
this solution corresponds to target values of
0 t = 0 ,
class 1: t 1 = 0 t 2 = 0 , class 2: t3 = 4
0 0 1 1

class 3: t 5 = 1 t 6 = 1 , class 4: t 7 = 1 t 8 = 1 .
0 0 1 1
We can now select the weight vectors:
4-25
1w = 3 and 2w = 1 .
1 2
Note that the lengths of the weight vectors is not important, only their di-
rections. They must be orthogonal to the decision boundaries. Now we can
calculate the bias by picking a point on a boundary and satisfying Eq.
(4.15):
b 1 = 1w p = 3 1 0 = 1 ,
T
b 2 = 2w p = 1 2 0 = 0 .
T
1
3
Figure P4.5 Decision Regions for Problem P4.3

In matrix form we have
T
1w
W = = 3 1 and b = 1 ,
T 1 2 0
2w
which completes our design.
P4.4 Solve the following classification problem with the perceptron

rule. Apply each input vector in order, for as many repetitions as
it takes to ensure that the problem is solved. Draw a graph of the
problem only after you have found a solution.
4-26
Solved Problems
2 t = 0 p = 1 t = 1 p = 2 t = 0 p = 1 t = 1
p1 = 1 2 2 3 3 4 4
2 2 2 1
Use the initial weights and bias:
W0 = 0 0 b0 = 0 .
We start by calculating the perceptrons output a for the first input vector
p 1 , using the initial weights and bias.
a = hardlim W 0 p 1 + b 0

4
= hardlim 0 0 2 + 0 = hardlim 0 = 1
2
The output a does not equal the target value t 1 , so we use the perceptron
rule to find new weights and biases based on the error.
e = t1 a = 0 1 = 1
T
W 1 = W 0 + ep 1 = 0 0 + 1 2 2 = 2 2
b 1 = b 0 + e = 0 + 1 = 1
We now apply the second input vector p 2 , using the updated weights and
bias.

= hardlim 2 2 1 1 = hardlim 1 = 1
2
This time the output a is equal to the target t 2 . Application of the percep-
tron rule will not result in any changes.
W2 = W1
b2 = b1
We now apply the third input vector.
4-27

2
The output in response to input vector p 3 is equal to the target t 3 , so there

will be no changes.
W3 = W2
b3 = b2
We now move on to the last input vector p 4 .

1
This time the output a does not equal the appropriate target t 4 . The per-
ceptron rule will result in a new set of values for W and b .
e = t4 a = 1 0 = 1
T
W 4 = W 3 + ep 4 = 2 2 + 1 1 1 = 3 1
b4 = b3 + e = 1 + 1 = 0
We now must check the first vector p 1 again. This time the output a is
equal to the associated target t 1 .

2
Therefore there are no changes.
W5 = W4
b5 = b4
4-28
Solved Problems
The second presentation of p 2 results in an error and therefore a new set

of weight and bias values.

2
Here are those new values:
e = t2 a = 1 0 = 1
T
W 6 = W 5 + ep 2 = 3 1 + 1 1 2 = 2 3 4
b 6 = b 5 + e = 0 + 1 = 1.
Cycling through each input vector once more results in no errors.

a = hardlim W 6 p 3 + b 6 = hardlim 2 3 2 + 1 = 0 = t 3
2

1

2

2
Therefore the algorithm has converged. The final solution is:
W = 2 3 b = 1.
Now we can graph the training data and the decision boundary of the solu-
tion. The decision boundary is given by
n = Wp + b = w 1 1 p 1 + w 1 2 p 2 + b = 2p 1 3p 2 + 1 = 0 .
To find the p 2 intercept of the decision boundary, set p 1 = 0 :
b 1 1
p 2 = ---------- = ------ = --- if p 1 = 0
w 1 2 3 3
.
4-29
To find the p 1 intercept, set p 2 = 0 :
b 1 1
p 1 = ---------- = ------ = --- if p 2 = 0
w 1 1 2 2
.
The resulting decision boundary is illustrated in Figure P4.6.
Figure P4.6 Decision Boundary for Problem P4.4

Note that the decision boundary falls across one of the training vectors.
This is acceptable, given the problem definition, since the hard limit func-
tion returns 1 when given an input of 0, and the target for the vector in
question is indeed 1.
P4.5 Consider again the four-class decision problem that we introduced

in Problem P4.3. Train a perceptron network to solve this problem
using the perceptron learning rule.
If we use the same target vectors that we introduced in Problem P4.3, the
training set will be:
1 t = 0 p = 1 t = 0 p = 2 t = 0
p1 = 1 2 2 3 3
1 0 2 0 1 1
2 t = 0 p = 1 t = 1 p = 2 t = 1
p4 = 4 5 5 6 6
0 1 2 0 1 0
1 t = 1 p = 2 t = 1 .
p7 = 7 8 8
1 1 2 1
Lets begin the algorithm with the following initial weights and biases:
W 0 = 1 0 , b 0 = 1 .
0 1 1
The first iteration is
4-30
Solved Problems
a = hardlim (W 0 p 1 + b 0 ) = hardlim ( 1 0 1 + 1 ) = 1 ,
0 1 1 1 1
e = t1 a = 0 1 = 1 ,
0 1 1
W 1 = W 0 + ep 1 = 1 0 + 1 1 1 = 0 1 ,
T
0 1 1 1 0
b 1 = b 0 + e = 1 + 1 = 0 .
1 1 0 4
The second iteration is
1 0 2 0 0
e = t2 a = 0 0 = 0 ,
0 0 0
T
W 2 = W 1 + ep 2 = 0 1 + 0 0 1 ,
1 2 =
1 0 0 1 0
b2 = b1 + e = 0 + 0 = 0 .
0 0 0
The third iteration is
1 0 1 0 0
e = t3 a = 0 1 = 1 ,
1 0 1
4-31
T
W 3 = W 2 + ep 3 = 0 1 + 1 2 0 ,
2 1 =
1 0 1 1 1
b 3 = b 2 + e = 0 + 1 = 1 .
0 1 1
Iterations four through eight produce no changes in the weights.
W8 = W7 = W6 = W5 = W4 = W3
b8 = b7 = b6 = b 5 = b 4 = b3
The ninth iteration produces
1 1 1 1 1
e = t1 a = 0 0 = 0 ,
0 1 1
W 9 = W 8 + ep 1 = 2 0 + 0 1 1 = 2 0 ,
T
1 1 1 0 2
b 9 = b 8 + e = 1 + 0 = 1 .
1 1 0
At this point the algorithm has converged, since all input patterns will be
correctly classified. The final decision boundaries are displayed in Figure
P4.7. Compare this result with the network we designed in Problem P4.3.
4-32
Solved Problems
1
3
Figure P4.7 Final Decision Boundaries for Problem P4.5
4-33
Epilogue
In this chapter we have introduced our first learning rule the perceptron
learning rule. It is a type of learning called supervised learning, in which
the learning rule is provided with a set of examples of proper network be-
havior. As each input is applied to the network, the learning rule adjusts
the network parameters so that the network output will move closer to the
target.
The perceptron learning rule is very simple, but it is also quite powerful.
We have shown that the rule will always converge to a correct solution, if
such a solution exists. The weakness of the perceptron network lies not
with the learning rule, but with the structure of the network. The standard
perceptron is only able to classify vectors that are linearly separable. We
will see in Chapter 11 that the perceptron architecture can be generalized
to multilayer perceptrons, which can solve arbitrary classification prob-
lems. The backpropagation learning rule, which is introduced in Chapter
11, can be used to train these networks.
In Chapters 3 and 4 we have used many concepts from the field of linear
algebra, such as inner product, projection, distance (norm), etc. We will find
in later chapters that a good foundation in linear algebra is essential to our
understanding of all neural networks. In Chapters 5 and 6 we will review
some of the key concepts from linear algebra that will be most important in
our study of neural networks. Our objective will be to obtain a fundamental
understanding of how neural networks work.
4-34
Further Reading
Further Reading
[BaSu83] A. Barto, R. Sutton and C. Anderson, Neuron-like adap-
tive elements can solve difficult learning control problems,
IEEE Transactions on Systems, Man and Cybernetics, Vol.
13, No. 5, pp. 834846, 1983.
A classic paper in which a reinforcement learning algo-
rithm is used to train a neural network to balance an in-
verted pendulum.
[Brog91] W. L. Brogan, Modern Control Theory, 3rd Ed., Englewood
Cliffs, NJ: Prentice-Hall, 1991.
A well-written book on the subject of linear systems. The 4
first half of the book is devoted to linear algebra. It also has
good sections on the solution of linear differential equa-
tions and the stability of linear and nonlinear systems. It
has many worked problems.
Biophysics, Vol. 5, pp. 115133, 1943.
neuron, in which a weighted sum of input signals is com-
ron fires.
MIT Press, 1969.
A landmark book that contains the first rigorous study de-
voted to determining what a perceptron network is capable
of learning. A formal treatment of the perceptron was need-
ed both to explain the perceptrons limitations and to indi-
cate directions for overcoming them. Unfortunately, the
book pessimistically predicted that the limitations of per-
ceptrons indicated that the field of neural networks was a
dead end. Although this was not true, it temporarily cooled
research and funding for research for several years.
logical Review, Vol. 65, pp. 386408, 1958.
This paper presents the first practical artificial neural net-
work the perceptron.
4-35
[Rose61] F. Rosenblatt, Principles of Neurodynamics, Washington

DC: Spartan Press, 1961.
One of the first books on neurocomputing.
[WhSo92] D. White and D. Sofge (Eds.), Handbook of Intelligent Con-
trol, New York: Van Nostrand Reinhold, 1992.
Collection of articles describing current research and appli-
cations of neural networks and fuzzy logic to control sys-
tems.
4-36
Exercises
Exercises
E4.1 Consider the classification problem defined below:
1 t = 1 p = 0 t = 1 p = 1 t = 1 p = 1 t = 0
p1 = 1 2 2 3 3 4 4
1 0 1 0
0 t = 0.
p5 = 5
1
i. Draw a diagram of the single-neuron perceptron you would use to

solve this problem. How many inputs are required?
ii. Draw a graph of the data points, labeled according to their targets.
Is this problem solvable with the network you defined in part (i)?
Why or why not?
E4.2 Consider the classification problem defined below.
1 t = 1 p = 1 t = 1 p = 0 t = 0 p = 1 t = 0 .
p1 = 1 2 2 3 3 4 4
1 1 0 0
i. Design a single-neuron perceptron to solve this problem. Design the

network graphically, by choosing weight vectors that are orthogonal
to the decision boundaries.
ii. Test your solution with all four input vectors.
2+2
ans = iii. Classify the following input vectors with your solution. You can ei-
4 ther perform the calculations manually or with MATLAB.
p5 = 2 p6 = 1 p7 = 0 p8 = 1
0 1 1 2
iv. Which of the vectors in part (iii) will always be classified the same
way, regardless of the solution values for W and b ? Which may
vary depending on the solution? Why?
E4.3 Solve the classification problem in Exercise E4.2 by solving inequalities (as
in Problem P4.2), and repeat parts (ii) and (iii) with the new solution. (The
solution is more difficult than Problem P4.2, since you cant isolate the
weights and biases in a pairwise manner.)
4-37
E4.4 Solve the classification problem in Exercise E4.2 by applying the percep-
tron rule to the following initial parameters, and repeat parts (ii) and (iii)
with the new solution.
W0 = 0 0 b0 = 0
E4.5 Prove mathematically (not graphically) that the following problem is un-
solvable for a two-input/single-neuron perceptron.
1 t = 1 p = 1 t = 0 p = 1 t = 1 p = 1 t = 0
p1 = 1 2 2 3 3 4 4
1 1 1 1
(Hint: start by rewriting the input/target requirements as inequalities that

constrain the weight and bias values.)
E4.6 We have four categories of vectors.

Category I: 1 1 , Category II: 0 1
1 0 2 2

Category III: 2 2 , Category IV: 1 0
0 1 1 1
i. Design a two-neuron perceptron network (single layer) to recognize

these four categories of vectors. Sketch the decision boundaries.
iii. Suppose the following vector is to be added to Category I.
1
3
Perform one iteration of the perceptron learning rule with this vec-
tor. (Start with the weights you determined in part i.) Draw the new
decision boundaries.
E4.7 We have two categories of vectors. Category I consists of
0 1 0
.
0 0 1
4-38
Exercises
Category II consists of
1 0 2
.
1 2 0
i. Design a single-neuron perceptron network to recognize these two

categories of vectors.
iii. Sketch the decision boundary.
iv. If we add the following vector to Category I, will your network clas-
sify it correctly? Demonstrate by computing the network response.
4
3
0
v. Can your weight matrix and bias be modified so your network can
classify this new vector correctly (while continuing to classify the
other vectors correctly)? Explain.
E4.8 We want to train a perceptron network with the following training set:
1 t = 0 p = 0 t = 0 p = 1 t = 1 .
p1 = 1 2 2 3 3
1 0 1
The initial weight matrix and bias are
W 0 = 1 0 , b 0 = 0.5 .
i. Plot the initial decision boundary, weight vector and input patterns.
Which patterns are correctly classified using the initial weight and
bias?
ii. Train the network with the perceptron rule. Present each input vec-
tor once, in the order shown.
iii. Plot the final decision boundary, and demonstrate graphically
which patterns are correctly classified.
iv. Will the perceptron rule (given enough iterations) always learn to
correctly classify the patterns in this training set, no matter what
initial weights we use? Explain.
4-39
E4.9 We want to train a perceptron network using the following training set:
1 t = 0 p = 1 t = 0 p = 1 t = 1 ,
p1 = 1 2 2 3 3
0 2 2
starting from the initial conditions
W0 = 0 1 , b0 = 1 .
i. Sketch the initial decision boundary, and show the weight vector
and the three training input vectors, p 1 p 2 p 3 . Indicate the class of
each input vector, and show which ones are correctly classified by
the initial decision boundary.
ii. Present the input p 1 to the network, and perform one iteration of
the perceptron learning rule.
iii. Sketch the new decision boundary and weight vector, and again in-
dicate which of the three input vectors are correctly classified.
iv. Present the input p 2 to the network, and perform one more itera-
tion of the perceptron learning rule.
v. Sketch the new decision boundary and weight vector, and again in-
dicate which of the three input vectors are correctly classified.
vi. If you continued to use the perceptron learning rule, and presented
all of the patterns many times, would the network eventually learn
to correctly classify the patterns? Explain your answer. (This part
does not require any calculations.)
E4.10 The symmetric hard limit function is sometimes used in perceptron net-
works, instead of the hard limit function. Target values are then taken
a = hardlims (n) from the set [-1, 1] instead of [0, 1].
i. Write a simple expression that maps numbers in the ordered set [0,
1] into the ordered set [-1, 1]. Write the expression that performs
the inverse mapping.
n = Wp + b ii. Consider two single-neuron perceptrons with the same weight and
bias values. The first network uses the hard limit function ([0, 1]
values), and the second network uses the symmetric hard limit func-
tion. If the two networks are given the same input p , and updated
with the perceptron learning rule, will their weights continue to
have the same value?
iii. If the changes to the weights of the two neurons are different, how
do they differ? Why?
4-40
Exercises
iv. Given initial weight and bias values for a standard hard limit per-
ceptron, create a method for initializing a symmetric hard limit per-
ceptron so that the two neurons will always respond identically
when trained on identical data.
E4.11 The vectors in the ordered set defined below were obtained by measuring
the weight and ear lengths of toy rabbits and bears in the Fuzzy Wuzzy An-
imal Factory. The target values indicate whether the respective input vec-
2+2 tor was taken from a rabbit (0) or a bear (1). The first element of the input
ans = vector is the weight of the toy, and the second element is the ear length.
4
1 t = 0 p = 1 t = 0 p = 2 t = 0 p = 2 t = 0
p1 = 1 2 2 3 3 4 4
4 5 4 5
3 t = 1 p = 3 t = 1 p = 4 t = 1 p = 4 t = 1
p5 = 5 6 6 7 7 8 8
1 2 1 2
i. Use MATLAB to initialize and train a network to solve this practi-

cal problem.
ii. Use MATLAB to test the resulting weight and bias values against
the input vectors.
iii. Add input vectors to the training set to ensure that the decision
boundary of any solution will not intersect one of the original input
vectors (i.e., to ensure only robust solutions are found). Then retrain
the network. Your method for adding the input vectors should be
general purpose (not designed specifically for this problem).
E4.12 Consider again the four-category classification problem described in Prob-

lems P4.3 and P4.5. Suppose that we change the input vector p 3 to
p3 = 2 .
2
i. Is the problem still linearly separable? Demonstrate your answer
graphically.
2+2
ans = ii. Use MATLAB to initialize and train a network to solve this prob-
4 lem. Explain your results.
iii. If p 3 is changed to
p3 = 2
1.5
4-41
is the problem linearly separable?

iv. With the p 3 from (iii), use MATLAB to initialize and train a net-
work to solve this problem. Explain your results.
E4.13 One variation of the perceptron learning rule is

new old T
W = W + ep
new old
b = b + e
where is called the learning rate. Prove convergence of this algorithm.

Does the proof require a limit on the learning rate? Explain.
4-42
Objectives
5 Signal and Weight

Vector Spaces
Objectives 5-1
Linear Vector Spaces 5-2
Linear Independence 5-4
Spanning a Space 5-5
Inner Product 5-6
Norm 5-7
Orthogonality 5-7
Gram-Schmidt Orthogonalization 5-8
Vector Expansions 5-9
Reciprocal Basis Vectors 5-10
Epilogue 5-26
Exercises 5-28
Objectives
It is clear from Chapters 3 and 4 that it is very useful to think of the inputs
and outputs of a neural network, and the rows of a weight matrix, as vec-
tors. In this chapter we want to examine these vector spaces in detail and
to review those properties of vector spaces that are most helpful when an-
alyzing neural networks. We will begin with general definitions and then
apply these definitions to specific neural network problems. The concepts
that are discussed in this chapter and in Chapter 6 will be used extensively
throughout the remaining chapters of this book. They are critical to our un-
derstanding of why neural networks work.
5-1
5 Signal and Weight Vector Spaces
Theory and Examples

Linear algebra is the core of the mathematics required for understanding
neural networks. In Chapters 3 and 4 we saw the utility of representing the
inputs and outputs of neural networks as vectors. In addition, we saw that
it is often useful to think of the rows of a weight matrix as vectors in the
same vector space as the input vectors.
Recall from Chapter 3 that in the Hamming network the rows of the weight
matrix of the feedforward layer were equal to the prototype vectors. In fact,
the purpose of the feedforward layer was to calculate the inner products be-
tween the prototype vectors and the input vector.
In the single neuron perceptron network we noted that the decision bound-
ary was always orthogonal to the weight matrix (a row vector).
In this chapter we want to review the basic concepts of vector spaces (e.g.,
inner products, orthogonality) in the context of neural networks. We will
begin with a general definition of vector spaces. Then we will present the
basic properties of vectors that are most useful for neural network applica-
tions.
One comment about notation before we begin. All of the vectors we have
discussed so far have been ordered n-tuples (columns) of real numbers and
are represented by bold small letters, e.g.,
T
x = x1 x2 xn . (5.1)
n
These are vectors in , the standard n-dimensional Euclidean space. In
this chapter we will also be talking about more general vector spaces than
n
. These more general vectors will be represented with a script typeface,
as in x . We will show in this chapter how these general vectors can often
be represented by columns of numbers.
Linear Vector Spaces

What do we mean by a vector space? We will begin with a very general def-
inition. While this definition may seem abstract, we will provide many con-
crete examples. By using a general definition we can solve a larger class of
problems, and we can impart a deeper understanding of the concepts.
Vector Space Definition. A linear vector space, X , is a set of elements (vectors) defined
over a scalar field, F , that satisfies the following conditions:
1. An operation called vector addition is defined such that if x X ( x is

an element of X ) and y X , then x + y X .
5-2
2. x+y = y+x.
3. x + y + z = x + y + z .
4. There is a unique vector 0 X , called the zero vector, such that

x + 0 = x for all x X .
5. For each vector x X there is a unique vector in X, to be called x , such

that x + x = 0 .
6. An operation, called multiplication, is defined such that for all scalars

a F , and all vectors x X , a x X .
x2 7. For any x X , 1 x = x (for scalar 1 ).
8. For any two scalars a F and b F , and any x X , a b x = ab x .

2 5
9. a + b x = ax + bx .
x1
10. a x + y = a x + a y .
To illustrate these conditions, lets investigate a few sample sets and deter-
mine whether or not they are vector spaces. First consider the standard
2
x+y two-dimensional Euclidean space, , shown in the upper left figure. This
is clearly a vector space, and all ten conditions are satisfied for the stan-
dard definitions of vector addition and scalar multiplication.
y x 2 2
What about subsets of ? What subsets of are also vector spaces (sub-
spaces)? Consider the boxed area ( X ) in the center left figure. Does it sat-
X isfy all ten conditions? No. Clearly even condition 1 is not satisfied. The
vectors x and y shown in the figure are in X , but x + y is not. From this
example it is clear that no bounded sets can be vector spaces.
2
Are there any subsets of that are vector spaces? Consider the line ( X )
x2
shown in the bottom left figure. (Assume that the line extends to infinity in
both directions.) Is this line a vector space? We leave it to you to show that
indeed all ten conditions are satisfied. Will any such infinite line satisfy the
ten conditions? Well, any line that passes through the origin will work. If
it does not pass through the origin then condition 4, for instance, would not
x1 be satisfied.
X
In addition to the standard Euclidean spaces, there are other sets that also
satisfy the ten conditions of a vector space. Consider, for example, the set
2
P of all polynomials of degree less than or equal to 2. Two members of this
set would be
5-3
x = 2 + t + 4t 2
y = 1 + 5t . (5.2)
If you are used to thinking of vectors only as columns of numbers, these
may seem to be strange vectors indeed. However, recall that to be a vector
space, a set need only satisfy the ten conditions we presented. Are these
2
conditions satisfied for the set P ? If we add two polynomials of degree less
than or equal to 2, the result will also be a polynomial of degree less than
or equal to 2. Therefore condition 1 is satisfied. We can also multiply a poly-
nomial by a scalar without changing the order of the polynomial. Therefore
condition 6 is satisfied. It is not difficult to show that all ten conditions are
2
satisfied, showing that P is a vector space.
f(t) Consider the set C 0 1 of all continuous functions defined on the interval
[0, 1]. Two members of this set would be
1
x = sin t
t
y = e 2t . (5.3)
Another member of the set is shown in the figure to the left.
The sum of two continuous functions is also a continuous function, and a
scalar times a continuous function is a continuous function. The set C 0 1
is also a vector space. This set is different than the other vector spaces we
have discussed; it is infinite dimensional. We will define what we mean by
dimension later in this chapter.
Linear Independence
Now that we have defined what we mean by a vector space, we will inves-
tigate some of the properties of vectors. The first properties are linear de-
pendence and linear independence.
Consider n vectors x 1 x 2 x n . If there exist n scalars a 1 a 2 a n , at
least one of which is nonzero, such that
a1 x1 + a2 x2 + + an xn = 0, (5.4)
then the x i are linearly dependent.

The converse statement would be: If a 1 x 1 + a 2 x 2 + + a n x n = 0 implies
Linear Independence that each a i = 0 , then x i is a set of linearly independent vectors.
5-4
Spanning a Space
Note that these definitions are equivalent to saying that if a set of vectors
is independent then no vector in the set can be written as a linear combi-
nation of the other vectors.
2 As an example of independence, consider the pattern recognition problem
+2
of Chapter 3. The two prototype patterns (orange and apple) were given by:
1 1
p1 = 1 , p2 = 1 . (5.5)
1 1
Let a 1 p 1 + a 2 p 2 = 0 , then
a1 + a2 0
a1 + a2 = 0 , (5.6)
a1 + a2 0
but this can only be true if a 1 = a 2 = 0 . Therefore p 1 and p 2 are linearly 5

independent.
2
+2
Consider vectors from the space P2 of polynomials of degree less than or
equal to 2. Three vectors from this space would be
x 1 = 1 + t + t2 , x2 = 2 + 2t + t2 , x 3 = 1 + t . (5.7)
Note that if we let a 1 = 1 , a 2 = 1 and a 3 = 1 , then
a1 x1 + a2 x2 + a3 x3 = 0. (5.8)
Therefore these three vectors are linearly dependent.
Spanning a Space
Next we want to define what we mean by the dimension (size) of a vector
space. To do so we must first define the concept of a spanning set.
Let X be a linear vector space and let u 1 u 2 u m be a subset of general
vectors in X. This subset spans X if and only if for every vector x X there
exist scalars x 1 x 2 x n such that x = x 1 u 1 + x 2 u 2 + + x m u m . In other
words, a subset spans a space if every vector in the space can be written as
a linear combination of the vectors in the subset.
The dimension of a vector space is determined by the minimum number of
vectors it takes to span the space. This leads to the definition of a basis set.
Basis Set A basis set for X is a set of linearly independent vectors that spans X. Any
basis set contains the minimum number of vectors required to span the
5-5
space. The dimension of X is therefore equal to the number of elements in

the basis set. Any vector space can have many basis sets, but each one must
contain the same number of elements. (See [Stra80] for a proof of this fact.)
Take, for example, the linear vector space P2. One possible basis for this
space is
u1 = 1 , u2 = t , u3 = t2 . (5.9)
Clearly any polynomial of degree two or less can be created by taking a lin-
ear combination of these three vectors. Note, however, that any three inde-
pendent vectors from P2 would form a basis for this space. One such
alternate basis is:
u1 = 1 , u2 = 1 + t , u3 = 1 + t + t2 . (5.10)
Inner Product
From our brief encounter with neural networks in Chapters 3 and 4, it is
clear that the inner product is fundamental to the operation of many neural
networks. Here we will introduce a general definition for inner products
and then give several examples.
Inner Product Any scalar function of x and y can be defined as an inner product, (x,y) , pro-
vided that the following properties are satisfied:
1. (x,y) = (y,x) .
2. (x,a y 1 + b y 2) = a (x,y 1) + b (x,y 2) .
3. (x,x) 0 , where equality holds if and only if x is the zero vector.
The standard inner product for vectors in Rn is
x y = x1 y1 + x2 y2 + + xn yn ,
T
(5.11)
but this is not the only possible inner product. Consider again the set C[0, 1]
of all continuous functions defined on the interval [0, 1]. Show that the fol-
lowing scalar function is man inner product (see Problem P5.6).
1
(x,y) = x t y t dt (5.12)
0
5-6
Norm
Norm
The next operation we need to define is the norm, which is based on the con-
cept of vector length.
Norm A scalar function x is called a norm if it satisfies the following properties:
1. x 0.
2. x = 0 if and only if x = 0 .
3. ax = a x for scalar a.
4. x+y x + y .
There are many functions that would satisfy these conditions. One common
norm is based on the inner product:
12
x = (x,x) . (5.13)
n
For Euclidean spaces, , this yields the norm with which we are most fa-
miliar:
5
12
x1 + x2 + + xn .
T 2 2 2
x = x x = (5.14)
In neural network applications it is often useful to normalize the input vec-

tors. This means that p i = 1 for each input vector.
Using the norm and the inner product we can generalize the concept of an-
Angle gle for vector spaces of dimension greater than two. The angle between
two vectors x and y is defined by
(x,y)
cos = --------------- . (5.15)
x y
Orthogonality
Now that we have defined the inner product operation, we can introduce
the important concept of orthogonality.
Orthogonality Two vectors x y X are said to be orthogonal if (x,y) = 0 .
Orthogonality is an important concept in neural networks. We will see in
Chapter 7 that when the prototype vectors of a pattern recognition problem
are orthogonal and normalized, a linear associator neural network can be
trained, using the Hebb rule, to achieve perfect recognition.
In addition to orthogonal vectors, we can also have orthogonal spaces. A
vector x X is orthogonal to a subspace X 1 if x is orthogonal to every vec-
5-7
tor in X 1 . This is typically represented as x X 1 . A subspace X 1 is orthog-

onal to a subspace X 2 if every vector in X 1 is orthogonal to every vector in
p3
X 2 . This is represented by X 1 X 2 .
The figure to the left illustrates the two orthogonal spaces that were used
in the perceptron example of Chapter 3. (See Figure 3.4.) The p 1 p 3 plane
p2 3
is a subspace of , which is orthogonal to the p 2 axis (which is another
3
subspace of ). The p 1 p 3 plane was the decision boundary of a percep-
tron network. In Solved Problem P5.1 we will show that the perceptron de-
p1 cision boundary will be a vector space whenever the bias value is zero.
Gram-Schmidt Orthogonalization
There is a relationship between orthogonality and independence. It is pos-
sible to convert a set of independent vectors into a set of orthogonal vectors
that spans the same vector space. The standard procedure to accomplish
this is called Gram-Schmidt orthogonalization.
Assume that we have n independent vectors y 1 y 2 y n . From these vec-
tors we want to obtain n orthogonal vectors v 1 v 2 v n . The first orthog-
onal vector is chosen to be the first independent vector:
v1 = y1 . (5.16)
To obtain the second orthogonal vector we use y 2 , but subtract off the por-
tion of y 2 that is in the direction of v 1 . This leads to the equation
v2 = y2 a v1 , (5.17)
where a is chosen so that v2 is orthogonal to v 1 . This requires that

(v 1,v 2) = (v 1,y 2 a v 1) = (v 1,y 2) a (v 1,v 1) = 0 , (5.18)
or
(v 1,y 2)
-.
a = --------------- (5.19)
(v 1,v 1)
Therefore to find the component of y 2 in the direction of v 1 , a v 1 , we need

to find the inner product between the two vectors. We call a v 1 the projec-
Projection tion of y 2 on the vector v 1 .
If we continue this process, the kth step will be
k1
(v i,y k)
v k = y k --------------
- vi . (5.20)
(v i,v i)
i=1
5-8
Vector Expansions
To illustrate this process, we consider the following independent vectors in

2 2
+2 :
y1 = 2 , y2 = 1 . (5.21)
1 2
The first orthogonal vector would be
v1 = y1 = 2 . (5.22)
1
The second orthogonal vector is calculated as follows:
1
T 2 1
v1 y2 2
- v 1 = 1 --------------------- 2 = 1 1.6 = 0.6 .
v 2 = y 2 ----------
T
(5.23)
v1 v1 2 2 1 2 0.8 1.2
2 1
1
See Figure 5.1 for a graphical representation of this process.
y2 v2
y2
y1, v1
av1
Figure 5.1 Gram-Schmidt Orthogonalization Example

Orthonormal We could convert v 1 and v 2 to a set of orthonormal (orthogonal and nor-
malized) vectors by dividing each vector by its norm.
To experiment with this orthogonalization process, use the Neural Network
Design Demonstration Gram-Schmidt (nnd5gs).
Vector Expansions
Note that we have been using a script font ( x ) to represent general vectors
n
and bold type ( x ) to represent vectors in , which can be written as col-
umns of numbers. In this section we will show that general vectors in finite
5-9
dimensional vector spaces can also be written as columns of numbers and

n
therefore are in some ways equivalent to vectors in .
If a vector space X has a basis set v 1 v 2 v n , then any x X has a
Vector Expansion unique vector expansion:
n
x = xi vi = x1 v1 + x2 v2 + + xn vn . (5.24)
i=1
Therefore any vector in a finite dimensional vector space can be represent-

ed by a column of numbers:
T
x = x1 x2 xn . (5.25)
This x is a representation of the general vector x . Of course in order to in-

terpret the meaning of x we need to know the basis set. If the basis set
changes, x will change, even though it still represents the same general
vector x . We will discuss this in more detail in the next subsection.
If the vectors in the basis set are orthogonal ( (v i,v j) = 0 , i j ) it is very
easy to compute the coefficients in the expansion. We simply take the inner
product of v j with both sides of Eq. (5.24):
n n
(v j,x) = (v j, x i v i) = x i (vj,vi) = x j (v j,v j) . (5.26)

i=1 i=1
Therefore the coefficients of the expansion are given by
(v j,x)
-.
x j = ------------- (5.27)
(v j,v j)
When the vectors in the basis set are not orthogonal, the computation of the
coefficients in the vector expansion is more complex. This case is covered in
the following subsection.
Reciprocal Basis Vectors

If a vector expansion is required and the basis set is not orthogonal, the re-
ciprocal basis vectors are introduced. These are defined by the following
equations:
(r i,v j) = 0 ij
(5.28)
= 1 i = j,
5-10
Vector Expansions
Reciprocal Basis Vectors where the basis vectors are v 1 v 2 v n and the reciprocal basis vectors
are r 1 r 2 r n .
If the vectors have been represented by columns of numbers (through vec-
tor expansion), and the standard inner product is used
T
(r i,v j) = r i v j , (5.29)
then Eq. (5.28) can be represented in matrix form as

T
R B = I, (5.30)
where
B = v1 v2 vn , (5.31)
R = r1 r2 rn . (5.32)
Therefore R can be found from

T 1
R = B , (5.33)
and the reciprocal basis vectors can be obtained from the columns of R.
Now consider again the vector expansion
x = x1 v1 + x2 v2 + + xn vn . (5.34)
Taking the inner product of r1 with both sides of Eq. (5.34) we obtain
(r 1,x) = x 1 (r 1,v 1) + x 2 (r 1,v 2) + + x n (r 1,v n) . (5.35)
By definition
(r 1,v 2) = (r 1,v 3) = = (r 1,v n) = 0
(r 1,v 1) = 1 .
(5.36)
Therefore the first coefficient of the expansion is
x 1 = (r 1,x) , (5.37)
and in general
x j = (r j,x) . (5.38)
5-11
As an example, consider the two basis vectors

2
+2
v1 = 2 , v2 = 1 .
s s
(5.39)
1 2
Suppose that we want to expand the vector
0
s
x = 3 (5.40)
---
v2 2
in terms of the two basis vectors. (We are using the superscript s to indi-
cate that these columns of numbers represent expansions of the vectors in
s2 v1 2
terms of the standard basis in . The elements of the standard basis are
indicated in the adjacent figure as the vectors s 1 and s 2 . We need to use
this explicit notation in this example because we will be expanding the vec-
s1 tors in terms of two different basis sets.)
The first step in the vector expansion is to find the reciprocal basis vectors.
2 1 2 1
1 --- --- --- ---
R = 2 1 3 3 3 3 .
T
= r1 = r2 = (5.41)
1 2 1 2 1 2
--- --- --- ---
3 3 3 3
Now we can find the coefficients in the expansion.
0
x1 = r1 x = 2 1 1
v T s
--- --- 3 = ---
3 3 --- 2
2
0
x2 = r2 x = 1 2
v T s
--- --- 3 = 1 (5.42)
3 3 ---
2
or, in matrix form,
2 1
--- --- 0 1
v T s 3 1 s 3 ---
x = R x = B x = 3 = 2 . (5.43)
1 2 ---
--- --- 2 1
3 3
So that
5-12
Vector Expansions
x = 1--- v 1 + 1 v 2 , (5.44)
2
as indicated in Figure 5.2.
v2
x
v1
v2
- v1/2
Figure 5.2 Vector Expansion

Note that we now have two different vector expansions for
s v
x , represented
by x and x . In other words,
x = 0 s 1 + 3--- s 2 = 1--- v 1 + 1 v 2 . (5.45)

2 2
When we represent a general vector as a column of numbers we need to

know what basis set was used for the expansion. In this text, unless other-
wise stated, assume the standard basis set was used.
Eq. (5.43) shows the relationship between the two different representations
v 1 s
of x , x = B x . This operation, called a change of basis, will become very
important in later chapters for the performance analysis of certain neural
networks.
To experiment with the vector expansion process, use the Neural Network
Design Demonstration Reciprocal Basis (nnd5rb).
5-13
Summary of Results

Definition. A linear vector space, X, is a set of elements (vectors) defined
over a scalar field, F, that satisfies the following conditions:
1. An operation called vector addition is defined such that if x X and

y X , then x + y X .
2. x+y = y+x.
3. x + y + z = x + y + z .
4. There is a unique vector 0 X , called the zero vector, such that

x + 0 = x for all x X .
5. For each vector x X there is a unique vector in X, to be called x , such

that x + x = 0 .
6. An operation, called multiplication, is defined such that for all scalars

a F , and all vectors x X , a x X .
7. For any x X , 1 x = x (for scalar 1 ).
8. For any two scalars a F and b F , and any x X , a b x = ab x .
9. a + b x = ax + bx .
10. a x + y = a x + a y .
Linear Independence
Consider n vectors x 1 x 2 x n . If there exist n scalars a 1 a 2 a n , at
least one of which is nonzero, such that
a1 x1 + a2 x2 + + an xn = 0,
then the x i are linearly dependent.
5-14
Summary of Results
Spanning a Space
Let X be a linear vector space and let u 1 u 2 u m be a subset of vectors
in X. This subset spans X if and only if for every vector x X there exist
scalars x 1 x 2 x n such that x = x 1 u 1 + x 2 u 2 + + x m u m .
Inner Product
Any scalar function of x and y can be defined as an inner product, (x,y), pro-
vided that the following properties are satisfied.
1. (x,y) = (y,x) .
2. (x,a y 1 + b y 2) = a (x,y 1) + b (x,y 2) .
Norm
A scalar function
1. x 0.
x is called a norm if it satisfies the following properties:
5
2. x = 0 if and only if x = 0 .
3. ax = a x for scalar a.
4. x+y x + y .
Angle
The angle between two vectors x and y is defined by
(x,y)
cos = --------------- .
x y
Orthogonality
Two vectors x y X are said to be orthogonal if (x,y) = 0 .
Gram-Schmidt Orthogonalization
Assume that we have n independent vectors y 1 y 2 y n . From these vec-
tors we will obtain n orthogonal vectors v 1 v 2 v n .
v1 = y1
5-15
k1
(v i,y k)
v k = yk --------------
- vi ,
( v i ,v i )
i=1
where
(v i,y k)
--------------- v i
(v i,v i)
is the projection of y k on v i .
Vector Expansions
n
x = xi vi = x1 v1 + x2 v2 + + xn vn .
i=1
For orthogonal vectors,
(v j,x)
x j = -------------
-
(v j,v j)
Reciprocal Basis Vectors

(r i,v j) = 0 ij
= 1 i = j
x j = (r j,x) .
To compute the reciprocal basis vectors:
B = v1 v2 vn ,
R = r1 r2 rn ,
T 1
R = B .
In matrix form:
v 1 s
x = B x .
5-16
Solved Problems
Solved Problems
P5.1 Consider the single-neuron perceptron network shown in Figure
P5.1. Recall from Chapter 3 (see Eq. (3.6)) that the decision bound-
ary for this network is given by Wp + b = 0 . Show that the decision
boundary is a vector space if b = 0 .

- Title - Layer
p a
Rx1 W 1x1
1xR n
1x1
1 b
R 1x1 1
- Exp (Wp
a = hardlims - + b)
Figure P5.1 Single-Neuron Perceptron

To be a vector space the boundary must satisfy the ten conditions given at
the beginning of this chapter. Condition 1 requires that when we add two
vectors together the sum remains in the vector space. Let p 1 and p 2 be two
vectors on the decision boundary. To be on the boundary they must satisfy
Wp 1 = 0 Wp 2 = 0 .
If we add these two equations together we find
W p1 + p2 = 0 .
Therefore the sum is also on the decision boundary.
Conditions 2 and 3 are clearly satisfied. Condition 4 requires that the zero
vector be on the boundary. Since W0 = 0 , the zero vector is on the decision
boundary. Condition 5 implies that if p is on the boundary, then p must
also be on the boundary. If p is on the boundary, then
Wp = 0 .
If we multiply both sides of this equation by -1 we find
W p = 0 .
Therefore condition 5 is satisfied.
5-17
Condition 6 will be satisfied if for any p on the boundary ap is also on the

boundary. This can be shown in the same way as condition 5. Just multiply
both sides of the equation by a instead of by 1.
W ap = 0
Conditions 7 through 10 are clearly satisfied. Therefore the perceptron de-
cision boundary is a vector space.
P5.2 Show that the set Y of nonnegative ( f t 0 ) continuous functions

is not a vector space.
This set violates several of the conditions required of a vector space. For ex-
ample, there are no negative vectors, so condition 5 cannot be satisfied. Al-
so, consider condition 6. The function f t = t is a member of Y . Let
a = 2 . Then
a f 2 = 2 2 = 4 0 .
Therefore a f t is not a member of Y , and condition 6 is not satisfied.
P5.3 Which of the following sets of vectors are independent? Find the
dimension of the vector space spanned by each set.
1 1 1
i. 1 0 2
1 1 1

ii. sin t cos t 2 cos t + ---
4
1 1 1
iii. 1 0 2
1 1 1
1 1 1
i. We can solve this problem several ways. First, lets assume that the
vectors are dependent. Then we can write
1 1 1 0
a1 1 + a2 0 + a3 2 = 0 .
1 1 1 0
5-18
Solved Problems
If we can solve for the coefficients and they are not all zero, then the vectors
are dependent. By inspection we can see that if we let a 1 = 2 , a 2 = 1 and
a 3 = 1 , then the equation is satisfied. Therefore the vectors are depen-
dent.
n
Another approach, when we have n vectors in , is to write the above
equation in matrix form:
1 1 1 a1 0
1 0 2 a2 = 0
1 1 1 a3 0
If the matrix in this equation has an inverse, then the solution will require
that all coefficients be zero; therefore the vectors are independent. If the
matrix is singular (has no inverse), then a nonzero set of coefficients will
work, and the vectors are dependent. The test, then, is to create a matrix
using the vectors as columns. If the determinant of the matrix is zero (sin-
gular matrix), then the vectors are dependent; otherwise they are indepen-
dent. Using the Laplace expansion [Brog91] on the first column, the 5
determinant of this matrix is
1 1 1
1 0 2 = 1 0 2 + 1 1 1 + 1 1 1 = 2+0+2 = 0
1 1 1 1 0 2
1 1 1
Therefore the vectors are dependent.

The dimension of the space spanned by the vectors is two, since any two of
the vectors can be shown to be independent.
ii. By using some trigonometric identities we can write

cos t + --- = ------- sin t + ------- cos t .
1 1
4 2 2
Therefore the vectors are dependent. The dimension of the space spanned
by the vectors is two, since no linear combination of sin t and cos t is iden-
tically zero.
iii. This is similar to part (i), except that the number of vectors is less than
4
the size of the vector space they are drawn from (three vectors in ). In
this case the matrix made up of the vectors will not be square, so we will
not be able to compute a determinant. However, we can use something
called the Gramian [Brog91]. It is the determinant of a matrix whose i, j
element is the inner product of vector i and vector j. The vectors are depen-
dent if and only if the Gramian is zero.
5-19
For our problem the Gramian would be
(x 1,x 1) (x 1,x 2) (x 1,x 3)

G = (x 2,x 1) (x 2,x 2) (x 2,x 3) ,
(x 3,x 1) (x 3,x 2) (x 3,x 3)
where
1 1 1
x1 = 1 x2 = 0 x3 = 2 .
1 1 1
1 1 1
Therefore
4 3 5
G = 3 3 3 = 4 3 3 + 3 3 5 + 5 3 5 = 48 18 30 = 0 .
3 7 37 3 3
5 3 7
We can also show that these vectors are dependent by noting
1 1 1 0
2 1 1 0 1 2 = 0 .
1 1 1 0
1 1 1 0
The dimension of the space must therefore be less than 3. We can show that
x 1 and x 2 are independent, since
G = 4 3 = 40.
3 3
Therefore the dimension of the space is 2.
P5.4 Recall from Chapters 3 and 4 that one-layer perceptrons can only
be used to recognize patterns that are linearly separable (can be
separated by a linear boundary see Figure 3.3). If two patterns
are linearly separable, are they always linearly independent?
No, these are two unrelated concepts. Take the following simple example.
Consider the two input perceptron shown in Figure P5.2.
5-20
Solved Problems
Suppose that we want to separate the two vectors

Wp + b = 0
p 1 = 0.5 p 2 = 1.5 .
0.5 1.5
p2
If we choose the weights and offsets to be w 11 = 1 , w 12 = 1 and b = 2 ,
then the decision boundary ( Wp + b = 0 ) is shown in the figure to the left.
p1 Clearly these two vectors are linearly separable. However, they are not lin-
early independent since p 2 = 3p 1 .

- Title - Neuron
p1 w11
n a

SxR
w12 b
p2
1
5
- Exp -(Wp + b)
a = hardlims
Figure P5.2 Two-Input Perceptron
P5.5 Using the following basis vectors, find an orthogonal set using
Gram-Schmidt orthogonalization.
1 1 0
y1 = 1 y2 = 0 y3 = 1
1 0 0
Step 1.
1
v1 = y1 = 1
1
5-21
Step 2.
1
1 1 1 0
T
v1 y2 1 0 1 1 13 23
v 2 = y 2 ----------
T
- v 1 = 0 -------------------------
- 1 = 0 1 3 = 13
v1 v1 1
0 1 0 13 1 3
1 1 1 1
1
Step 3.
T T
v1 y3 v2 y3
v 3 = y 3 ----------
T
- v 1 ----------
T
- v2
v1 v1 v2 v2
0 0
1 1 1 1 2 3 1 3 1 3 1
0 0 1 0 23
v 3 = 1 -------------------------- 1 ---------------------------------------------------------------- 1 3
0 1 1 2 3 1 3
1 1 1 1 2 3 1 3 1 3 1 3
1 1 3
0 13 1 3 0
v3 = 1 1 3 1 6 = 1 2
0 13 16 1 2
P5.6 Consider the vector space of all polynomials defined on the inter-
1
val [-1, 1]. Show that (x,y) = x(t) y t dt is a valid inner product.
1
An inner product must satisfy the following properties.

1. (x,y) = (y,x)
1 1
(x,y) = x(t) y t dt = y t x(t)dt = (y,x)

1 1
2. (x,a y 1 + b y 2) = a (x,y 1) + b (x,y 2)
5-22
Solved Problems
1 1 1
(x,a y 1 + b y 2) = x(t) a y1 t + b y2 t dt = a x(t) y 1 t dt + b x(t) y2 t dt

1 1 1
= a (x,y 1) + b (x,y 2)

1 1
2
(x,x) = x t x t dt = x (t) dt 0
1 1
Equality holds here only if x(t) = 0 for 1 t 1 , which is the zero vector.
P5.7 Two vectors from the vector space described in the previous prob-
lem (polynomials defined on the interval [-1, 1]) are 1 + t and 1 t .
Find an orthogonal set of vectors based on these two vectors.
Step 1. 5
v1 = y1 = 1 + t
Step 2.
(v 1,y 2)
v 2 = y 2 ---------------
- v1
(v 1,v 1)
where
1 1
3
1 + t 1 t dt = t ---3 = --- --- = ---

t 2 2 4
(v 1,y 2) =
3 3 3
1
1
1
3 1
1 + t
= --- 0 = --- .
8 8
1 + t dt = -----------------
2
(v 1,v 1) = -
3 3 3
1
1
Therefore
3
v2 = 1 t 4--------- 1 3
- 1 + t = --- --- t .
83 2 2
5-23
T
P5.8 Expand x = 6 9 9 in terms of the following basis set.
1 1 1
v1 = 1 v2 = 2 v3 = 3
1 3 2
The first step is to calculate the reciprocal basis vectors.
5 1 1
--- --- ---
3 3 3
1 1 1
= 1 1 2
1
B = 1 2 3 B --- --- ---
3 3 3
1 3 2
1 2 1
--- --- ---
3 3 3
Therefore taking the rows of B-1,
53 1 3 1 3
r1 = 1 3 r2 = 1 3 r3 = 23 .
1 3 23 1 3
The coefficients in the expansion are calculated
6
x1 = r1 x = 5 1 1
v T
--- ------ ------ 9 = 4
3 3 3
9
6
x 2 = r 2 x = -----1- -----1- 2
v T
--- 9 = 1
3 3 3
9
6
x 3 = r 3 x = -----1- 2 1
v T
--- ------ 9 = 1
3 3 3
9
,
and the expansion is written
5-24
Solved Problems
1 1 1
v v v
x = x1 v1 + x2 v2 + x3 v3 = 4 1 + 1 2 + 1 3 .
1 3 2
We can represent the process in matrix form:
5 1 1
--- --- ---
3 3 3
6 4
x = B x = 1 1 2
v 1
--- --- --- 9 = 1 .
3 3 3
9 1
1 2 1
--- --- ---
3 3 3
v
Recall that both x and x are representations of the same vector, but are
expanded in terms of different basis sets. (It is assumed that x uses the
standard basis set, unless otherwise indicated.)
5
5-25
Epilogue
This chapter has presented a few of the basic concepts of vector spaces, ma-
terial that is critical to the understanding of how neural networks work.
This subject of vector spaces is very large, and we have made no attempt to
cover all its aspects. Instead, we have presented those concepts that we feel
are most relevant to neural networks. The topics covered here will be revis-
ited in almost every chapter that follows.
The next chapter will continue our investigation of the topics of linear al-
gebra most relevant to neural networks. There we will concentrate on lin-
ear transformations and matrices.
5-26
Further Reading
Further Reading
This is a well-written book on the subject of linear systems.
The first half of the book is devoted to linear algebra. It also
has good sections on the solution of linear differential equa-
[Stra76] G. Strang, Linear Algebra and Its Applications, New York:
Academic Press, 1980.
Strang has written a good basic text on linear algebra.
Many applications of linear algebra are integrated into the
text.
5-27
Exercises
E5.1 Consider again the perceptron described in Problem P5.1. If b 0 , show
that the decision boundary is not a vector space.
E5.2 What is the dimension of the vector space described in Problem P5.1?
E5.3 Consider the set of all continuous functions that satisfy the condition
f 0 = 0 . Show that this is a vector space.
E5.4 Show that the set of 2 2 matrices is a vector space.
E5.5 Consider a perceptron network, with the following weights and bias.
W = 1 0 1 , b = 0 .
i. Write out the equation for the decision boundary.

ii. Show that the decision boundary is a vector space. (Demonstrate
that the 10 criteria are satisfied for any point on the boundary.)
iii. What is the dimension of the vector space?
iv. Find a basis set for the vector space.
E5.6 The three parts to this question refer to subsets of the set of real-valued
continuous functions defined on the interval [0,1]. Tell which of these sub-
sets are vector spaces. If the subset is not a vector space, identify which of
the 10 criteria are not satisfied.
i. All functions such that f 0.5 = 2 .
ii. All functions such that f 0.75 = 0 .
iii. All functions such that f 0.5 = f 0.75 3 .
E5.7 The next three questions refer to subsets of the set of real polynomials de-
2
fined over the real line (e.g., 3 + 2t + 6t ). Tell which of these subsets are
vector spaces. If the subset is not a vector space, identify which of the 10
criteria are not satisfied.
i. Polynomials of degree 5 or less.
ii. Polynomials that are positive for positive t.
iii. Polynomials that go to zero as t goes to zero.
5-28
Exercises
E5.8 Which of the following sets of vectors are independent? Find the dimension
of the vector space spanned by each set. (Verify your answers to parts (i)
2+2
and (iv) using the MATLAB function rank.)
ans =
4
1 1 1
i. 2 0 2
3 1 1
ii. sin t cos t cos 2t
iii. 1 + t 1t
1 1 3
iv. 2 0 4
2 0 4
1 1 3
E5.9 Recall the apple and orange pattern recognition problem of Chapter 3. Find
the angles between each of the prototype patterns (orange and apple) and
the test input pattern (oblong orange). Verify that the angles make intui-
tive sense.
1 1 1
p 1 = 1 orange p2 = 1 apple p = 1
1 1 1
E5.10 Using the following basis vectors, find an orthogonal set using Gram-
Schmidt orthogonalization. (Check your answer using MATLAB.)
2+2 1 1 1
ans = y1 = 0 y2 = 1 y3 = 1
4
0 0 1
E5.11 Consider the vector space of all piecewise continuous functions on the in-
terval [0, 1]. The set f 1 f 2 f 3 , which is defined in Figure E15.1, contains
three vectors from this vector space.
i. Show that this set is linearly independent.
ii. Generate an orthogonal set using the Gram-Schmidt procedure. The
inner product is defined to be
5-29
(f,g) = f t g t dt .
0
f1(t) f2(t) f3(t)
1 1 1
1 1
1
-1 -1 -1
Figure E15.1 Basis Set for Exercise E5.11
E5.12 Consider the vector space of all piece wise continuous functions on the in-
terval [0,1]. The set f 1 f 2 , which is defined in Figure E15.2, contains two
vectors from this vector space.
f2(t)
f1(t)
2
1 1
1 1
-1 -1
-
Figure E15.2 Basis Set for Exercise E5.12

i. Generate an orthogonal set using the Gram-Schmidt procedure. The
inner product is defined to be
1
0
5-30
Exercises
g (t)
3
2 h (t)
1 1
1
1
-1
-2
-3
Figure E15.3 Vectors vectors g and h for Exercise E5.12 part ii.
ii. Expand the vectors g and h in Figure E15.3 in terms of the orthog-
onal set you created in Part 1. Explain any problems you find.
E5.13 Consider the set of polynomials of degree 1 or less. This is a linear vector
space. One basis set for this space is
{u 1 = 1, u 2 = t }
Using this basis set, the polynomial y = 2 + 4t can be represented as
y = 2
u
Consider the new basis set
{v 1 = 1 + t , v 2 = 1 t }
Use reciprocal basis vectors to find the representation of y in terms of this

new basis set.
E5.14 A vector x can be expanded in terms of the basis vectors {v 1, v 2} as
x = 1 v1 + 1 v2
The vectors v 1 and v2 can be expanded in terms of the basis vectors {s1, s 2}
as
v1 = 1 s1 1 s2
v2 = 1 s1 + 1 s2
5-31
i. Find the expansion for x in terms of the basis vectors {s 1, s 2} .

ii. A vector y can be expanded in terms of the basis vectors {s 1, s 2} as
y = 1 s1 + 1 s2 .
Find the expansion of y in terms of the basis vectors {v 1, v 2} .
E5.15 Consider the vector space of all continuous functions on the interval [0,1].
The set f 1 f 2 , which is defined in the figure below, contains two vectors
from this vector space.
f1 t = 1 0t1 f2 t = 1 t 0t1
1 1
t t
1 1
Figure E15.4 Independent Vectors for Exercise E5.15

i. From these two vectors, generate an orthogonal set g 1 g 2 using
the Gram-Schmidt procedure. The inner product is defined to be
1
0
Plot the two orthogonal vectors g 1 and g2 as functions of time.

ii. Expand the following vector h in terms of the orthogonal set you
created in part i., using Eq. (5.27). Demonstrate that the expansion
is correct by reproducing h as a combination of g 1 and g 2 .
h t = t 1 0t1
t
1
-1
Figure E15.5 Vector h for Exercise E5.15
E5.16 Consider the set of all complex numbers. This can be considered a vector
space, because it satisfies the ten defining properties. We can also define
5-32
Exercises
an inner product for this vector space x y = Re x Re y + Im x Im y ,

where Re x is the real part of x , and Im x is the imaginary part of x .
This leads to the following definition for norm: x = x y .
i. Consider the following basis set for the vector space described
above: v 1 = 1 + 2j , v 2 = 2 + j . Using the Gram-Schmidt method,
find an orthogonal basis set.
ii. Using your orthogonal basis set from part i., find vector expansions
for u 1 = 1 j , u 2 = 1 + j , and x = 3 + j . This will allow you to write
x , u1 , and u2 as a columns of numbers x , u 1 and u 2 .
iii. We now want to represent the vector x using the basis set u 1 u 2 .
Use reciprocal basis vectors to find the expansion for x in terms of
the basis vectors u 1 u 2 . This will allow you to write x as a new
u
column of numbers x .
iv. Show that the representations for x that you found in parts ii. and
u
iii. are equivalent (the two columns of numbers x and x both rep-
resent the same vector x ).
E5.17 Consider the vectors defined in Figure E15.6. The set s 1 s 2 is the stan-
dard basis set. The set u 1 u 2 is an alternate basis set. The vector x is a
vector that we wish to represent with respect to the two basis sets.
s2
u1
s1
u2
Figure E15.6 Vector Definitions for Exercise E5.17

i. Write the expansion for x in terms of the standard basis s 1 s2 .
ii. Write the expansions for u 1 and u 2 in terms of the standard basis
s 1 s 2
5-33
iii. Using reciprocal basis vectors, write the expansion for x in terms of
the basis u 1 u 2 .
iv. Draw sketches, similar to Figure 5.2, that demonstrate that the ex-
pansions of part i. and part iii. are equivalent.
E5.18 Consider the set of all functions that can be written in the form A sin t + .
This set can be considered a vector space, because it satisfies the ten defin-
ing properties.
i. Consider the following basis set for the vector space described
above: v 1 = sin t , v 2 = cos t . Represent the vector
x = 2 sin t + 4 cos t as a column of numbers x v (find the vector ex-
pansion), using this basis set.
ii. Using your basis set from part i., find vector expansions for
u1 = 2 sin t + cos t , u 2 = 3 sin t .
iii. We now want to represent the vector x of part i., using the basis set
u 1 u 2 . Use reciprocal basis vectors to find the expansion for x in
terms of the basis vectors u 1 u 2 . This will allow you to write x as
u
a new column of numbers x .
iv. Show that the representations for x that you found in parts i. and
v u
iii. are equivalent (the two columns of numbers x and x both rep-
resent the same vector x ).
E5.19 Suppose that we have three vectors: x y z X . We want to add some mul-
tiple of y to x, so that the resulting vector is orthogonal to z.
i. How would you determine the appropriate multiple of y to add to x?
ii. Verify your results in part i. using the following vectors.
x = 1 y = 1 z = 0.5
0 0.5 1
iii. Use a sketch to illustrate your results from part ii.
T
E5.20 Expand x = 1 2 2 in terms of the following basis set. (Verify your an-
swer using MATLAB.)
2+2
ans =
4 1 1 1
v1 = 1 v2 = 1 v3 = 1
0 2 0
5-34
Exercises
12
E5.21 Find the value of a that makes x a y a minimum. (Use x = (x,x) .)
Show that for this value of a the vector z = x a y is orthogonal to y and
that
2 2 2
x ay + ay = x .
(The vector ay is the projection of x on y.) Draw a diagram for the
case where x and y are two-dimensional. Explain how this concept
is related to Gram-Schmidt orthogonalization.
5-35
Objectives
6 Linear Transformations for

Neural Networks
Objectives 6-1
Linear Transformations 6-2
Matrix Representations 6-3
Change of Basis 6-6
Eigenvalues and Eigenvectors 6-10
Diagonalization 6-13
Epilogue 6-29
Exercises 6-31
Objectives
This chapter will continue the work of Chapter 5 in laying out the mathe-
matical foundations for our analysis of neural networks. In Chapter 5 we
reviewed vector spaces; in this chapter we investigate linear transforma-
tions as they apply to neural networks.
As we have seen in previous chapters, the multiplication of an input vector
by a weight matrix is one of the key operations that is performed by neural
networks. This operation is an example of a linear transformation. We
want to investigate general linear transformations and determine their
fundamental characteristics. The concepts covered in this chapter, such as
eigenvalues, eigenvectors and change of basis, will be critical to our under-
standing of such key neural network topics as performance learning (in-
cluding the Widrow-Hoff rule and backpropagation) and Hopfield network
convergence.
6-1
Theory and Examples

Recall the Hopfield network that was discussed in Chapter 3. (See Figure
6.1.) The output of the network is updated synchronously according to the
equation
a t + 1 = satlin Wa t + b . (6.1)
Notice that at each iteration the output of the network is again multiplied
by the weight matrix W. What is the effect of this repeated operation? Can
we determine whether or not the output of the network will converge to
some steady state value, go to infinity, or oscillate? In this chapter we will
lay the foundation for answering these questions, along with many other
questions about neural networks discussed in this book.
Initial
p
Sx1 W n(t + 1) a(t + 1) a(t)
SxS
Sx1 Sx1 D Sx1
1 b
Sx1
S S
Linear Transformations
We begin with some general definitions.
Transformation A transformation consists of three parts:
1. a set of elements X = x i , called the domain,
2. a set of elements Y = y i , called the range, and
3. a rule relating each x i X to an element y i Y .
6-2
Matrix Representations
Linear Transformation A transformation A is linear if:

1. for all x 1 x 2 X , A x 1 + x2 = A x1 + A x2 ,
2. for all x X , a R , A ax = aA x .
A(x ) x Consider, for example, the transformation obtained by rotating vectors in
2
by an angle , as shown in the figure to the left. The next two figures
illustrate that property 1 is satisfied for rotation. They show that if you
want to rotate a sum of two vectors, you can rotate each vector first and
then sum them. The fourth figure illustrates property 2. If you want to ro-
tate a scaled vector, you can rotate it first and then scale it. Therefore ro-
tation is a linear operation.
As we mentioned at the beginning of this chapter, matrix multiplication is
x 1+x 2 an example of a linear transformation. We can also show that any linear
transformation between two finite-dimensional vector spaces can be repre-
x2 sented by a matrix (just as in the last chapter we showed that any general
vector in a finite-dimensional vector space can be represented by a column
of numbers). To show this we will use most of the concepts covered in the
x1 previous chapter.
Let v 1 v 2 v n be a basis for vector space X , and let u 1 u 2 u m be
a basis for vector space Y . This means that for any two vectors x X and
yY
A(x 1 + x 2)
n m
A(x 1)
x = x i v i and y = y i u i . (6.2)
i=1 i=1
A(x 2) Let A be a linear transformation with domain X and range Y ( A :X Y ).

Then
Ax = y (6.3)
can be written
A(ax )
= aA(x ) ax
n

m
A x j v j = yi ui . (6.4)

j=1 i=1
A(x ) x
Since A is a linear operator, Eq. (6.4) can be written
n m
xj A vj = yi ui . (6.5)
j=1 i=1
6-3
Since the vectors A v j are elements of Y , they can be written as linear

combinations of the basis vectors for Y :
m
A v j = a ij u i . (6.6)
i=1
(Note that the notation used for the coefficients of this expansion, a ij , was
not chosen by accident.) If we substitute Eq. (6.6) into Eq. (6.5) we obtain
n m m
xj a ij u i = yi ui . (6.7)
j=1 i=1 i=1
The order of the summations can be reversed, to produce

m n m
ui a ij xj = yi ui . (6.8)
i=1 j=1 i=1
This equation can be rearranged, to obtain

m n

u i a ij x j y i =

0. (6.9)
i=1 j=1
Recall that since the u i form a basis set they must be independent. This
means that each coefficient that multiplies u i in Eq. (6.9) must be identi-
cally zero (see Eq. (5.4)), therefore
n
aij x j = yi . (6.10)
j=1
This is just matrix multiplication, as in
a 11 a 12 a 1n x 1 y1
a 21 a 22 a 2n x 2 y2 (6.11)
=

a m1 a m2 a mn x n ym
We can summarize these results: For any linear transformation between

two finite-dimensional vector spaces there is a matrix representation. When
we multiply the matrix times the vector expansion for the domain vector x ,
we obtain the vector expansion for the transformed vector y .
6-4
Keep in mind that the matrix representation is not unique (just as the rep-
resentation of a general vector by a column of numbers is not unique see
Chapter 5). If we change the basis set for the domain or for the range, the
matrix representation will also change. We will use this fact to our advan-
tage in later chapters.
2 As an example of a matrix representation, consider the rotation transfor-
+2
mation. Lets find a matrix representation for that transformation. The key
step is given in Eq. (6.6). We must transform each basis vector for the do-
main and then expand it in terms of the basis vectors of the range. In this
2
example the domain and the range are the same ( X = Y = ), so to keep
s2 A(x) things simple we will use the standard basis for both ( u i = v i = s i ), as
shown in the adjacent figure.
The first step is to transform the first basis vector and expand the resulting
x transformed vector in terms of the basis vectors. If we rotate s 1 counter-
clockwise by the angle we obtain
2
s1
A s 1 = cos s1 + sin s 2 = a i1 s i = a 11 s 1 + a 21 s 2 , (6.12)
i=1
as can be seen in the middle left figure. The two coefficients in this expan-
sion make up the first column of the matrix representation.
cos() A(s1) The next step is to transform the second basis vector. If we rotate s 2 coun-
terclockwise by the angle we obtain
sin()
2
A s2 = sin s 1 + cos s 2 = a i2 s i = a 12 s 1 + a 22 s 2 , (6.13)

s1
i=1
as can be seen in the lower left figure. From this expansion we obtain the
second column of the matrix representation. The complete matrix repre-
s2 sentation is thus given by
A(s2) -sin()
A = cos sin . (6.14)

cos() sin cos
Verify for yourself that when you multiply a vector by the matrix of Eq.
(6.14), the vector is rotated by an angle .
In summary, to obtain the matrix representation of a transformation we
use Eq. (6.6). We transform each basis vector for the domain and expand it
in terms of the basis vectors of the range. The coefficients of each expansion
produce one column of the matrix.
6-5
To graphically investigate the process of creating a matrix representation,

use the Neural Network Design Demonstration Linear Transformations
(nnd6lt).
Change of Basis
We notice from the previous section that the matrix representation of a lin-
ear transformation is not unique. The representation will depend on what
basis sets are used for the domain and the range of the transformation. In
this section we will illustrate exactly how a matrix representation changes
as the basis sets are changed.
Consider a linear transformation A :X Y . Let v 1 v 2 v n be a basis
for vector space X , and let u 1 u 2 u m be a basis for vector space Y.
Therefore, any vector x X can be written
n
x = xi vi , (6.15)
i=1
and any vector y Y can be written

m
y = yi ui . (6.16)
i=1
So if
Ax = y (6.17)
the matrix representation will be
a 11 a 12 a 1n x 1 y1
a 21 a 22 a 2n x 2 y2
=

a m1 a m2 a mn x n ym
, (6.18)
or
Ax = y . (6.19)
Now suppose that we use different basis sets for X and Y . Let
t 1 t 2 t n be the new basis for X , and let w 1 w 2 w m be the new
basis for Y . With the new basis sets, the vector x X is written
6-6
Change of Basis
x = x' i t i , (6.20)
i=1
and the vector y Y is written

m
y = y' i w i . (6.21)
i=1
This produces a new matrix representation:
a' 11 a' 12 a' 1n x' 1 y' 1

a' 21 a' 22 a' 2n x' 2 y' 2
= , (6.22)

a' m1 a' m2 a' mn x' n y' m
or
A'x' = y' . (6.23)
What is the relationship between A and A' ? To find out, we need to find
the relationship between the two basis sets. First, since each t i is an ele- 6
ment of X , they can be expanded in terms of the original basis for X :
n
t i = t ji v j . (6.24)
j=1
Next, since each w i is an element of Y , they can be expanded in terms of

the original basis for Y :
m
w i = w ji u j . (6.25)
j=1
Therefore, the basis vectors can be written as columns of numbers:
t 1i w 1i
t 2i w 2i . (6.26)
ti = wi =

t ni w mi
Define a matrix whose columns are the ti:
6-7
Bt = t1 t2 tn . (6.27)
Then we can write Eq. (6.20) in matrix form:
x = x' 1 t 1 + x' 2 t 2 + + x' n t n = B t x' . (6.28)
This equation demonstrates the relationships between the two different

representations for the vector x . (Note that this is effectively the same as
Eq. (5.43). You may want to revisit our discussion of reciprocal basis vec-
tors in Chapter 5.)
Now define a matrix whose columns are the wi:
Bw = w1 w2 wm . (6.29)
This allows us to write Eq. (6.21) in matrix form,
y = B w y' , (6.30)
which then demonstrates the relationships between the two different rep-
resentations for the vector y.
Now substitute Eq. (6.28) and Eq. (6.30) into Eq. (6.19):
AB t x' = B w y' . (6.31)

1
If we multiply both sides of this equation by B w we obtain
1
B w AB t x' = y' . (6.32)
A comparison of Eq. (6.32) and Eq. (6.23) yields the following operation for
Change of Basis a change of basis:
1
A' = B w AB t . (6.33)
This key result, which describes the relationship between any two matrix
Similarity Transform representations of a given linear transformation, is called a similarity
transform [Brog91]. It will be of great use to us in later chapters. It turns
out that with the right choice of basis vectors we can obtain a matrix rep-
resentation that reveals the key characteristics of the linear transforma-
tion it represents. This will be discussed in the next section.
2 As an example of changing basis sets, lets revisit the vector rotation exam-
+2
ple of the previous section. In that section a matrix representation was de-
veloped using the standard basis set {s1, s2} . Now lets find a new
representation using the basis {t1, t2} , which is shown in the adjacent fig-
6-8
Change of Basis
ure. (Note that in this example the same basis set is used for both the do-
main and the range.)
t2 The first step is to expand t 1 and t 2 in terms of the standard basis set, as
s2 in Eq. (6.24) and Eq. (6.25). By inspection of the adjacent figure we find:
t1 t 1 = s 1 + 0.5 s 2 , (6.34)
s1 t2 = s1 + s2 . (6.35)
Therefore we can write
t1 = 1 t2 = 1 . (6.36)
0.5 1
Now we can form the matrix
Bt = t1 t2 = 1 1 , (6.37)
0.5 1
and, because we are using the same basis set for both the domain and the
range of the transformation,
Bw = Bt = 1 1 . (6.38)
0.5 1
We can now compute the new matrix representation from Eq. (6.33):
1
A' = B w AB t = 2 3 2 3 cos sin 1 1
1 3 2 3 sin cos 0.5 1
1 3 sin + cos 4 3 sin
= 5 .
--- sin 1 3 sin + cos
6
(6.39)
Take, for example, the case where = 30.
A' = 1.033 0.667 , (6.40)

0.417 0.699
and
6-9
A = 0.866 0.5 . (6.41)

0.5 0.866
To check that these matrices are correct, lets try a test vector
x = 1 , which corresponds to x' = 1 . (6.42)

0.5 0
(Note that the vector represented by x and x' is t 1 , a member of the second
basis set.) The transformed test vector would be
y = Ax = 0.866 0.5 1 = 0.616 , (6.43)

0.5 0.866 0.5 0.933
which should correspond to
y' = A'x' = 1.033 0.667 1 = 1.033 . (6.44)

0.416 0.699 0 0.416
How can we test to see if y' does correspond to y ? Both should be represen-
tations of the same vector, y , in terms of two different basis sets; y uses
the basis {s1, s2} and y' uses the basis {t1, t2} . In Chapter 5 we used the re-
ciprocal basis vectors to transform from one representation to another (see
Eq. (5.43)). Using that concept we have
t2 y = A( x )
s2 1
t1 = x y' = B 1 y = 1 1 0.616 = 2 3 2 3 0.616 = 1.033 , (6.45)
0.5 1 0.933 1 3 2 3 0.933 0.416
s1 which verifies our previous result. The vectors are displayed in the figure
to the left. Verify graphically that the two representations, y and y' , given
by Eq. (6.43) and Eq. (6.44), are reasonable.
Eigenvalues and Eigenvectors

In this final section we want to discuss two key properties of linear trans-
formations: eigenvalues and eigenvectors. Knowledge of these properties
will allow us to answer some key questions about neural network perfor-
mance, such as the question we posed at the beginning of this chapter, con-
cerning the stability of Hopfield networks.
Eigenvalues Lets first define what we mean by eigenvalues and eigenvectors. Consider
Eigenvectors a linear transformation A :X X . (The domain is the same as the range.)
Those vectors z X that are not equal to zero and those scalars that sat-
isfy
6-10
Az = z (6.46)
are called eigenvectors ( z ) and eigenvalues ( ), respectively. Notice that

the term eigenvector is a little misleading, since it is not really a vector but
a vector space, since if z satisfies Eq. (6.46), then a z will also satisfy it.
Therefore an eigenvector of a given transformation represents a direction,
s2 A(x) such that any vector in that direction, when transformed, will continue to
point in the same direction, but will be scaled by the eigenvalue. As an ex-
ample, consider again the rotation example used in the previous sections.
Is there any vector that, when rotated by 30, continues to point in the
x same direction? No; this is a case where there are no real eigenvalues. (If
we allow complex scalars, then two eigenvalues exist, as we will see later.)
s1 How can we compute the eigenvalues and eigenvectors? Suppose that a ba-
sis has been chosen for the n-dimensional vector space X . Then the matrix
representation for Eq. (6.46) can be written
Az = z , (6.47)
or
A I z = 0 . (6.48)
This means that the columns of A I are dependent, and therefore the
determinant of this matrix must be zero:
A I = 0 . (6.49)
This determinant is an nth-order polynomial. Therefore Eq. (6.49) always

has n roots, some of which may be complex and some of which may be re-
peated.
2 As an example, lets revisit the rotation example. If we use the standard ba-
+2
sis set, the matrix of the transformation is
A = cos sin . (6.50)

sin cos
We can then write Eq. (6.49) as
cos sin = 0, (6.51)

sin cos
or
2 2 2 2
2 cos + cos + sin = 2 cos + 1 = 0 . (6.52)
6-11
The roots of this equation are
1 = cos + j sin 2 = cos j sin . (6.53)
Therefore, as we predicted, this transformation has no real eigenvalues (if

sin 0 ). This means that when any real vector is transformed, it will
point in a new direction.
2 Consider another matrix:
+2
A = 1 1 . (6.54)
0 2
To find the eigenvalues we must solve
1 1 = 0, (6.55)
0 2
or
2
+ 3 + 2 = + 1 + 2 = 0 , (6.56)
and the eigenvalues are
1 = 1 2 = 2 . (6.57)
To find the eigenvectors we must solve Eq. (6.48), which in this example be-
comes
1 1 z = 0 . (6.58)
0 2 0
We will solve this equation twice, once using 1 and once using 2 . Begin-
ning with 1 we have
0 1 z = 0 1 z 11 = 0 (6.59)
1
0 1 0 1 z 21 0
or
z 21 = 0 , no constraint on z 11 . (6.60)
Therefore the first eigenvector will be
6-12
z1 = 1 , (6.61)
0
or any scalar multiple. For the second eigenvector we use 2 :
1 1 z = 1 1 z 12 = 0 , (6.62)
2
00 0 0 z 22 0
or
z 22 = z 12 . (6.63)
Therefore the second eigenvector will be
z2 = 1 , (6.64)
1
or any scalar multiple.
To verify our results we consider the following:
Az 1 = 1 1 1 = 1 = 1 1 = 1 z 1 , (6.65)
0 2 0 0 0
Az 2 = 1 1 1 = 2 = 2 1 = 2 z 2 . (6.66)
0 2 1 2 1
To test your understanding of eigenvectors, use the Neural Network Design

Demonstration Eigenvector Game (nnd6eg).
Diagonalization
Whenever we have n distinct eigenvalues we are guaranteed that we can
find n independent eigenvectors [Brog91]. Therefore the eigenvectors
make up a basis set for the vector space of the transformation. Lets find
the matrix of the previous transformation (Eq. (6.54)) using the eigenvec-
tors as the basis vectors. From Eq. (6.33) we have
A' = B AB = 1 1 1 1 1 1 = 1 0 .
1
(6.67)
0 1 0 2 0 1 0 2
Note that this is a diagonal matrix, with the eigenvalues on the diagonal.
This is not a coincidence. Whenever we have distinct eigenvalues we can
diagonalize the matrix representation by using the eigenvectors as the ba-
6-13
Diagonalization sis vectors. This diagonalization process is summarized in the following.

Let
B = z1 z2 zn , (6.68)
where {z 1 z 2 , z n} are the eigenvectors of a matrix A. Then
1 0 0
1 0 2 0
B AB =

0 0 n
, (6.69)
where { 1 2 , n} are the eigenvalues of the matrix A.
This result will be very helpful as we analyze the performance of several
neural networks in later chapters.
6-14
Summary of Results
Summary of Results
Transformations
A transformation consists of three parts:
1. a set of elements X = x i , called the domain,
2. a set of elements Y = y i , called the range, and
3. a rule relating each x i X to an element y i Y .
Linear Transformations
A transformation A is linear if:
1. for all x 1 x 2 X , A x 1 + x2 = A x1 + A x2 ,
2. for all x X , a R , A ax = aA x .
Let v 1 v 2 v n be a basis for vector space X , and let u 1 u 2 u m be
a basis for vector space Y . Let A be a linear transformation with domain
6
X and range Y :
Ax = y .
The coefficients of the matrix representation are obtained from
m
A v j = aij u i .
i=1
Change of Basis
Bt = t1 t2 tn
Bw = w1 w2 wm
1
A' = B w AB t
6-15

Az = z
A I = 0
Diagonalization
B = z1 z2 zn ,
where {z 1 z 2 , z n} are the eigenvectors of a square matrix A .
1 0 0
1 0 2 0
B AB =

0 0 n
6-16
Solved Problems
Solved Problems
P6.1 Consider the single-layer network shown in Figure P6.1, which has
a linear transfer function. Is the transformation from the input
vector to the output vector a linear transformation?
Inputs Linear Layer

- Title -
p
Rx1 W
SxR n a
Sx1 Sx1
1 b
R Sx1 S
a = purelin (Wp + b)

The network equation is
a = A p = Wp + b .
In order for this transformation to be linear it must satisfy
1. A p1 + p2 = A p1 + A p2 ,
2. A ap = a A p .
Lets test condition 1 first.
A p 1 + p 2 = W p 1 + p 2 + b = Wp 1 + Wp 2 + b .
Compare this with
A p 1 + A p 2 = Wp1 + b + Wp 2 + b = Wp 1 + Wp 2 + 2b .
Clearly these two expressions will be equal only if b = 0 . Therefore this
network performs a nonlinear transformation, even though it has a linear
transfer function. This particular type of nonlinearity is called an affine
transformation.
6-17
P6.2 We discussed projections in Chapter 5. Is a projection a linear

transformation?
The projection of a vector x onto a vector v is computed as
(x,v)
y = A x = ----------- v ,
(v,v)
where (x,v) is the inner product of x with v .

We need to check to see if this transformation satisfies the two conditions
for linearity. Lets start with condition 1:
(x 1 + x 2,v) (x 1,v) + (x 2,v) (x 1,v) (x 2,v)

A x1 + x2 = ------------------------
- v = ---------------------------------
- v = -------------
- v + -------------
-v
(v,v) (v,v) (v,v) (v,v)
= A x 1 + A x2 .
(Here we used linearity properties of inner products.) Checking condition 2:
(a x,v) a (x,v)
A a x = --------------- v = --------------- v = a A x .
(v,v) (v,v)
Therefore projection is a linear operation.
P6.3 Consider the transformation A created by reflecting a vector x in

2
about the line x 1 + x 2 = 0 , as shown in Figure P6.2. Find the ma-
2
trix of this transformation relative to the standard basis in .
s2
s1
A(x)
Figure P6.2 Reflection Transformation
The key to finding the matrix of a transformation is given in Eq. (6.6):
6-18
Solved Problems
A v j = aij u i .
i=1
We need to transform each basis vector of the domain and then expand the
result in terms of the basis vectors for the range. Each time we do the ex-
s1 pansion we get one column of the matrix representation. In this case the
basis set for both the domain and the range is {s 1, s 2} . So lets transform s 1
first. If we reflect s 1 about the line x 1 + x 2 = 0 , we find
2
A(s1) = -s2
A s1 = s 2 = ai1 s i = a 11 s 1 + a21 s 2 = 0 s 1 + 1 s 2
i=1
(as shown in the top left figure), which gives us the first column of the ma-
trix. Next we transform s 2 :
s2 2
A s 2 = s1 = a i2 s i = a 12 s 1 + a 22 s 2 = 1 s 1 + 0 s 2
i=1
(as shown in the second figure on the left), which gives us the second col-
A(s2) = -s1 umn of the matrix. The final result is
0 1 .
1 0
T
Lets test our result by transforming the vector x = 1 1 :
Ax = 0 1 1 = 1 .
1 0 1 1
This is indeed the reflection of x about the line x 1 + x 2 = 0 , as we can see

in Figure P6.3.
6-19
s2 x
s1
A(x )
Figure P6.3 Test of Reflection Operation

(Can you guess the eigenvalues and eigenvectors of this transformation?
Use the Neural Network Design Demonstration Linear Transformations
(nnd6lt) to investigate this graphically. Compute the eigenvalues and
eigenvectors, using the MATLAB function eig, and check your guess.)
P6.4 Consider the space of complex numbers. Let this be the vector
space X , and let the basis for X be {1 + j, 1 j} . Let A :X X be the
conjugation operator (i.e., A x = x ).
i. Find the matrix of the transformation A relative to the basis
set given above.
ii. Find the eigenvalues and eigenvectors of the transforma-
tion.
iii. Find the matrix representation for A relative to the eigen-
vectors as the basis vectors.
i. To find the matrix of the transformation, transform each of the basis
vectors (by finding their conjugate):
A v1 = A 1 + j = 1 j = v 2 = a 11 v 1 + a 21 v 2 = 0 v 1 + 1 v2 ,
A v 2 = A 1 j = 1 + j = v 1 = a 12 v 1 + a 22 v 2 = 1 v 1 + 0 v2 .
This gives us the matrix representation
A = 0 1 .
1 0
ii. To find the eigenvalues, we need to use Eq. (6.49):
6-20
Solved Problems
A I = 1 2
= 1 = 1 + 1 = 0 .
1
So the eigenvalues are: 1 = 1 , 2 = 1 . To find the eigenvectors, use Eq.

(6.48):
A I z = 1 z = 0 .
1 0
For = 1 = 1 this gives us
1 1 z = 1 1 z 11 = 0 ,
1
1 1 1 1 z 21 0
or
z 11 = z 21 .
z1 = 1 ,
1
or any scalar multiple. For the second eigenvector we use = 2 = 1 :
1 1 z = 1 1 z 12 = 0 ,
1
11 1 1 z 22 0
or
z 12 = z 22 .
Therefore the second eigenvector is
z2 = 1 ,
1

Note that while these eigenvectors can be represented as columns of num-
bers, in reality they are complex numbers. For example:
z1 = 1 v1 + 1 v2 = 1 + j + 1 j = 2 ,
6-21
z 2 = 1 v1 + 1 v 2 = 1 + j 1 j = 2j .
Checking that these are indeed eigenvectors:
A z 1 = 2 = 2 = 1 z 1 ,
A z 2 = 2j = 2j = 2 z 2 .
iii. To perform a change of basis we need to use Eq. (6.33):
1 1
A' = B w AB t = B AB ,
where
B = z1 z2 = 1 1 .
1 1
(We are using the same basis set for the range and the domain.) Therefore
we have
1 0
A' = 0.5 0.5 0 1 1 1 = 1 0 = .
0.5 0.5 1 0 1 1 0 1 0 2
As expected from Eq. (6.69), we have diagonalized the matrix representa-

tion.
P6.5 Diagonalize the following matrix:
A = 2 2 .
1 3
The first step is to find the eigenvalues:
A I = 2 2 2
= 5 + 4 = 1 4 = 0 ,
1 3
so the eigenvalues are 1 = 1 , 2 = 4 . To find the eigenvectors,
A I z = 2 2 z = 0 .
1 3 0
For = 1 = 1
6-22
Solved Problems
1 2 z = 1 2 z 11 = 0 ,
1
1 2 1 2 z 21 0
or
z 11 = 2z 21 .
z1 = 2 ,
1

For = 2 = 4
2 2 z = 2 2 z 12 = 0 ,
1
1 1 1 1 z 22 0
or
z 12 = z 22 .
Therefore the second eigenvector will be
z2 = 1 ,
1

To diagonalize the matrix we use Eq. (6.69):
1
A' = B AB ,
where
B = z1 z2 = 2 1 .
1 1
Therefore we have
6-23
1 1
--- ---
A' = 3 3 2 2 2 1 = 1 0 = 1 0 .
1 2 1 3 1 1 0 4 0 2
--- ---
3 3
3 2
P6.6 Consider a transformation A :R R whose matrix representation
relative to the standard basis sets is
A = 3 1 0 .
0 0 1
Find the matrix for this transformation relative to the basis sets:
2 0 0

T = 0 1 2 W = 1 0 .
0 2
1 0 3
The first step is to form the matrices
2 0 0
Bt = 0 1 2 Bw = 1 0 .
0 2
1 0 3
Now we use Eq. (6.33) to form the new matrix representation:

1
A' = B w AB t ,
1 0 2 0 0 6 1 2
3 1 0
A' = 1 0 1 2 = 1 3 .
0 --- 0 0 1 --- 0 ---
2 1 0 3 2 2
Therefore this is the matrix of the transformation with respect to the basis
sets T and W .
P6.7 Consider a transformation A : 2 2 . One basis set for 2 is given

as V = v 1 v 2 .
i. Find the matrix of the transformation A relative to the basis
set V if it is given that
6-24
Solved Problems
A v1 = v1 + 2 v2 ,
A v2 = v1 + v2 .
ii. Consider a new basis set W = w 1 w 2 . Find the matrix of

the transformation A relative to the basis set W if it is given
that
w1 = v1 + v2 ,
w2 = v1 v2 .
i. Each of the two equations gives us one column of the matrix, as defined
in Eq. (6.6). Therefore the matrix is
A = 1 1 .
2 1
ii. We can represent the W basis vectors as columns of numbers in terms

of the V basis vectors:
w1 = 1 w2 = 1 .
1 1
We can now form the basis matrix that we need to perform the similarity
transform:
Bw = 1 1 .
1 1
The new matrix representation can then be obtained from Eq. (6.33):
1
A' = B w AB w ,
1 1 5 1
--- --- --- ---
A' = 2 2 1 1 1 1 = 2 2 .
1 1 2 1 1 1 1 1
--- --- --- ---
2 2 2 2
6-25
2
P6.8 Consider the vector space P of all polynomials of degree less than
2
or equal to 2. One basis for this vector space is V = 1 t t . Consid-
er the differentiation transformation D .
i. Find the matrix of this transformation relative to the basis
set V .
ii. Find the eigenvalues and eigenvectors of the transforma-
tion.
i. The first step is to transform each of the basis vectors:
D 1 = 0 = 0 1 + 0 t + 0 t 2 ,
D t = 1 = 1 1 + 0 t + 0 t 2 ,
D t2 = 2t = 0 1 + 2 t + 0 t 2 .
The matrix of the transformation is then given by
0 1 0
D = 0 0 2 .
0 0 0
ii. To find the eigenvalues we must solve
1 0
3
D I = 0 2 = = 0 .
0 0
Therefore all three eigenvalues are zero. To find the eigenvectors we need
to solve
1 0 0
D I z = 0 2 z = 0 .
0 0 0
For = 0 we have
0 1 0 z1 0
0 0 2 z2 = 0 .
0 0 0 z3 0
6-26
Solved Problems
This means that
z2 = z3 = 0 .
Therefore we have a single eigenvector:
1
z = 0 .
0
Therefore the only polynomial whose derivative is a scaled version of itself

is a constant (a zeroth-order polynomial).
2 2
P6.9 Consider a transformation A :R R . Two examples of trans-
formed vectors are given in Figure P6.4. Find the matrix represen-
tation of this transformation relative to the standard basis set.
x1
x2
s2
A(x 1)
s1
A(x 2)
Figure P6.4 Transformation for Problem P6.9

For this problem we do not know how the basis vectors are transformed, so
we cannot use Eq. (6.6) to find the matrix representation. However, we do
know how two vectors are transformed, and we do know how those vectors
can be represented in terms of the standard basis set. From Figure P6.4 we
can write the following equations:
A 2 = 1 , A 1 = 2 .
2 0 1 1
We then put these two equations together to form
A 2 1 = 1 2 .
2 1 0 1
6-27
So that
1 1 3 5
1 --- --- --- ---
A = 1 2 2 1 = 1 2 4 4 = 4 4 .
0 1 2 1 0 1 1 1 1 1
--- --- --- ---
2 2 2 2
This is the matrix representation of the transformation with respect to the

standard basis set.
This procedure is used in the Neural Network Design Demonstration Linear
Transformations (nnd6lt).
6-28
Epilogue
Epilogue
In this chapter we have reviewed those properties of linear transforma-
tions and matrices that are most important to our study of neural net-
works. The concepts of eigenvalues, eigenvectors, change of basis
(similarity transformation) and diagonalization will be used again and
again throughout the remainder of this text. Without this linear algebra
background our study of neural networks could only be superficial.
In the next chapter we will use linear algebra to analyze the operation of
one of the first neural network training algorithms the Hebb rule.
6-29
Further Reading
Academic Press, 1980.
Strang has written a good basic text on linear algebra.
Many applications of linear algebra are integrated into the
text.
6-30
Exercises
Exercises
E6.1 Is the operation of transposing a matrix a linear transformation?
E6.2 Consider again the neural network shown in Figure P6.1. Show that if the
bias vector b is equal to zero then the network performs a linear operation.
E6.3 Consider the linear transformation illustrated in Figure E6.1.

i. Find the matrix representation of this transformation relative to
the standard basis set.
ii. Find the matrix of this transformation relative to the basis set
v 1 v 2 .
v2
A(v1)
s2 v1
s1
A(v2)
Figure E6.1 Example Transformation for Exercise E6.3
E6.4 Consider the space of complex numbers. Let this be the vector space X , and
let the basis for X be {1 + j, 1 j} . Let A :X X be the operation of multi-
plication by 1 + j (i.e., A x = 1 + j x ).
i. Find the matrix of the transformation A relative to the basis set
given above.
ii. Find the eigenvalues and eigenvectors of the transformation.
iii. Find the matrix representation for A relative to the eigenvectors as
the basis vectors.
2+2
ans = iv. Check your answers to parts (ii) and (iii) using MATLAB.
4
6-31
2 3
E6.5 Consider a transformation A :P P , from the space of second-order poly-
nomials to the space of third-order polynomials, which is defined by the fol-
lowing:
x = a0 + a1 t + a2 t2 ,
A x = a0 t + 1 + a1 t + 1 2 + a2 t + 1 3 .
Find the matrix representation of this transformation relative to the basis
2 2 3 2 3
sets V = 1 t t V = 1 t t t .
E6.6 Consider the vector space of polynomials of degree two or less. These poly-
2
nomials have the form f t = a 0 + a 1 t + a 2 t . Now consider the transforma-
tion in which the variable t is replaced by t + 1 . (for example,
2 2 2
t + 2t + 3 t + 1 + 2 t + 1 + 3 = t + 4t + 6 )
i. Find the matrix of this transformation with respect to the basis set
2
1 t 1 t .
ii. Find the eigenvalues and eigenvectors of the transformation. Show
the eigenvectors as columns of numbers and as functions of time
(polynomials).
E6.7 Consider the space of functions of the form sin t + . One basis set for
this space is V = sin t cos t . Consider the differentiation transformation
D.
i. Find the matrix of the transformation D relative to the basis set V .
ii. Find the eigenvalues and eigenvectors of the transformation. Show
the eigenvectors as columns of numbers and as functions of t .
iii. Find the matrix of the transformation relative to the eigenvectors as
basis vectors.
2t
E6.8 Consider the vector space of functions of the form + e . One basis set
2t 2t
for this vector space is V = 1 + e 1 e . Consider the differentiation
transformation D .
i. Find the matrix of the transformation D relative to the basis set V ,
using Eq. (6.6).
2t
ii. Verify the operation of the matrix on the function 2e .
iii. Find the eigenvalues and eigenvectors of the transformation. Show
the eigenvectors as columns of numbers (with respect to the basis
set V ) and as functions of t .
6-32
Exercises
iv. Find the matrix of the transformation relative to the eigenvectors as

basis vectors.
E6.9 Consider the set of all 2x2 matrices. This set is a vector space, which we will
call X (yes, matrices can be vectors). If M is an element of this vector space,
T
define the transformation A :X X , such that A M = M + M . Consider
the following basis set for the vector space X.
v1 = 1 0 , v2 = 0 1 , v3 = 0 0 , v4 = 0 0
0 0 0 0 1 0 0 1
i. Find the matrix representation of the transformation A relative to

the basis set v 1 v 2 v 3 v 4 (for both domain and range) (using Eq.
(6.6)).
ii. Verify the operation of the matrix representation from part i. on the
element of X given below. (Verify that the matrix multiplication pro-
duces the same result as the transformation.)
12
01
iii. Find the eigenvalues and eigenvectors of the transformation. You do

not need to use the matrix representation that you found in part i.
You can find the eigenvalues and eigenvectors directly from the def-
inition of the transformation. Your eigenvectors should be 2x2 ma-
trices (elements of the vector space X). This does not require much
computation. Use the definition of eigenvector in Eq. (6.46).
1 2
E6.10 Consider a transformation A :P P , from the space of first degree poly-
nomials into the space of second degree polynomials. The transformation is
defined as follows
A a + bt = at + b--- t 2
2
2 1
(e.g., A 2 + 6t = 2t + 3t ). One basis set for P is U = 1 t . One basis for
2 2
P is V = 1 t t .
the basis sets U and V , using Eq. (6.6).
ii. Verify the operation of the matrix on the polynomial 6 + 8t . (Verify
that the matrix multiplication produces the same result as the
transformation.)
6-33
iii. Using a similarity transform, find the matrix of the transformation

with respect to the basis sets S = 1 + t 1 t and V .
E6.11 Let D be the differentiation operator ( D f = df dt ), and use the basis set
5t 5t
u 1 u 2 = e te
for both the domain and the range of the transformation D.

i. Show that the transformation D is linear.,
ii. Find the matrix of this transformation relative to the basis shown
above.
iii. Find the eigenvalues and eigenvectors of the transformation D.
E6.12 A certain linear transformation has the following eigenvalues and eigen-
vectors (represented in terms of the standard basis set):
1 = 1 , z = 1 = 2
z1 = 1 2 2
2 2
i. Find the matrix representation of the transformation, relative to

ii. Find the matrix representation of the transformation relative to the
eigenvectors as the basis vectors.
2 2
E6.13 Consider a transformation A : . In the figure below, we show a set
of basis vectors V = v 1 v 2 and the transformed basis vectors.
A(v1) v1 v2
A(v2)
Figure E6.2 Definition of Transformation for Exercise E6.13
6-34
Exercises
i. Find the matrix representation of this transformation with respect

to the basis vectors V = v 1 v 2 .
ii. Find the matrix representation of this transformation with respect
to the standard basis vectors.
iii. Find the eigenvalues and eigenvectors of this transformation.
Sketch the eigenvectors and their transformations.
iv. Find the matrix representation of this transformation with respect
to the eigenvectors as the basis vectors.
2 3
E6.14 Consider the vector spaces P and P of second-order and third-order poly-
nomials. Find the matrix representation of the integration transformation
I :P 2 P 3 , relative to the basis sets V 2 = 1 t t 2 V 3 = 1 t t2 t 3 .
2 2
E6.15 A certain linear transformation A : has a matrix representation
relative to the standard basis set of
A = 1 2 .
3 4
Find the matrix representation of this transformation relative to the new

basis set: 6

V = 1 2 .
3 5
E6.16 We know that a certain linear transformation A :R 2 R 2 has eigenvalues

and eigenvectors given by
1 = 1 z1 = 1 2 = 2 z2 = 1 .
1 2
(The eigenvectors are represented relative to the standard basis

set.)
ii. Find the matrix representation relative to the new basis
6-35

V = 1 1 .
1 1
E6.17 Consider the transformation A created by projecting a vector x onto the line
shown in Figure E6.3. An example of the transformation is shown in the
figure.
i. Using Eq. (6.6), find the matrix representation of this transforma-
tion relative to the standard basis set s 1 s 2 .
ii. Using your answer to part i, find the matrix representation of this
transformation relative to the basis set v 1 v 2 shown in Figure
E6.3.
iii. What are the eigenvalues and eigenvectors of this transformation?
Sketch the eigenvectors and their transformations.
v1 x
s2
y =A(x )
s1
v2
Figure E6.3 Definition of Transformation for Exercise E6.17
2
E6.18 Consider the following basis set for :
6-36
Exercises

V = v 1 v 2 = 1 1 .
1 2
(The basis vectors are represented relative to the standard basis set.)
i. Find the reciprocal basis vectors for this basis set.
2 2
ii. Consider a transformation A : . The matrix representation
2
for A relative to the standard basis in is
A = 0 1 .
2 3
Find the expansion of Av 1 in terms of the basis set V . (Use the re-
ciprocal basis vectors.)
iii. Find the expansion of Av 2 in terms of the basis set V .
iv. Find the matrix representation for A relative to the basis V . (This
step should require no further computation.)
6-37
Objectives
7 Supervised Hebbian Learning
Objectives 7-1
Linear Associator 7-3
The Hebb Rule 7-4
Performance Analysis 7-5
Pseudoinverse Rule 7-7
Application 7-10
Variations of Hebbian Learning 7-12
Epilogue 7-29
Exercises 7-31
Objectives
The Hebb rule was one of the first neural network learning laws. It was
proposed by Donald Hebb in 1949 as a possible mechanism for synaptic
modification in the brain and since then has been used to train artificial
neural networks.
In this chapter we will use the linear algebra concepts of the previous two
chapters to explain why Hebbian learning works. We will also show how
the Hebb rule can be used to train neural networks for pattern recognition.
7-1
Theory and Examples

Donald O. Hebb was born in Chester, Nova Scotia, just after the turn of the
century. He originally planned to become a novelist, and obtained a degree
in English from Dalhousie University in Halifax in 1925. Since every first-
rate novelist needs to have a good understanding of human nature, he be-
gan to study Freud after graduation and became interested in psychology.
He then pursued a masters degree in psychology at McGill University,
where he wrote a thesis on Pavlovian conditioning. He received his Ph.D.
from Harvard in 1936, where his dissertation investigated the effects of
early experience on the vision of rats. Later he joined the Montreal Neuro-
logical Institute, where he studied the extent of intellectual changes in
brain surgery patients. In 1942 he moved to the Yerkes Laboratories of Pri-
mate Biology in Florida, where he studied chimpanzee behavior.
In 1949 Hebb summarized his two decades of research in The Organization
of Behavior [Hebb49]. The main premise of this book was that behavior
could be explained by the action of neurons. This was in marked contrast
to the behaviorist school of psychology (with proponents such as B. F. Skin-
ner), which emphasized the correlation between stimulus and response and
discouraged the use of any physiological hypotheses. It was a confrontation
between a top-down philosophy and a bottom-up philosophy. Hebb stated
his approach: The method then calls for learning as much as one can about
what the parts of the brain do (primarily the physiologists field), and relat-
ing the behavior as far as possible to this knowledge (primarily for the psy-
chologist); then seeing what further information is to be had about how the
total brain works, from the discrepancy between (1) actual behavior and (2)
the behavior that would be predicted from adding up what is known about
the action of the various parts.
The most famous idea contained in The Organization of Behavior was the
postulate that came to be known as Hebbian learning:
Hebbs Postulate When an axon of cell A is near enough to excite a cell B and repeatedly or
persistently takes part in firing it, some growth process or metabolic change
takes place in one or both cells such that As efficiency, as one of the cells fir-
ing B, is increased.
This postulate suggested a physical mechanism for learning at the cellular
level. Although Hebb never claimed to have firm physiological evidence for
his theory, subsequent research has shown that some cells do exhibit Heb-
bian learning. Hebbs theories continue to influence current research in
neuroscience.
As with most historic ideas, Hebbs postulate was not completely new, as
he himself emphasized. It had been foreshadowed by several others, includ-
ing Freud. Consider, for example, the following principle of association
stated by psychologist and philosopher William James in 1890: When two
7-2
Linear Associator
brain processes are active together or in immediate succession, one of

them, on reoccurring tends to propagate its excitement into the other.
Linear Associator
Hebbs learning law can be used in combination with a variety of neural
network architectures. We will use a very simple architecture for our initial
presentation of Hebbian learning. In this way we can concentrate on the
learning law rather than the architecture. The network we will use is the
Linear Associator linear associator, which is shown in Figure 7.1. (This network was intro-
duced independently by James Anderson [Ande72] and Teuvo Kohonen
[Koho72].)
Inputs Linear Layer
p n a
Rx1
W Sx1 Sx1
SxR
R S
a = purelin (Wp)
Figure 7.1 Linear Associator

The output vector a is determined from the input vector p according to:
a = Wp , (7.1)
or
R
ai = w ij p j . (7.2)
j=1
The linear associator is an example of a type of neural network called an

Associative Memory associative memory. The task of an associative memory is to learn Q pairs
of prototype input/output vectors:
{p 1, t 1} {p 2 , t 2} {p Q, t Q} . (7.3)
In other words, if the network receives an input p = p q then it should pro-

duce an output a = t q , for q = 1 2 Q . In addition, if the input is
changed slightly (i.e., p = p q + ) then the output should only be changed
slightly (i.e., a = t q + ).
7-3
The Hebb Rule

How can we interpret Hebbs postulate mathematically, so that we can use
it to train the weight matrix of the linear associator? First, lets rephrase
the postulate: If two neurons on either side of a synapse are activated si-
multaneously, the strength of the synapse will increase. Notice from Eq.
(7.2) that the connection (synapse) between input p j and output a i is the
weight w ij . Therefore Hebbs postulate would imply that if a positive p j
produces a positive a i then w ij should increase. This suggests that one
mathematical interpretation of the postulate could be
The Hebb Rule w ijnew = w ijold + f i a iq g j p jq , (7.4)
where p jq is the jth element of the qth input vector p q ; a iq is the ith ele-
ment of the network output when the qth input vector is presented to the
network; and is a positive constant, called the learning rate. This equa-
tion says that the change in the weight w ij is proportional to a product of
functions of the activities on either side of the synapse. For this chapter we
will simplify Eq. (7.4) to the following form
w ijnew = w ijold + a iq p jq . (7.5)
Note that this expression actually extends Hebbs postulate beyond its
strict interpretation. The change in the weight is proportional to a product
of the activity on either side of the synapse. Therefore, not only do we in-
crease the weight when both p j and a i are positive, but we also increase
the weight when they are both negative. In addition, this implementation
of the Hebb rule will decrease the weight whenever p j and a i have opposite
sign.
The Hebb rule defined in Eq. (7.5) is an unsupervised learning rule. It does
not require any information concerning the target output. In this chapter
we are interested in using the Hebb rule for supervised learning, in which
the target output is known for each input vector. (We will revisit the unsu-
pervised Hebb rule in Chapter 13.) For the supervised Hebb rule we substi-
tute the target output for the actual output. In this way, we are telling the
algorithm what the network should do, rather than what it is currently do-
ing. The resulting equation is
w ijnew = w ijold + t iq p jq , (7.6)
where t iq is the ith element of the qth target vector t q . (We have set the
learning rate to one, for simplicity.)
Notice that Eq. (7.6) can be written in vector notation:
new old T
W = W + tq pq . (7.7)
7-4
The Hebb Rule
If we assume that the weight matrix is initialized to zero and then each of
the Q input/output pairs are applied once to Eq. (7.7), we can write
Q
+ + tQ pQ = tq pq .
T T T T
W = t1 p1 + t2 p2 (7.8)
q=1
This can be represented in matrix form:
T
p1
T
W = t 1 t 2 t Q p 2 = TP ,
T
(7.9)

T
pQ
where
T = t1 t2 tQ P = p1 p2 pQ . (7.10)
Performance Analysis
Lets analyze the performance of Hebbian learning for the linear associa-
tor. First consider the case where the p q vectors are orthonormal (orthog-
onal and unit length). If p k is input to the network, then the network
output can be computed
Q Q
T
tq pq pk .
T
a = Wp k = t q p q p k = (7.11)

q=1 q=1
Since the p q are orthonormal,

T
pq pk = 1 q = k
= 0 q k. (7.12)
Therefore Eq. (7.11) can be rewritten
a = Wp k = t k . (7.13)
The output of the network is equal to the target output. This shows that, if
the input prototype vectors are orthonormal, the Hebb rule will produce the
correct output for each input.
7-5
But what about non-orthogonal prototype vectors? Lets assume that each
p q vector is unit length, but that they are not orthogonal. Then Eq. (7.11)
becomes
Error
tq pq pk .
T
a = Wp k = t k + (7.14)
qk
Because the vectors are not orthogonal, the network will not produce the
correct output. The magnitude of the error will depend on the amount of
correlation between the prototype input patterns.
2 As an example, suppose that the prototype input/output vectors are
+2

0.5 0.5
0.5 t = 1 0.5 1
p1 = 1 p2 = t2 = . (7.15)
0.5 1 0.5 1
0.5 0.5

(Check that the two input vectors are orthonormal.)

The weight matrix would be
W = TP =
T 1 1 0.5 0.5 0.5 0.5 = 1 0 0 1 . (7.16)
1 1 0.5 0.5 0.5 0.5 0 1 1 0
If we test this weight matrix on the two prototype inputs we find
0.5
Wp 1 = 1 0 0 1 0.5 = 1 , (7.17)
0 1 1 0 0.5 1
0.5
and
0.5
Wp 2 = 1 0 0 1 0.5 = 1 . (7.18)
0 1 1 0 0.5 1
0.5
Success!! The outputs of the network are equal to the targets.
7-6
Pseudoinverse Rule
Now lets revisit the apple and orange recognition problem described in
2
+2 Chapter 3. Recall that the prototype inputs were
1 1
p 1 = 1 orange p2 = 1 apple . (7.19)
1 1
(Note that they are not orthogonal.) If we normalize these inputs and
choose as desired outputs -1 and 1, we obtain
0.5774 0.5774

1
p = 0.5774 t 1 = 1 p 2 = 0.5774 t 2 = 1 . (7.20)

0.5774 0.5774
Our weight matrix becomes
W = TP = 1 1 0.5774 0.5774 0.5774 = 0 1.1548 0 .

T
(7.21)
0.5774 0.5774 0.5774
So, if we use our two prototype patterns,
0.5774
Wp 1 = 0 1.1548 0 0.5774 = 0.6668 , (7.22)
0.5774
7
0.5774
Wp 2 = 0 1.1548 0 0.5774 = 0.6668 . (7.23)
0.5774
The outputs are close, but do not quite match the target outputs.
Pseudoinverse Rule
When the prototype input patterns are not orthogonal, the Hebb rule pro-
duces some errors. There are several procedures that can be used to reduce
these errors. In this section we will discuss one of those procedures, the
pseudoinverse rule.
Recall that the task of the linear associator was to produce an output of t q
for an input of p q . In other words,
Wp q = t q q = 1 2 Q . (7.24)
7-7
If it is not possible to choose a weight matrix so that these equations are

exactly satisfied, then we want them to be approximately satisfied. One ap-
proach would be to choose the weight matrix to minimize the following per-
formance index:
Q

2
F W = t q Wp q . (7.25)
q=1
If the prototype input vectors p q are orthonormal and we use the Hebb rule
to find W, then F(W) will be zero. When the input vectors are not orthogo-
nal and we use the Hebb rule, then F(W) will be not be zero, and it is not
clear that F(W) will be minimized. It turns out that the weight matrix that
will minimize F(W) is obtained by using the pseudoinverse matrix, which
we will define next.
First, lets rewrite Eq. (7.24) in matrix form:
WP = T , (7.26)
where
T = t1 t2 tQ P = p1 p2 pQ . (7.27)
Then Eq. (7.25) can be written

2 2
F W = T WP = E , (7.28)
where
E = T WP , (7.29)
and
eij .
2 2
E = (7.30)
i j
Note that F(W) can be made zero if we can solve Eq. (7.26). If the P matrix
has an inverse, the solution is
1
W = TP . (7.31)
However, this is rarely possible. Normally the p q vectors (the columns of

P) will be independent, but R (the dimension of p q ) will be larger than Q
(the number of p q vectors). Therefore, P will not be a square matrix, and
no exact inverse will exist.
7-8
Pseudoinverse Rule
It has been shown [Albe72] that the weight matrix that minimizes Eq.
Pseudoinverse Rule (7.25) is given by the pseudoinverse rule:
+
W = TP , (7.32)
+
where P is the Moore-Penrose pseudoinverse. The pseudoinverse of a real
matrix P is the unique matrix that satisfies
+
PP P = P,
+ + +
P PP = P ,
T
(7.33)
+ +
P P = P P ,
+ + T
PP = PP .
When the number, R, of rows of P is greater than the number of columns,

Q, of P , and the columns of P are independent, then the pseudoinverse can
be computed by
+ T 1 T
P = P P P . (7.34)
2 To test the pseudoinverse rule (Eq. (7.32)), consider again the apple and or-
+2
ange recognition problem. Recall that the input/output prototype vectors
are
1 1

1
p = 1 t1 = 1 p2 = 1 t2 = 1 . (7.35)

1 1
(Note that we do not need to normalize the input vectors when using the
pseudoinverse rule.)
The weight matrix is calculated from Eq. (7.32):
+
1 1
+
W = TP = 1 1 1 1 , (7.36)

1 1
where the pseudoinverse is computed from Eq. (7.34):

1
1 T
P = P P P = 3 1 1 1 1 = 0.25 0.5 0.25 .
+ T
(7.37)
1 3 1 1 1 0.25 0.5 0.25
7-9
This produces the following weight matrix:
W = TP = 1 1 0.25 0.5 0.25 = 0 1 0 .

+
(7.38)
0.25 0.5 0.25
Lets try this matrix on our two prototype patterns.
1
Wp 1 = 0 1 0 1 = 1 (7.39)
1
1
Wp 2 = 0 1 0 1 = 1 (7.40)
1
The network outputs exactly match the desired outputs. Compare this re-
sult with the performance of the Hebb rule. As you can see from Eq. (7.22)
and Eq. (7.23), the Hebbian outputs are only close, while the pseudoinverse
rule produces exact results.
Application
Now lets see how we might use the Hebb rule on a practical, although
greatly oversimplified, pattern recognition problem. For this problem we
will use a special type of associative memory the autoassociative memo-
Autoassociative Memory ry. In an autoassociative memory the desired output vector is equal to the
input vector (i.e., t q = p q ). We will use an autoassociative memory to store
a set of patterns and then to recall these patterns, even when corrupted
patterns are provided as input.
The patterns we want to store are shown to the left. (Since we are designing
an autoassociative memory, these patterns represent the input vectors and
the targets.) They represent the digits {0, 1, 2} displayed in a 6X5 grid. We
need to convert these digits to vectors, which will become the prototype pat-
p1,t1 p2,t2 p3,t3 terns for our network. Each white square will be represented by a -1, and
each dark square will be represented by a 1. Then, to create the input vec-
tors, we will scan each 6X5 grid one column at a time. For example, the first
prototype pattern will be
T
p1 = 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 . (7.41)
The vector p 1 corresponds to the digit 0, p 2 to the digit 1, and p 3 to the

digit 2. Using the Hebb rule, the weight matrix is computed
7-10
Application
T T T
W = p1 p1 + p2 p2 + p3 p3 . (7.42)
(Note that p q replaces t q in Eq. (7.8), since this is autoassociative memory.)

Because there are only two allowable values for the elements of the proto-
type vectors, we will modify the linear associator so that its output ele-
ments can only take on values of -1 or 1. We can do this by replacing
the linear transfer function with a symmetrical hard limit transfer func-
tion. The resulting network is displayed in Figure 7.2.
Inputs Sym. Hard Limit Layer
p n a
30x1
W 30x1 30x1
30x30
30 30
a = hardlims (Wp)
Figure 7.2 Autoassociative Network for Digit Recognition

Now lets investigate the operation of this network. We will provide the net-
work with corrupted versions of the prototype patterns and then check the
network output. In the first test, which is shown in Figure 7.3, the network
is presented with a prototype pattern in which the lower half of the pattern
is occluded. In each case the correct pattern is produced by the network.
Figure 7.3 Recovery of 50% Occluded Patterns

In the next test we remove even more of the prototype patterns. Figure 7.4
illustrates the result of removing the lower two-thirds of each pattern. In
this case only the digit 1 is recovered correctly. The other two patterns
produce results that do not correspond to any of the prototype patterns.
This is a common problem in associative memories. We would like to design
networks so that the number of such spurious patterns would be mini-
mized. We will come back to this topic again in Chapter 18, when we dis-
cuss recurrent associative memories.
7-11
Figure 7.4 Recovery of 67% Occluded Patterns

In our final test we will present the autoassociative network with noisy ver-
sions of the prototype pattern. To create the noisy patterns we will random-
ly change seven elements of each pattern. The results are shown in Figure
7.5. For these examples all of the patterns were correctly recovered.
Figure 7.5 Recovery of Noisy Patterns

To experiment with this type of pattern recognition problem, use the Neural
Network Design Demonstration Supervised Hebb (nnd7sh).
Variations of Hebbian Learning

There have been a number of variations on the basic Hebb rule. In fact,
many of the learning laws that will be discussed in the remainder of this
text have some relationship to the Hebb rule.
One of the problems of the Hebb rule is that it can lead to weight matrices
having very large elements if there are many prototype patterns in the
training set. Consider again the basic rule:
new old T
W = W + tq pq . (7.43)
A positive parameter , called the learning rate, can be used to limit the
amount of increase in the weight matrix elements, if the learning rate is
less than one, as in:
new old T
W = W + t q p q . (7.44)
We can also add a decay term, so that the learning rule behaves like a
smoothing filter, remembering the most recent inputs more clearly:
new old T old old T
W = W + t q p q W = 1 W + t q p q , (7.45)
where is a positive constant less than one. As approaches zero, the

learning law becomes the standard rule. As approaches one, the learning
7-12
law quickly forgets old inputs and remembers only the most recent pat-
terns. This keeps the weight matrix from growing without bound.
The idea of filtering the weight changes and of having an adjustable learn-
ing rate are important ones, and we will discuss them again in Chapters
10, 12, 15, 16, 18 and 19.
If we modify Eq. (7.44) by replacing the desired output with the difference
between the desired output and the actual output, we get another impor-
tant learning rule:
new old T
W = W + t q a q p q . (7.46)
This is sometimes known as the delta rule, since it uses the difference be-
tween desired and actual output. It is also known as the Widrow-Hoff algo-
rithm, after the researchers who introduced it. The delta rule adjusts the
weights so as to minimize the mean square error (see Chapter 10). For this
reason it will produce the same results as the pseudoinverse rule, which
minimizes the sum of squares of errors (see Eq. (7.25)). The advantage of
the delta rule is that it can update the weights after each new input pattern
is presented, whereas the pseudoinverse rule computes the weights in one
step, after all of the input/target pairs are known. This sequential updating
allows the delta rule to adapt to a changing environment. The delta rule
will be discussed in detail in Chapter 10.
The basic Hebb rule will be discussed again, in a different context, in Chap-
ter 13. In the present chapter we have used a supervised form of the Hebb
rule. We have assumed that the desired outputs of the network, t q , are
known, and can be used in the learning rule. In the unsupervised Hebb
rule, which is discussed in Chapter 13, the actual network output is used
instead of the desired network output, as in:
new old T
W = W + a q p q , (7.47)
where a q is the output of the network when p q is given as the input (see
also Eq. (7.5)). This unsupervised form of the Hebb rule, which does not re-
quire knowledge of the desired output, is actually a more direct interpreta-
tion of Hebbs postulate than is the supervised form discussed in this
chapter.
7-13
Summary of Results
Hebbs Postulate
When an axon of cell A is near enough to excite a cell B and repeatedly or
Linear Associator
Inputs Linear Layer
p n a
Rx1
W Sx1 Sx1
SxR
R S
a = purelin (Wp)
The Hebb Rule

w ijnew = w ijold + t qi p qj
W = t1 p1 + t2 p2 + + tQ pQ
T T T
T
p1
T
W = t 1 t 2 t Q p 2 = TP
T

T
pQ
Pseudoinverse Rule
+
W = TP
7-14
Summary of Results
When the number, R , of rows of P is greater than the number of columns,

Q , of P and the columns of P are independent, then the pseudoinverse can
be computed by
+ T 1 T
P = P P P .

Filtered Learning
(See Chapter 14)
new old T
W = 1 W + t q p q
Delta Rule
(See Chapter 10)
new old T
W = W + t q a q p q
Unsupervised Hebb
(See Chapter 13)
new old T
W = W + a q p q
7-15
Solved Problems
P7.1 Consider the linear associator shown in Figure P7.1.
Inputs Linear Layer
p n a
4x1
W 2x1 2x1
2x4
4 2
a = purelin (Wp)

Let the input/output prototype vectors be

1 1
1 t = 1 1 1
p1 = 1 p2 = t2 = .
1 1 1 1
1 1

i. Use the Hebb rule to find the appropriate weight matrix for
this linear associator.
ii. Repeat part (i) using the pseudoinverse rule.
iii. Apply the input p 1 to the linear associator using the weight
matrix of part (i), then using the weight matrix of part (ii).
i. The first step is to create the P and T matrices of Eq. (7.10):
1 1
P = 1
1 , T = 1 1 .
1 1 1 1
1 1
Then the weight matrix can be computed using Eq. (7.9):
7-16
Solved Problems
h
W = TP =
T 1 1 1 1 1 1 = 2 0 0 2 .
1 1 1 1 1 1 0 2 2 0
ii. For the pseudoinverse rule we use Eq. (7.32):

+
W = TP .
Since the number of rows of P , four, is greater than the number of columns
of P , two, and the columns of P are independent, then the pseudoinverse
can be computed by Eq. (7.34):
+ T 1 T
P = P P P .
1
1 1
1
P = 1 1 1 1 1 1
+ 1 1 1 1 = 4 0 1 1 1 1
1 1 1 1 1 1
1 1 1 1 0 4 1 1 1 1
1 1
1 1 1 1 1
--- 0 --- --- --- ---
= 4 1 1 1 1 = 4 4 4 4
1 1 1 1 1 1 1 1 1
0 --- --- --- --- ---
4 4 4 4 4
The weight matrix can now be computed:
1 1 1 1 1 1
--- --- --- --- --- 0 0 ---
p
W = TP = 1 1+4 4 4 4 = 2 2 .
1 1 1 1 1 1 1 1
--- --- --- --- 0 --- --- 0
4 4 4 4 2 2
iii. We now test the two weight matrices.
1
h
W p1 = 2 0 0 2 1 = 4 t
1
0 2 2 0 1 4
1
7-17
1 1 1
--- 0 0 ---
W p1 = 2 2 1 = 1 = t
p
1
1 1 1 1
0 --- --- 0
2 2 1
Why didnt the Hebb rule produce the correct results? Well, consider again
Eq. (7.11). Since p 1 and p 2 are orthogonal (check that they are) this equa-
tion can be written
h T
W p1 = t1 p1 p1 ,
T
but the p 1 vector is not normalized, so p 1 p 1 1 . Therefore the output of
the network will not be t 1 .
The pseudoinverse rule, on the other hand, is guaranteed to minimize
2

2
t q Wp q ,
q=1
which in this case can be made equal to zero.
P7.2 Consider the prototype patterns shown to the left.

i. Are these patterns orthogonal?
p1 p2
ii. Design an autoassociator for these patterns. Use the Hebb
rule.
iii. What response does the network give to the test input pat-
tern, p t , shown to the left?
pt i. The first thing we need to do is to convert the patterns into vectors.
Lets assign any solid square the value 1 and any open square the value -1.
Then to convert from the two-dimensional pattern to a vector we will scan
the pattern column by column. (We could use rows if we wished.) The two
prototype vectors then become:
T T
p1 = 1 1 1 1 1 1 p2 = 1 1 1 1 1 1 .
To test orthogonality we take the inner product of p 1 and p 2 :
7-18
Solved Problems
1
1
p1 p2 = 1 1 1 1 1 1 1 = 0 .
T
1
1
1
Therefore they are orthogonal. (Although they are not normalized since
T T
p 1 p 1 = p 2 p 2 = 6 .)
ii. We will use an autoassociator like the one in Figure 7.2, except that the
number of inputs and outputs to the network will be six. To find the weight
matrix we use the Hebb rule:
T
W = TP ,
where
1 1
1 1
P = T = 1
1 .
1 1
1 1
1 1
Therefore the weight matrix is
1 1 2 0 2 0 2 0
1 1 0 2 0 2 0 2
W = TP = 1
1 1 1 1 1 1 1 = 2 0 2 0 2 0 .
T
1 1 1 1 1 1 1 1 0 2 0 2 0 2
1 1 2 0 2 0 2 0
1 1 0 2 0 2 0 2
iii. To apply the test pattern to the network we convert it to a vector:

T
pt = 1 1 1 1 1 1 .
The network response is then
7-19
2 0 2 0 2 0 1

0 2 0 2 0 2 1
2 0 2 0 2 0 1
a = hardlims Wp t = hardlims
0 2 0 2 0 2 1

2 0 2 0 2 0 1
0 2 0 2 0 2 1
2 1

6 1
2 1 = p .
a = hardlims = 2
6 1

2 1
6 1
Is this a satisfactory response? How would we want the network to respond

to this input pattern? The network should produce the prototype pattern
that is closest to the input pattern. In this case the test input pattern, p t ,
has a Hamming distance of 1 from p 2 , and a distance of 2 from p 1 . There-
fore the network did produce the correct response. (See Chapter 3 for a dis-
cussion of Hamming distance.)
Note that in this example the prototype vectors were not normalized. This
did not cause the same problem with network performance that we saw in
Problem P7.1, because of the hardlims nonlinearity. It forces the network
output to be either 1 or -1. In fact, most of the interesting and useful prop-
erties of neural networks are due to the effects of nonlinearities.
P7.3 Consider an autoassociation problem in which there are three pro-

totype patterns (shown below as p 1 , p 2 , p 3 ). Design autoassociative
networks to recognize these patterns, using both the Hebb rule
and the pseudoinverse rule. Check their performance on the test
pattern p t shown below.
1 1 1 1
1 1 1 1
1 1 1 1
p1 = 1 p2 = 1 p3 = 1 pt = 1
1 1 1 1
1 1 1 1
1 1 1 1
7-20
Solved Problems
This problem is a little tedious to work out by hand, so lets use MATLAB.
2+2 First we create the prototype vectors.
ans =
4
p1=[ 1 1 -1 -1 1 1 1]';
p2=[ 1 1 1 -1 1 -1 1]';
p3=[-1 1 -1 1 1 -1 1]';
P=[p1 p2 p3];
Now we can compute the weight matrix using the Hebb rule.
wh=P*P';
To check the network we create the test vector.

pt=[-1 1 -1 -1 1 -1 1]';
The network response is then calculated.

ah=hardlims(wh*pt);
ah'
ans =
1 1 -1 -1 1 -1 1
Notice that this response does not match any of the prototype vectors. This
is not surprising since the prototype patterns are not orthogonal. Lets try
the pseudoinverse rule.
pseu=inv(P'*P)*P'; 7
wp=P*pseu;
ap=hardlims(wp*pt);
ap'
ans =
-1 1 -1 1 1 -1 1
Note that the network response is equal to p 3 . Is this the correct response?
As usual, we want the response to be the prototype pattern closest to the
input pattern. In this case p t is a Hamming distance of 2 from both p 1 and
p 2 , but only a distance of 1 from p 3 . Therefore the pseudoinverse rule pro-
duces the correct response.
Try other test inputs to see if there are additional cases where the pseudo-
inverse rule produces better results than the Hebb rule.
7-21
P7.4 Consider the three prototype patterns shown to the left.

i. Use the Hebb rule to design a perceptron network that will
p1 p2 p3 recognize these three patterns.
ii. Find the response of the network to the pattern p t shown to
the left. Is the response correct?
pt
i. We can convert the patterns to vectors, as we did in previous problems,
to obtain:
1 1 1 1
p1 = 1
p2 = 1 p3 = 1 pt = 1 .

1 1 1 1
1 1 1 1
We now need to choose the desired output vectors for each prototype input
vector. Since there are three prototype vectors that we need to distinguish,
we will need two elements in the output vector. We can choose the three de-
sired outputs to be:
t1 = 1 t2 = 1 t3 = 1 .
1 1 1
(Note that this choice was arbitrary. Any distinct combination of 1 and -1
could have been chosen for each vector.)
The resulting network is shown in Figure P7.2.
p n a
4x1
W 2x1 2x1
2x4
4 2
a = hardlims (Wp)
Figure P7.2 Perceptron Network for Problem P7.4

The next step is to determine the weight matrix using the Hebb rule.
7-22
Solved Problems
1 1 1 1
W = TP = 1 1 1 3 1 1 1
T
1 1 1 1 =
1 1 1 1 3 1 1
1 1 1 1
ii. The response of the network to the test input pattern is calculated as
follows.
1

a = hardlims Wp t = hardlims 3 1 1 1 1
1 3 1 1 1

1

= hardlims 2 = 1 p 1 .
2 1
So the response of the network indicates that the test input pattern is clos-
est to p 1 . Is this correct? Yes, the Hamming distance to p 1 is 1, while the
distance to p 2 and p 3 is 3.
P7.5 Suppose that we have a linear autoassociator that has been de-
signed for Q orthogonal prototype vectors of length R using the
Hebb rule. The vector elements are either 1 or -1.
i. Show that the Q prototype patterns are eigenvectors of the
weight matrix.
ii. What are the other ( R - Q ) eigenvectors of the weight ma-
7
trix?
i. Suppose the prototype vectors are:
p 1 p 2 p Q .
Since this is an autoassociator, these are both the input vectors and the de-
sired output vectors. Therefore
T = p1 p2 pQ P = p1 p2 pQ .
If we then use the Hebb rule to calculate the weight matrix we find
Q
pq pq ,
T T
W = TP =
q=1
7-23
from Eq. (7.8). Now, if we apply one of the prototype vectors as input to the
network we obtain
Q Q
T
pq pq pk .
T
a = Wp k = p q p q p k =

q=1 q=1
Because the patterns are orthogonal, this reduces to

T
a = pk pk pk .
And since every element of p k must be either -1 or 1, we find that
a = pk R .
To summarize the results:
Wp k = Rp k ,
which implies that p k is an eigenvector of W and R is the corresponding

eigenvalue. Each prototype vector is an eigenvector with the same eigen-
value.
ii. Note that the repeated eigenvalue R has a Q -dimensional eigenspace
associated with it: the subspace spanned by the Q prototype vectors. Now
consider the subspace that is orthogonal to this eigenspace. Every vector in
this subspace should be orthogonal to each prototype vector. The dimension
of the orthogonal subspace will be R - Q . Consider the following arbitrary
basis set for this orthogonal space:
z 1 z 2 z R Q .
If we apply any one of these basis vectors to the network we obtain:

Q Q
T
pq pq zk
T
a = Wz k = p q p q z k = = 0,

q=1 q=1
since each z k is orthogonal to every p q . This implies that each z k is an

eigenvector of W with eigenvalue 0.
To summarize, the weight matrix W has two eigenvalues, R and 0. This
means that any vector in the space spanned by the prototype vectors will
be amplified by R , whereas any vector that is orthogonal to the prototype
vectors will be set to 0. We will revisit this concept when we discuss the per-
formance of the Hopfield network in Chapter 18.
7-24
Solved Problems
P7.6 The networks we have used so far in this chapter have not includ-
ed a bias vector. Consider the problem of designing a perceptron
network (Figure P7.3) to recognize the following patterns:
p1 = 1 p2 = 2 .
1 2

- Title - Layer
p
2x1
W
1x2
n a
1x1 1x1
1 b
2 1x1 1
- Exp (Wp
a = hardlims - + b)

i. Why is a bias required to solve this problem?
ii. Use the pseudoinverse rule to design a network with bias to
solve this problem.
i. Recall from Chapters 3 and 4 that the decision boundary for the percep-
tron network is the line defined by:
Wp + b = 0 .
If there is no bias, then b = 0 and the boundary is defined by:
Wp = 0 ,
which is a line that must pass through the origin. Now consider the two
Wp = 0 vectors, p 1 and p 2 , which are given in this problem. They are shown graph-
ically in the figure to the left, along with an arbitrary decision boundary
p2 that passes through the origin. It is clear that no decision boundary that
passes through the origin could separate these two vectors. Therefore a
bias is required to solve this problem.
p1 ii. To use the pseudoinverse rule (or the Hebb rule) when there is a bias
term, we should treat the bias as another weight, with an input of 1 (as is
shown in all of the network figures). We then augment the input vectors
with a 1 as the last element:
7-25
1 2
p' 1 = 1 p' 2 = 2 .
1 1
Lets choose the desired outputs to be
t1 = 1 t2 = 1 ,
so that
1 2
P = 1 2 , T = 1 1 .
1 1
We now form the pseudoinverse matrix:

1
1 2

1
P = 1 1 1 1 2 1 1 1 = 3 5 1 1 1 = 0.5 0.5 2 .
+

2 2 1 2 2 1 5 9 2 2 1 0.5 0.5 1
1 1
The augmented weight matrix is then computed:
W' = TP = 1 1 0.5 0.5 2 = 1 1 3 .

+
0.5 0.5 1
We can then pull out the standard weight matrix and bias:
W = 1 1 b = 3.
The decision boundary for this weight and bias is shown in the Figure P7.4.
This boundary does separate the two prototype vectors.
7-26
Solved Problems
p2
Wp + b = 0
p1
Figure P7.4 Decision Boundary for Solved Problem P7.6
P7.7 In all of our pattern recognition examples thus far, we have repre-
sented patterns as vectors by using 1 and -1 to represent dark
and light pixels (picture elements), respectively. What if we were
to use 1 and 0 instead? How should the Hebb rule be changed?
First, lets introduce some notation to distinguish the two different repre-
sentations (usually referred to as the bipolar {-1, 1} representation and the
binary {0, 1} representation). The bipolar representation of the prototype
input/output vectors will be denoted
{p 1, t 1} {p 2, t 2} {p Q, t Q} ,
and the binary representation will be denoted
{p' 1, t' 1} {p' 2, t' 2} {p' Q, t' Q} .
The relationship between the two representations is given by:
1 1
p' q = --- p q + --- 1 p q = 2p' q 1 ,
2 2
where 1 is a vector of ones.

Next, we determine the form of the binary associative network. We will use
the network shown in Figure P7.5. It is different than the bipolar associa-
tive network, as shown in Figure 7.2, in two ways. First, it uses the hardlim
nonlinearity rather than hardlims, since the output should be either 0 or
1. Secondly, it uses a bias vector. It requires a bias vector because all binary
vectors will fall into one quadrant of the vector space, so a boundary that
passes through the origin will not always be able to divide the patterns.
(See Problem P7.6.)
The next step is to determine the weight matrix and the bias vector for this
network. If we want the binary network of Figure P7.5 to have the same
7-27
effective response as a bipolar network (as in Figure 7.2), then the net in-
put, n , should be the same for both networks:
W'p' + b = Wp .
Inputs Hard Limit Layer
p'
Rx1 W'
SxR n a
Sx1 Sx1
1 b
R Sx1 S
a = hardlim (Wp'+ b)
Figure P7.5 Binary Associative Network

This will guarantee that whenever the bipolar network produces a 1 the
binary network will produce a 1, and whenever the bipolar network pro-
duces a -1 the binary network will produce a 0.
If we then substitute for p' as a function of p we find:
W' --- p + --- 1 + b = --- W'p + --- W'1 + b = Wp .

1 1 1 1
2 2 2 2
Therefore, to produce the same results as the bipolar network, we should

choose
W' = 2W b = W1 ,
where W is the bipolar weight matrix.
7-28
Epilogue
Epilogue
We had two main objectives for this chapter. First, we wanted to introduce
one of the most influential neural network learning rules: the Hebb rule.
This was one of the first neural learning rules ever proposed, and yet it con-
tinues to influence even the most recent developments in network learning
theory. Second, we wanted to show how the performance of this learning
rule could be explained using the linear algebra concepts discussed in the
two preceding chapters. This is one of the key objectives of this text. We
want to show how certain important mathematical concepts underlie the
operation of all artificial neural networks. We plan to continue to weave to-
gether the mathematical ideas with the neural network applications, and
hope in the process to increase our understanding of both.
We will again revisit the Hebb rule in Chapters 15 and 21. In Chapter 21
we will use the Hebb rule in the design of a recurrent associative memory
network the Hopfield network.
The next two chapters introduce some mathematics that are critical to our
understanding of the two learning laws covered in Chapters 10 and 11.
Those learning laws fall under a subheading called performance learning,
because they attempt to optimize the performance of the network. In order
to understand these performance learning laws, we need to introduce some
basic concepts in optimization. As with the material on the Hebb rule, our
understanding of these topics in optimization will be greatly aided by our
previous work in linear algebra.
7-29
Further Reading
[Albe72] A. Albert, Regression and the Moore-Penrose Pseudoin-
verse, New York: Academic Press, 1972.
Alberts text is the major reference for the theory and basic
properties of the pseudoinverse. Proofs are included for all
major pseudoinverse theorems.
[Ande72] J. Anderson, A simple neural network generating an inter-
active memory, Mathematical Biosciences, vol. 14, pp.
197220, 1972.
[Hebb49] D. O. Hebb, The Organization of Behavior, New York:
Wiley, 1949.
The main premise of this seminal book is that behavior can
be explained by the action of neurons. In it, Hebb proposes
one of the first learning laws, which postulated a mecha-
nism for learning at the cellular level.
pendently.
7-30
Exercises
Exercises
E7.1 Consider the prototype patterns given to the left.
p1 p2 i. Are p 1 and p 2 orthogonal?

ii. Use the Hebb rule to design an autoassociator network for these
patterns.
iii. Test the operation of the network using the test input pattern p t
shown to the left. Does the network perform as you expected? Ex-
pt plain.
E7.2 Repeat Exercise E7.1 using the pseudoinverse rule.
E7.3 Use the Hebb rule to determine the weight matrix for a perceptron network
(shown in Figure E7.1) to recognize the patterns shown to the left.
p1 p2
p n a
6x1
W 1x1 1x1
1x6
6 1
a = hardlims (Wp)
Figure E7.1 Perceptron Network for Exercise E7.3
E7.4 In Problem P7.7 we demonstrated how networks can be trained using the
Hebb rule when the prototype vectors are given in binary (as opposed to bi-
polar) form. Repeat Exercise E7.1 using the binary representation for the
prototype vectors. Show that the response of this binary network is equiv-
alent to the response of the original bipolar network.
E7.5 Show that an autoassociator network will continue to perform if we zero

the diagonal elements of a weight matrix that has been determined by the
Hebb rule. In other words, suppose that the weight matrix is determined
from:
T
W = PP QI ,
7-31
where Q is the number of prototype vectors. (Hint: show that the prototype
vectors continue to be eigenvectors of the new weight matrix.)
E7.6 We have three input/output prototype vector pairs:
1 t = 1 p = 1 t = 1 p = 0 t = 1 .
p1 = 1 2 2 3 3
0 1 1
i. Show that this problem cannot be solved unless the network uses a
bias.
ii. Use the pseudoinverse rule to design a network for these prototype
vectors. Verify that the network correctly transforms the prototype
vectors.
E7.7 Consider the reference patterns and targets given below. We want to use
these data to train a linear associator network.
2 t = 4 t = 2 t =
p1 = 1 26 p2 = 2 26 p3 = 3 26
4 2 2
i. Use the Hebb rule to find the weights of the network.

ii. Find and sketch the decision boundary for the network with the
Hebb rule weights.
iii. Use the pseudo-inverse rule to find the weights of the network. Be-
cause the number, R, of rows of P is less than the number of col-
umns, Q, of P , the pseudoinverse can be computed by
+ T T 1
P = P PP .
iv. Find and sketch the decision boundary for the network with the
pseudo-inverse rule weights.
v. Compare (discuss) the decision boundaries and weights for each of
the methods (Hebb and pseudo-inverse).
E7.8 Consider the three prototype patterns shown in Figure E7.2.

i. Are these patterns orthogonal? Demonstrate.
ii. Use the Hebb rule to determine the weight matrix for a linear au-
toassociator to recognize these patterns.
iii. Draw the network diagram.
7-32
Exercises
iv. Find the eigenvalues and eigenvectors of the weight matrix. (Do not
solve the equation W I = 0 . Use an analysis of the Hebb
rule.)
-1
p1 p2 p3 1
Figure E7.2 Prototype Patterns for Exercise E7.8
E7.9 Suppose that we have the following three reference patterns and their tar-
gets.
3 t = 6 t = 6 t =
p1 = 1 75 p2 = 2 75 p3 = 3 75
6 3 3
i. Draw the network diagram for a linear associator network that

could be trained on these patterns.
ii. Use the Hebb rule to find the weights of the network.
iii. Find and sketch the decision boundary for the network with the
Hebb rule weights. Does the boundary separate the patterns? Dem-
onstrate.
iv. Use the pseudo-inverse rule to find the weights of the network. De-
scribe the difference between this boundary and the Hebb rule
boundary.
E7.10 We have the following input/output pairs:
1 t = 1 t =
p1 = 1 1 p2 = 2 1
1 1
i. Use the Hebb rule to determine the weight matrix for the percep-
tron network shown in Figure E7.3.
ii. Plot the resulting decision boundary. Is this a good decision
boundary? Explain.
iii. Repeat part i. using the Pseudoinverse rule.
iv. Will there be any difference in the operation of the network if the
Pseudoinverse weight matrix is used? Explain.
7-33
a
1x1
p n
2x1
W 1x1
1x2
2 1
a = hardlims(Wp)
Figure E7.3 Network for Exercise E7.10

2+2
ans = E7.11 One question we might ask about the Hebb and pseudoinverse rules is:
4 How many prototype patterns can be stored in one weight matrix? Test this
experimentally using the digit recognition problem that was discussed on
page 7-10. Begin with the digits 0 and 1. Add one digit at a time up to
6, and test how often the correct digit is reconstructed after randomly
changing 2, 4 and 6 pixels.
i. First use the Hebb rule to create the weight matrix for the digits 0
and 1. Then randomly change 2 pixels of each digit and apply the
noisy digits to the network. Repeat this process 10 times, and record
the percentage of times in which the correct pattern (without noise)
is produced at the output of the network. Repeat as 4 and 6 pixels of
each digit are modified. The entire process is then repeated when
the digits 0, 1 and 2 are used. This continues, one digit at a
time, until you test the network when all of the digits 0 through
6 are used. When you have completed all of the tests, you will be
able to plot three curves showing percentage error versus number of
digits stored, one curve each for 2, 4 and 6 pixel errors.
ii. Repeat part (i) using the pseudoinverse rule, and compare the re-
sults of the two rules.
7-34
Objectives
8 Performance Surfaces and

Optimum Points
Objectives 8-1
Taylor Series 8-2
Vector Case 8-4
Directional Derivatives 8-5
Minima 8-7
Necessary Conditions for Optimality 8-9
First-Order Conditions 8-10
Second-Order Conditions 8-11
Quadratic Functions 8-12
Eigensystem of the Hessian 8-13
Epilogue 8-34
Exercises 8-36
Objectives
This chapter lays the foundation for a type of neural network training tech-
nique called performance learning. There are several different classes of
network learning laws, including associative learning (as in the Hebbian
learning of Chapter 7) and competitive learning (which we will discuss in
Chapter 16). Performance learning is another important class of learning
law, in which the network parameters are adjusted to optimize the perfor-
mance of the network. In the next two chapters we will lay the groundwork
for the development of performance learning, which will then be presented
in detail in Chapters 1014. The main objective of the present chapter is to
investigate performance surfaces and to determine conditions for the exist-
ence of minima and maxima of the performance surface. Chapter 9 will fol-
low this up with a discussion of procedures to locate the minima or maxima.
8-1
Theory and Examples

There are several different learning laws that fall under the category of
Performance Learning performance learning. Two of these will be presented in this text. These
learning laws are distinguished by the fact that during training the net-
work parameters (weights and biases) are adjusted in an effort to optimize
the performance of the network.
There are two steps involved in this optimization process. The first step is
to define what we mean by performance. In other words, we must find a
Performance Index quantitative measure of network performance, called the performance in-
dex, which is small when the network performs well and large when the
network performs poorly. In this chapter, and in Chapter 9, we will assume
that the performance index is given. In Chapters 10, 11 and 13 we will dis-
cuss the choice of performance index.
The second step of the optimization process is to search the parameter
space (adjust the network weights and biases) in order to reduce the per-
formance index. In this chapter we will investigate the characteristics of
performance surfaces and set some conditions that will guarantee that a
surface does have a minimum point (the optimum we are searching for).
Thus, in this chapter we will obtain some understanding of what perfor-
mance surfaces look like. Then, in Chapter 9 we will develop procedures for
locating the optimum points.
Taylor Series
Let us say that the performance index that we want to minimize is repre-
sented by F x , where x is the scalar parameter we are adjusting. We will
assume that the performance index is an analytic function, so that all of its
Taylor Series Expansion derivatives exist. Then it can be represented by its Taylor series expansion
about some nominal point x :
d
F x = F x + F x x x
dx x = x
2
1 d
x x +
2
+ --- 2 F x
2 dx (8.1)
x = x
n
1 d
x x +
n
+ ----- n F x
n! d x
x = x
We will use the Taylor series expansion to approximate the performance

2 index, by limiting the expansion to a finite number of terms. For example,
+2
let
8-2
Taylor Series
F x = cos x . (8.2)
The Taylor series expansion for F x about the point x = 0 is
1 2
F x = cos x = cos 0 sin 0 x 0 --- cos 0 x 0
2
1
+ --- sin 0 x 0 +
3
6
1 2 1 4
= 1 --- x + ------ x +
2 24 (8.3)
The zeroth-order approximation of F x (using only the zeroth power of x )
is
F x F0 x = 1 . (8.4)
The second-order approximation is
1 2
F x F 2 x = 1 --- x . (8.5)
2
(Note that in this case the first-order approximation is the same as the ze-
roth-order approximation, since the first derivative is zero.)
The fourth-order approximation is
1 2 1 4
F x F 4 x = 1 --- x + ------ x . (8.6)
2 24
A graph showing F x and these three approximations is shown in Figure

8.1.
1.5
F0 x
1
F4 x
0.5
-0.5
Fx
-1 F2 x
-1.5
-6 -4 -2 0 2 4 6
Figure 8.1 Cosine Function and Taylor Series Approximations
8-3
From the figure we can see that all three approximations are accurate if x
is very close to x = 0 . However, as x moves farther away from x only the
higher-order approximations are accurate. The second-order approxima-
tion is accurate over a wider range than the zeroth-order approximation,
and the fourth-order approximation is accurate over a wider range than the
second-order approximation. An investigation of Eq. (8.1) explains this be-
havior. Each succeeding term in the series involves a higher power of
x x . As x gets closer to x , these terms will become geometrically
smaller.
We will use the Taylor series approximations of the performance index to
investigate the shape of the performance index in the neighborhood of pos-
sible optimum points.
To experiment with Taylor series expansions of the cosine function, use the
MATLAB Neural Network Design Demonstration Taylor Series (nnd8ts1).
Vector Case
Of course the neural network performance index will not be a function of a
scalar x . It will be a function of all of the network parameters (weights and
biases), of which there may be a very large number. Therefore, we need to
extend the Taylor series expansion to functions of many variables. Consid-
er the following function of n variables:
F x = F x 1 x 2 x n . (8.7)
The Taylor series expansion for this function, about the point x , will be

F x = F x + Fx x x + F x x x
x1 x = x 1 1 x2 x = x 2 2
2
1
++ x x + --- 2 F x x x
2
Fx
xn x = x n n 2 x x = x 1 1
1
2
1
+ --- Fx x x x 2 x 2 +
2 x 1 x 2 x = x 1 1
(8.8)
This notation is a bit cumbersome. It is more convenient to write it in ma-
trix form, as in:
F x = F x + F x x x
T
x = x
1
+ --- x x 2F x x x +
T
2 x = x (8.9)
Gradient where F x is the gradient, and is defined as
8-4
Directional Derivatives
T
F x =
Fx F x Fx , (8.10)
x1 x2 xn
Hessian and 2F x is the Hessian, and is defined as:
2 2 2

F x Fx Fx
x1
2 x 1 x 2 x 1 x n
2 2 2

Fx F x Fx .
2F x = x 2 x 1 x2
2 x 2 x n (8.11)

2 2 2

Fx Fx F x
x n x 1 x n x 2 xn
2
The gradient and the Hessian are very important to our understanding of
performance surfaces. In the next section we discuss the practical meaning
of these two concepts.
To experiment with Taylor series expansions of a function of two variables,
use the MATLAB Neural Network Design Demonstration Vector Taylor
Series(nnd8ts2).
The ith element of the gradient, F x x i , is the first derivative of the per-
formance index F along the x i axis. The ith element of the diagonal of the
2 2
Hessian matrix, F x x i , is the second derivative of the performance in-
dex F along the x i axis. What if we want to know the derivative of the func-
Directional Derivative
tion in an arbitrary direction? We let p be a vector in the direction along
which we wish to know the derivative. This directional derivative can be
8
computed from
T
p F x
---------------------- . (8.12)
p
The second derivative along p can also be computed:

T
p 2F x p
-.
---------------------------
2
(8.13)
p
8-5
To illustrate these concepts, consider the function

2
+2
2 2
F x = x 1 + 2x 2 . (8.14)
Suppose that we want to know the derivative of the function at the point
T T
x = 0.5 0.5 in the direction p = 2 1 . First we evaluate the gradient
at x :

F x
x1 2x 1
F x = = = 1 . (8.15)
x = x 4x 2 2
F x
x2 x = x
x = x
The derivative in the direction p can then be computed:
1
T
2 1
p F x 2 0
---------------------- = ------------------------ = ------- = 0 . (8.16)
p 5
2
1
Therefore the function has zero slope in the direction p from the point x .
Why did this happen? What can we say about those directions that have
zero slope? If we consider the definition of directional derivative in Eq.
(8.12), we can see that the numerator is an inner product between the di-
rection vector and the gradient. Therefore any direction that is orthogonal
to the gradient will have zero slope.
Which direction has the greatest slope? The maximum slope will occur
when the inner product of the direction vector and the gradient is a maxi-
mum. This happens when the direction vector is the same as the gradient.
(Notice that the magnitude of the direction vector has no effect, since we
normalize by its magnitude.) This effect is illustrated in Figure 8.2, which
shows a contour plot and a 3-D plot of F x . On the contour plot we see five
vectors starting from our nominal point x and pointing in different direc-
tions. At the end of each vector the first directional derivative is displayed.
The maximum derivative occurs in the direction of the gradient. The zero
derivative is in the direction orthogonal to the gradient (tangent to the con-
tour line).
To experiment with directional derivatives, use the MATLAB Neural Net-
work Design Demonstration Directional Derivatives (nnd8dd).
8-6
Minima
2.2 6
2.0
1
1.6
4
0.8
x2 0 0.0
2
0
-1 2
1 2
0 1
x2 0
-1 x1
-1
-2 -2
-2 -1 0 1 2 -2
x1
Figure 8.2 Quadratic Function and Directional Derivatives
Minima
Recall that the objective of performance learning will be to optimize the
network performance index. In this section we want to define what we
mean by an optimum point. We will assume that the optimum point is a
minimum of the performance index. The definitions can be easily modified
for maximization problems.
Strong Minimum Strong Minimum
The point x is a strong minimum of F x if a scalar 0 exists, such that
F x F x + x for all x such that x 0 .
In other words, if we move away from a strong minimum a small distance
in any direction the function will increase.
Global Minimum Global Minimum
The point x is a unique global minimum of F x if F x F x + x for all
x 0 .
8
For a simple strong minimum, x , the function may be smaller than F x
at some points outside a small neighborhood of x . Therefore this is some-
times called a local minimum. For a global minimum the function will be
larger than the minimum point at every other point in the parameter
space.
Weak Minimum Weak Minimum
The point x is a weak minimum of F x if it is not a strong minimum, and a sca-
lar 0 exists, such that F x F x + x for all x such that x 0 .
8-7
No matter which direction we move away from a weak minimum, the func-
tion cannot decrease, although there may be some directions in which the
function does not change.
2 As an example of local and global minimum points, consider the following
+2
scalar function:
4 2 1
F x = 3x 7x --- x + 6 . (8.17)
2
This function is displayed in Figure 8.3. Notice that it has two strong min-
imum points: at approximately -1.1 and 1.1. For both of these points the
function increases in a local neighborhood. The minimum at 1.1 is a global
minimum, since there is no other point for which the function is as small.
There is no weak minimum for this function. We will show a two-dimen-
sional example of a weak minimum later.
8
4 2 1
F x = 3x 7x --- x + 6
2
6
2 Local Minimum
Global Minimum
0
-2 -1 0 1 2
Figure 8.3 Scalar Example of Local and Global Minima

2 Now lets consider some vector cases. First, consider the following function:
+2
4
F x = x 2 x 1 + 8x 1 x 2 x 1 + x 2 + 3 . (8.18)
Contour Plot In Figure 8.4 we have a contour plot (a series of curves along which the
function value remains constant) and a 3-D surface plot for this function
(for function values less than 12). We can see that the function has two
strong local minimum points: one at (-0.42, 0.42), and the other at (0.55,
-0.55). The global minimum point is at (0.55, -0.55).
There is also another interesting feature of this function at (-0.13, 0.13). It
Saddle Point is called a saddle point because of the shape of the surface in the neighbor-
hood of the point. It is characterized by the fact that along the line x 1 = x 2
the saddle point is a local maximum, but along a line orthogonal to that line
it is a local minimum. We will investigate this example in more detail in
Problems P8.2 and P8.5.
8-8
Necessary Conditions for Optimality
This function is used in the MATLAB Neural Network Design Demonstra-

tion Vector Taylor Series (nnd8ts2).
2
12
0
4
0
-1 2
1 2
0 1
0
-1
-1
-2 -2
-2 -1 0 1 2 -2
Figure 8.4 Vector Example of Minima and Saddle Point

2 As a final example, consider the function defined in Eq. (8.19):
+2
2 2 2
F x = x 1 1.5x 1 x 2 + 2x 2 x 1 (8.19)
The contour and 3-D plots of this function are given in Figure 8.5. Here we
can see that any point along the line x 1 = 0 is a weak minimum.
2
1
6
4
0
0
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
Figure 8.5 Weak Minimum Example

Now that we have defined what we mean by an optimum (minimum) point,
lets identify some conditions that would have to be satisfied by such a
point. We will again use the Taylor series expansion to derive these condi-
tions:
8-9
F x = F x + x = F x + F x
T
x
x = x
(8.20)
1 T
+ --- x 2F x x + ,
2 x = x
where
x = x x . (8.21)
First-Order Conditions
If x is very small then the higher order terms in Eq. (8.20) will be neg-
ligible and we can approximate the function as
F x + x F x + F x
T
x . (8.22)
x = x
The point x is a candidate minimum point, which means that the function
should go up (or at least not go down) if x is not zero. For this to happen
the second term in Eq. (8.22) should not be negative. In other words
T
F x x 0 . (8.23)
x = x
However, if this term is positive,

T
F x x 0 , (8.24)
x = x
then this would imply that
F x x F x F x x F x .
T
(8.25)
x = x
But this is a contradiction, since x should be a minimum point. Therefore,

since Eq. (8.23) must be true, and Eq. (8.24) must be false, the only alter-
native must be that
T
F x x = 0 . (8.26)
x = x
Since this must be true for any x , we have
F x = 0. (8.27)
x = x
Therefore the gradient must be zero at a minimum point. This is a first-or-

der, necessary (but not sufficient) condition for x to be a local minimum
Stationary Points point. Any points that satisfy Eq. (8.27) are called stationary points.
8-10
Second-Order Conditions
Assume that we have a stationary point x . Since the gradient of F x is
zero at all stationary points, the Taylor series expansion will be
1 T
F x + x = F x + --- x 2F x x + . (8.28)
2 x = x
As before, we will consider only those points in a small neighborhood of x ,

so that x is small and F x can be approximated by the first two terms
in Eq. (8.28). Therefore a strong minimum will exist at x if
T
x 2F x x 0 . (8.29)
x = x
For this to be true for arbitrary x 0 requires that the Hessian matrix be
Positive Definite Matrix positive definite. (By definition, a matrix A is positive definite if
T
z Az 0 (8.30)
Positive Semidefinite for any vector z 0 . It is positive semidefinite if

T
z Az 0 (8.31)
for any vector z . We can check these conditions by testing the eigenvalues
of the matrix. If all eigenvalues are positive, then the matrix is positive def-
inite. If all eigenvalues are nonnegative, then the matrix is positive
semidefinite.)
Sufficient Condition A positive definite Hessian matrix is a second-order, sufficient condition for
a strong minimum to exist. It is not a necessary condition. A minimum can
still be strong if the second-order term of the Taylor series is zero, but the
third-order term is positive. Therefore the second-order, necessary condi-
tion for a strong minimum is that the Hessian matrix be positive semi-def-
inite.
2 To illustrate these conditions, consider the following function of two vari-
+2
ables:
4 2
F x = x1 + x2 . (8.32)
First, we want to locate any stationary points, so we need to evaluate the

gradient:
3
F x = 4x 1 = 0 . (8.33)
2x 2
8-11
Therefore the only stationary point is the point x = 0 . We now need to

test the second-order condition, which requires the Hessian matrix:
2
2F x x=0
= 12x 1 0 = 0 0 . (8.34)
0 2 0 2
x=0
This matrix is positive semidefinite, which is a necessary condition for

x = 0 to be a strong minimum point. We cannot guarantee from first-or-
der and second-order conditions that it is a minimum point, but we have
not eliminated it as a possibility. Actually, even though the Hessian matrix
is only positive semidefinite, x = 0 is a strong minimum point, but we
cannot prove it from the conditions we have discussed.
Just to summarize, the necessary conditions for x to be a minimum,
strong or weak, of F x are:
F x = 0 and 2F x positive semidefinite.

x = x x = x
The sufficient conditions for x to be a strong minimum point of F x are:
F x = 0 and 2F x positive definite.

x = x x = x
Quadratic Functions
We will find throughout this text that one type of performance index is uni-
versal the quadratic function. This is true because there are many appli-
cations in which the quadratic function appears, but also because many
functions can be approximated by quadratic functions in small neighbor-
hoods, especially near local minimum points. For this reason we want to
spend a little time investigating the characteristics of the quadratic func-
tion.
Quadratic Function The general form of a quadratic function is
1 T T
F x = --- x Ax + d x + c , (8.35)
2
where the matrix A is symmetric. (If the matrix is not symmetric it can be
replaced by a symmetric matrix that produces the same F x . Try it!)
To find the gradient for this function, we will use the following useful prop-
erties of the gradient:
T T
h x = x h = h , (8.36)
where h is a constant vector, and
8-12
Quadratic Functions
T T
x Qx = Qx + Q x = 2Qx (for symmetric Q) . (8.37)
We can now compute the gradient of F x :
F x = Ax + d , (8.38)
and in a similar way we can find the Hessian:
2F x = A . (8.39)
All higher derivatives of the quadratic function are zero. Therefore the first
three terms of the Taylor series expansion (as in Eq. (8.20)) give an exact
representation of the function. We can also say that all analytic functions
behave like quadratics over a small neighborhood (i.e., when x is small).
Eigensystem of the Hessian

We now want to investigate the general shape of the quadratic function. It
turns out that we can tell a lot about the shape by looking at the eigenval-
ues and eigenvectors of the Hessian matrix. Consider a quadratic function
that has a stationary point at the origin, and whose value there is zero:
1 T
F x = --- x Ax . (8.40)
2
The shape of this function can be seen more clearly if we perform a change
of basis (see Chapter 6). We want to use the eigenvectors of the Hessian
matrix, A , as the new basis vectors. Since A is symmetric, its eigenvectors
will be mutually orthogonal. (See [Brog91].) This means that if we make up
a matrix with the eigenvectors as the columns, as in Eq. (6.68):
B = z1 z2 zn , (8.41)
the inverse of the matrix will be the same as the transpose:

1 T
B = B . (8.42)
(This assumes that we have normalized the eigenvectors.)

If we now perform a change of basis, so that the eigenvectors are the basis
vectors (as in Eq. (6.69)), the new A matrix will be
1 0 0
T 0 2 0
A' = B AB = =

0 0 n
, (8.43)
8-13
where the i are the eigenvalues of A . We can also write this equation as
T
A = BB . (8.44)
We will now use the concept of the directional derivative to explain the
physical meaning of the eigenvalues and eigenvectors of A , and to explain
how they determine the shape of the surface of the quadratic function.
Recall from Eq. (8.13) that the second derivative of a function F x in the
direction of a vector p is given by
T T
p 2F x p p Ap
---------------------------
2
-.
- = -------------
2
(8.45)
p p
Now define
p = Bc , (8.46)
where c is the representation of the vector p with respect to the eigenvec-

tors of A . (See Eq. (6.28) and the discussion that follows.) With this defini-
tion, and Eq. (8.44), we can rewrite Eq. (8.45):
n
i ci
2
T T T T T
p Ap c B BB Bc c c
-------------
2
- = ---------------------------------------
T T
- = i------------------
- = -----------
T
=1
n
-. (8.47)
p c B Bc c c
ci
2
i=1
This result tells us several useful things. First, note that this second deriv-
ative is just a weighted average of the eigenvalues. Therefore it can never
be larger than the largest eigenvalue, or smaller than the smallest eigen-
value. In other words,
T
p Ap
2
- max .
min ------------- (8.48)
p
Under what condition, if any, will this second derivative be equal to the
largest eigenvalue? What if we choose
p = z max , (8.49)
where z max is the eigenvector associated with the largest eigenvalue, max ?
For this case the c vector will be
T T T
c = B p = B z max = 0 0 0 1 0 0 , (8.50)
8-14
Quadratic Functions
where the one occurs only in the position that corresponds to the largest
eigenvalue (i.e., c max = 1 ). This is because the eigenvectors are orthonor-
mal.
If we now substitute z max for p in Eq. (8.47) we obtain
n
i ci
2
T
z max Az max
---------------------------
2
i=1
- = max .
- = ------------------
n
(8.51)
z max
ci
2
i=1
So the maximum second derivative occurs in the direction of the eigenvec-

tor that corresponds to the largest eigenvalue. In fact, in each of the eigen-
vector directions the second derivatives will be equal to the corresponding
eigenvalue. In other directions the second derivative will be a weighted av-
erage of the eigenvalues. The eigenvalues are the second derivatives in the
directions of the eigenvectors.
The eigenvectors define a new coordinate system in which the quadratic
cross terms vanish. The eigenvectors are known as the principal axes of the
z1 function contours. The figure to the left illustrates these concepts in two di-
z2 (min) mensions. This figure illustrates the case where the first eigenvalue is
(max) smaller than the second eigenvalue. Therefore the minimum curvature
(second derivative) will occur in the direction of the first eigenvector. This
means that we will cross contour lines more slowly in this direction. The
maximum curvature will occur in the direction of the second eigenvector,
therefore we will cross contour lines more quickly in that direction.
One caveat about this figure: it is only valid when both eigenvalues have
the same sign, so that we have either a strong minimum or a strong maxi-
mum. For these cases the contour lines are always elliptical. We will pro-
vide examples later where the eigenvalues have opposite signs and where
one of the eigenvalues is zero.
2 For our first example, consider the following function:
+2
1 T
F x = x 1 + x 2 = --- x 2 0 x .
2 2
(8.52)
2 0 2
The Hessian matrix and its eigenvalues and eigenvectors are
2F x = 2 0 , 1 = 2 , z 1 = 1 , 2 = 2 , z 2 = 0 . (8.53)
0 2 0 1
8-15
(Actually, any two independent vectors could be the eigenvectors in this

case. There is a repeated eigenvalue, and its eigenvector is the plane.)
Since all the eigenvalues are equal, the curvature should be the same in all
directions, and therefore the function should have circular contours. Figure
8.6 shows the contour and 3-D plots for this function, a circular hollow.
2
2
0
0
-1 2
1 2
0 1
0
-1
-1
-2 -2
-2 -1 0 1 2 -2
Figure 8.6 Circular Hollow

2 Lets try an example with distinct eigenvalues. Consider the following qua-
+2
dratic function:
1 T
F x = x 1 + x 1 x 2 + x 2 = --- x 2 1 x
2 2
(8.54)
2 12
2F x = 2 1 , 1 = 1 , z 1 = 1 , 2 = 3 , z 2 = 1 . (8.55)
1 2 1 1
(As we discussed in Chapter 6, the eigenvectors are not unique, they can be
multiplied by any scalar.) In this case the maximum curvature is in the di-
rection of z 2 so we should cross contour lines more quickly in that direction.
Figure 8.7 shows the contour and 3-D plots for this function, an elliptical
hollow.
8-16
Quadratic Functions
0
1
0
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
Figure 8.7 Elliptical Hollow

2 What happens when the eigenvalues have opposite signs? Consider the fol-
+2
lowing function:
1 2 3 1 2 1 T
F x = --- x 1 --- x 1 x 2 --- x 2 = --- x 0.5 1.5 x . (8.56)
4 2 4 2 1.5 0.5
2F x = 0.5 1.5 , 1 = 1 , z 1 = 1 , 2 = 2 , z 2 = 1 . (8.57)

1.5 0.5 1 1
The first eigenvalue is positive, so there is positive curvature in the direc-

tion of z 1 . The second eigenvalue is negative, so there is negative curvature
in the direction of z 2 . Also, since the magnitude of the second eigenvalue is
greater than the magnitude of the first eigenvalue, we will cross contour
lines faster in the direction of z 2 .
Figure 8.8 shows the contour and 3-D plots for this function, an elongated
saddle. Note that the stationary point,
x = 0 , (8.58)
0
is no longer a strong minimum point, since the Hessian matrix is not posi-
tive definite. Since the eigenvalues are of opposite sign, we know that the
Hessian is indefinite (see [Brog91]). The stationary point is therefore a sad-
dle point. It is a minimum of the function along the first eigenvector (posi-
tive eigenvalue), but it is a maximum of the function along the second
eigenvector (negative eigenvalue).
8-17
0
-4
-8
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
Figure 8.8 Elongated Saddle

2 As a final example, lets try a case where one of the eigenvalues is zero. An
+2
example of this is given by the following function:
1 2 1 2 1 T
F x = --- x 1 x 1 x 2 + --- x 2 = --- x 1 1 x . (8.59)
2 2 2 1 1
2F x = 1 1 , = 2 , z = 1 , = 0 , z = 1 . (8.60)
1 1 2 2
1 1 1 1
The second eigenvalue is zero, so we would expect to have zero curvature

along z 2 . Figure 8.9 shows the contour and 3-D plots for this function, a sta-
tionary valley. In this case the Hessian matrix is positive semidefinite, and
we have a weak minimum along the line
x1 = x2 , (8.61)
corresponding to the second eigenvector.

For quadratic functions the Hessian matrix must be positive definite in or-
der for a strong minimum to exist. For higher-order functions it is possible
to have a strong minimum with a positive semidefinite Hessian matrix, as
we discussed previously in the section on minima.
8-18
Quadratic Functions
0
1
0
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
Figure 8.9 Stationary Valley

To experiment with other quadratic functions, use the MATLAB Neural
Network Design Demonstration Quadratic Function (nnd8qf).
At this point we can summarize some characteristics of the quadratic func-
tion.
1. If the eigenvalues of the Hessian matrix are all positive, the function
will have a single strong minimum.
2. If the eigenvalues are all negative, the function will have a single
strong maximum.
3. If some eigenvalues are positive and other eigenvalues are negative,
the function will have a single saddle point.
4. If the eigenvalues are all nonnegative, but some eigenvalues are zero,
then the function will either have a weak minimum (as in Figure 8.9)
or will have no stationary point (see Solved Problem P8.7).
5. If the eigenvalues are all nonpositive, but some eigenvalues are zero,
then the function will either have a weak maximum or will have no sta-
tionary point.
We should note that in this discussion we have assumed, for simplicity,
that the stationary point of the quadratic function was at the origin, and
that it had a zero value there. This requires that the terms d and c in Eq.
(8.35) both be zero. If c is nonzero then the function is simply increased in
magnitude by c at every point. The shape of the contours do not change.
When d is nonzero, and A is invertible, the shape of the contours are not
changed, but the stationary point of the function moves to
x = A d .
1
(8.62)
If A is not invertible (has some zero eigenvalues) and d is nonzero then

stationary points may not exist (see Solved Problem P8.7).
8-19
Summary of Results
Taylor Series
F x = F x + F x x x
T
x = x
1
+ --- x x 2F x x x +
T
2 x = x
Gradient
T
F x =
F x Fx Fx
x1 x2 xn
Hessian Matrix
2 2 2

Fx F x Fx
x1
2 x 1 x 2 x 1 x n
2 2 2

F x Fx Fx
2F x = x 2 x 1 x2
2 x 2 x n

2 2 2

F x F x Fx
x n x 1 x n x 2 2
xn
First Directional Derivative
T
p F x
----------------------
p
Second Directional Derivative

T
p 2F x p
---------------------------
2
-
p
8-20
Summary of Results
Minima
Strong Minimum
The point x is a strong minimum of F x if a scalar 0 exists, such that
F x F x + x for all x such that x 0 .
Global Minimum
The point x is a unique global minimum of F x if F x F x + x for all
x 0 .
Weak Minimum
The point x is a weak minimum of F x if it is not a strong minimum, and a sca-
lar 0 exists, such that F x F x + x for all x such that x 0 .

First-Order Condition
F x = 0 (Stationary Points)
x = x
Second-Order Condition
2F x 0 (Positive Semidefinite Hessian Matrix)

x = x
Quadratic Functions
1 T T
F x = --- x Ax + d x + c
2
Gradient 8
F x = Ax + d
Hessian
2F x = A
T
p Ap
min -------------
2
- max
p
8-21
Solved Problems
P8.1 In Figure 8.1 we illustrated 3 approximations to the cosine func-
tion about the point x = 0 . Repeat that procedure about the point
x = 2 .
The function we want to approximate is
F x = cos x .
The Taylor series expansion for F x about the point x = 2 is
2
F x = cos x = cos --- sin --- x --- --- cos --- x ---
1
2 2 2 2 2 2
3
+ --- sin --- x --- +
1
6 2 2
3 5
= x --- + --- x --- --------- x --- +
1 1
2 6 2 120 2
The zeroth-order approximation of F x is
F x F0 x = 0 .
The first-order approximation is

F x F 1 x = x --- = --- x .
2 2
(Note that in this case the second-order approximation is the same as the
first-order approximation, since the second derivative is zero.)
The third-order approximation is
3
F x F 3 x = x --- + --- x --- .
1
2 6 2
A graph showing F x and these three approximations is shown in Figure

P8.1. Note that in this case the zeroth-order approximation is very poor,
while the first-order approximation is accurate over a reasonably wide
range. Compare this result with Figure 8.1. In that case we were expanding
about a local maximum point, x = 0 , so the first derivative was zero.
Check the Taylor series expansions at other points using the Neural Net-
work Design Demonstration Taylor Series (nnd8ts1).
8-22
Solved Problems
1.5
F1 x
1
0.5
F0 x
0
-0.5
-1 Fx F3 x
-1.5
-6 -4 -2 0 2 4 6
Figure P8.1 Cosine Approximation About x = 2
P8.2 Recall the function that is displayed in Figure 8.4, on page 8-9. We
know that this function has two strong minima. Find the second-
order Taylor series expansions for this function about the two min-
ima.
The equation for this function is
4
F x = x 2 x 1 + 8x 1 x 2 x 1 + x 2 + 3 .
To find the second-order Taylor series expansion, we need to find the gra-
dient and the Hessian for F x . For the gradient we have

F x 3
x1 4 x 2 x 1 + 8x 2 1
F x = = ,
3
4 x 2 x 1 + 8x 1 + 1
F x
x2
and the Hessian matrix is
2 2

Fx F x
x1
2 x 1 x 2
2F x =
2 2

F x Fx
x 2 x 1 x2
2
2 2
12 x 2 x 1 12 x 2 x 1 + 8
=
2 2
12 x 2 x 1 + 8 12 x 2 x 1
8-23
1 T
One strong minimum occurs at x = 0.42 0.42 , and the other at
2 T
x = 0.55 0.55 . If we perform the second-order Taylor series expansion
of F x about these two points we obtain:
1 1 T 1 1 1 T 1
F x = F x + F x x x + --- x x 2F x x x
x=x 1
2 x = x1
T
1 8.42 0.42 x 0.42 .
= 2.93 + --- x 0.42
2 0.42 0.42 8.42 0.42
If we simplify this expression we find
1 T
F x = 4.49 3.7128 3.7128 x + --- x 8.42 0.42 x .
1
2 0.42 8.42
2
Repeating this process for x results in
1 T
F x = 7.41 11.781 11.781 x + --- x 14.71 6.71 x .
2
2 6.71 14.71
The original function and the two approximations are plotted in the follow-
ing figures.
Check the Taylor series expansions at other points using the Neural Net-
work Design Demonstration Vector Taylor Series (nnd8ts2).
12
0
4
0
-1 2
1 2
0 1
0
-1
-1
-2 -2 -2
-2 -1 0 1 2
Figure P8.2 Function F x for Problem P8.2
8-24
Solved Problems
12
0
4
0
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
1
2
12
0
4
0
-1 2
1 2
0 1
0
-1
-1
-2
-2 -1 0 1 2 -2 -2
2
P8.3 For the function F x given below, find the equation for the line
T
that is tangent to the contour line at x = 0 0 .
2 2 2
F x = 2 + x1 + 5 1 x1 x2
To solve this problem we can use the directional derivative. What is the de-
rivative of F x along a line that is tangent to a contour line? Since the con-
tour is a line along which the function does not change, the derivative of
F x should be zero in the direction of the contour. So we can get the equa-
tion for the tangent to the contour line by setting the directional derivative
equal to zero.
First we need to find the gradient:
8-25
2 2
2 2 + x 1 + 10 1 x 1 x 2 1 6 + 12x 1 + 10x 2
F x = = .
2 3
10 1 x 1 x 2 2x 2 20x 2 + 20x 1 x 2 + 20x 2
T
If we evaluate this at x = 0 0 , we obtain
F x = 6 .
0
Now recall that the equation for the derivative of F x in the direction of a
vector p is
T
p F x
---------------------- .
p
Therefore if we want the equation for the line that passes through
T
x = 0 0 and along which the derivative is zero, we can set the numer-
ator of the directional derivative in the direction of x to zero:
x F x = 0 ,
T
where x = x x . For this case we have
x 6 = 0 , or x 1 = 0 .
T
This result is illustrated in Figure P8.5.

2
12
0
4
0
-1 2
1 2
0 1
0
-1
-1
-2 -2 -2
-2 -1 0 1 2
Figure P8.5 Plot of F x for Problem P8.3
8-26
Solved Problems
P8.4 Consider the following fourth-order polynomial:
4 2 3 2
F x = x --- x 2x + 2x + 4 .
3
Find any stationary points and test them to see if they are minima.
To find the stationary points we set the derivative of F x to zero:
d 3 2
F x = 4x 2x 4x + 2 = 0 .
dx
We can use MATLAB to find the roots of this polynomial:

coef=[4 -2 -4 2];
stapoints=roots(coef);
stapoints
ans =
1.0000 -1.0000 0.5000
Now we need to check the second derivative at each of these points. The sec-
ond derivative of F x is
2
d 2
2
F x = 12x 4x 4 .
dx
If we evaluate this at each of the stationary points we find
d2 d2 d2
2 F 1 = 4 2 F 1 = 12 2 F 0.5 = 3 .
dx dx dx
Therefore we should have strong local minima at 1 and -1 (since the second
derivatives were positive), and a strong local maximum at 0.5 (since the
second derivative was negative). To find the global minimum we would
have to evaluate the function at the two local minima:
F 1 = 4.333 F 1 = 1.667 .
Therefore the global minimum occurs at -1. But are we sure that this is a
global minimum? What happens to the function as x or x ? In
this case, because the highest power of x has a positive coefficient and is
4
an even power ( x ), the function goes to at both limits. So we can safely
say that the global minimum occurs at -1. The function is plotted in Figure
P8.6.
8-27
4 2 3 2
F(x) = x --- x 2x + 2x + 4
6 3
0
-2 -1 0 1 2
Figure P8.6 Graph of F x for Problem P8.4
P8.5 Look back to the function of Problem P8.2. This function has three
stationary points:
x = 0.41878 , x = 0.134797 , x = 0.55358 .

1 2 3
0.41878 0.134797 0.55358
Test whether or not any of these points could be local minima.

From Problem P8.2 we know that the Hessian matrix for the function is
2 2
12 x 2 x 1 12 x 2 x 1 + 8
2F x = .
2 2
12 x 2 x 1 + 8 12 x 2 x 1
To test the definiteness of this matrix we can check the eigenvalues. If the
eigenvalues are all positive, the Hessian is positive definite, which guaran-
tees a strong minimum. If the eigenvalues are nonnegative, the Hessian is
positive semidefinite, which is consistent with either a strong or a weak
minimum. If one eigenvalue is positive and the other eigenvalue is nega-
tive, the Hessian is indefinite, which would signal a saddle point.
1
If we evaluate the Hessian at x , we find
2F x =
1 8.42 0.42 .
0.42 8.42
The eigenvalues of this matrix are
1 = 8.84 , 2 = 8.0 ,
1
therefore x must be a strong minimum point.
8-28
Solved Problems
2
2F x = 0.87 7.13 .
2
7.13 0.87
1 = 6.26 , 2 = 8.0 ,
2
therefore x must be a saddle point. In one direction the curvature is neg-
ative, and in another direction the curvature is positive. The negative cur-
vature is in the direction of the first eigenvector, and the positive curvature
is in the direction of the second eigenvector. The eigenvectors are
z1 = 1 and z = 1 .
2
1 1
(Note that this is consistent with our previous discussion of this function
on page 8-8.)
3
3
2F x = 14.7 6.71 .
6.71 14.7
1 = 21.42 , 2 = 8.0 ,
3
therefore x must be a strong minimum point.
Check these results using the Neural Network Design Demonstration Vector
Taylor Series (nnd8ts2).
P8.6 Lets apply the concepts in this chapter to a neural network prob-
lem. Consider the linear network shown in Figure P8.7. Suppose
that the desired inputs/outputs for the network are
p 1 = 2 t 1 = 0.5 p 2 = 1 t 2 = 0 .
Sketch the following performance index for this network:

2 2
F x = t1 a1 x + t2 a2 x .
8-29
Input Linear Neuron
w n a
p
b
1
a = purelin (wp + b)
Figure P8.7 Linear Network for Problem P8.6

The parameters of this network are w and b , which make up the parame-
ter vector
x = w .
b
We want to sketch the performance index F x . First we will show that the
performance index is a quadratic function. Then we will find the eigenvec-
tors and eigenvalues of the Hessian matrix and use them to sketch the con-
tour plot of the function.
Begin by writing F x as an explicit function of the parameter vector x :
2 2
F x = e1 + e2 ,
where
e 1 = t 1 wp 1 + b e 2 = t 2 wp 2 + b .
This can be written in matrix form:

T
Fx = e e ,
where
p1 1
e = t x = t Gx .
p2 1
The performance index can now be rewritten:

T T T T T
F x = t Gx t Gx = t t 2t Gx + x G Gx .
8-30
Solved Problems
If we compare this with Eq. (8.35):
1 T T
F x = --- x Ax + d x + c ,
2
we can see that the performance index for this linear network is a quadrat-
ic function, with
T T T
c = t t , d = 2 G t , and A = 2G G .
The gradient of the quadratic function is given in Eq. (8.38):

T T
F x = Ax + d = 2G Gx 2 G t .
The stationary point (also the center of the function contours) will occur
where the gradient is equal to zero:
1
x = A d = G G G t .
1 T T
For
p1 1 2 1 and t = 0.5
G = =
p2 1 1 1 0
we have
1
1 T
x = G G G t = 5 1 1 = 0.167 .
T
1 2 0.5 0.167
(Therefore the optimal network parameters are w = 0.167 and b = 0.167 .)

The Hessian matrix of the quadratic function is given by Eq. (8.39):
2F x = A = 2G G = 10 2 .
T
2 4
To sketch the contour plot we need the eigenvectors and eigenvalues of the
Hessian. For this case we find
1 , = 3.4 z = 0.3 .
1 = 10.6 z 1 = 2 2
0.3 1
Therefore we know that x is a strong minimum. Also, since the first eigen-
value is larger than the second, we know that the contours will be elliptical
and that the long axis of the ellipses will be in the direction of the second
8-31
eigenvector. The contours will be centered at x . This is demonstrated in

Figure P8.8.
2
3
0
0
-1 2
1 2
0 1
0
-1
-1
-2 -2 -2
-2 -1 0 1 2
Figure P8.8 Graph of Function for Problem P8.6
P8.7 There are quadratic functions that do not have stationary points.
This problem illustrates one such case. Consider the following
function:
1 T
F x = 1 1 x + --- x 1 1 x .
2 1 1
Sketch the contour plot of this function.

As with Problem P8.6, we need to find the eigenvalues and eigenvectors of
the Hessian matrix. By inspection of the quadratic function we see that the
Hessian matrix is
2F x = A = 1 1 . (8.63)
1 1
The eigenvalues and eigenvectors are
1 1 .
1 = 0 z 1 = , 2 = 2 z 2 =
1 1
Notice that the first eigenvalue is zero, so there is no curvature along the
first eigenvector. The second eigenvalue is positive, so there is positive cur-
vature along the second eigenvector. If we had no linear term in F x , the
plot of the function would show a stationary valley, as in Figure 8.9. In this
case we must find out if the linear term creates a slope in the direction of
the valley (the direction of the first eigenvector).
8-32
Solved Problems
The linear term is
F lin x = 1 1 x .
From Eq. (8.36) we know that the gradient of this term is
F lin x = 1 ,
1
which means that the linear term is increasing most rapidly in the direc-
tion of this gradient. Since the quadratic term has no curvature in this di-
rection, the overall function will have a linear slope in this direction.
Therefore F x will have positive curvature in the direction of the second
eigenvector and a linear slope in the direction of the first eigenvector. The
contour plot and the 3-D plot for this function are given in Figure P8.9.
2
-3
-1 2
1 2
0 1
0
-1
-1
-2 -2 -2
-2 -1 0 1 2
Figure P8.9 Falling Valley Function for Problem P8.7

Whenever any of the eigenvalues of the Hessian matrix are zero it is impos-
sible to solve for the stationary point of the quadratic function using
x = A d ,
1
since the Hessian matrix does not have an inverse. This lack of an inverse
could mean that we have a weak minimum point, as illustrated in Figure
8.9, or that there is no stationary point, as this example shows.
8-33
Epilogue
Performance learning is one of the most important classes of neural net-
work learning rules. With performance learning, network parameters are
adjusted to optimize network performance. In this chapter we have intro-
duced tools that we will need to understand performance learning rules. Af-
ter reading this chapter and solving the exercises, you should be able to:
i. Perform a Taylor series expansion and use it to approximate a func-
tion.
ii. Calculate a directional derivative.
iii. Find stationary points and test whether they could be minima.
iv. Sketch contour plots of quadratic functions.
We will be using these concepts in a number of succeeding chapters, includ-
ing the chapters on performance learning (914), the radial basis network
chapter (17) and the chapters on stability and Hopfield networks (2021).
In the next chapter we will build on the concepts we have covered here, to
design algorithms that will optimize performance functions. Then, in suc-
ceeding chapters, we will apply these algorithms to the training of neural
networks.
8-34
Further Reading
Further Reading
[Gill81] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimi-
zation, New York: Academic Press, 1981.
As the title implies, this text emphasizes the practical im-
plementation of optimization algorithms. It provides moti-
vation for the optimization methods, as well as details of
implementation that affect algorithm performance.
[Himm72] D. M. Himmelblau, Applied Nonlinear Programming, New
York: McGraw-Hill, 1972.
This is a comprehensive text on nonlinear optimization. It
covers both constrained and unconstrained optimization
problems. The text is very complete, with many examples
worked out in detail.
[Scal85] L. E. Scales, Introduction to Non-Linear Optimization, New
York: Springer-Verlag, 1985.
A very readable text describing the major optimization al-
gorithms, this text emphasizes methods of optimization
rather than existence theorems and proofs of convergence.
Algorithms are presented with intuitive explanations,
along with illustrative figures and examples. Pseudo-code
is presented for most algorithms.
8
8-35
Exercises
E8.1 Consider the following scalar function:
1
F x = -------------------------- .
3 3 1
x --- x ---
4 2
i. Find the second-order Taylor series approximation for F x about

the point x = 0.5 .
ii. Find the second-order Taylor series approximation for F x about
the point x = 1.1 .
iii. Plot F x and the two approximations and discuss their accuracy.
E8.2 Consider the following function of two variables:

2 2
2x 1 + 2x 2 + x 1 5x 2 + 10
F x = e .
i. Find the second-order Taylor series approximation for F x about

T
the point x = 0 0 .
ii. Find the stationary point for this approximation.

iii. Find the stationary point for F x . (Note that the exponent of F x
is simply a quadratic function.)
2+2
ans =
iv. Explain the difference between the two stationary points. (Use
4 MATLAB to plot the two functions.)
E8.3 For the following functions find the first and second directional derivatives
T T
from the point x = 1 1 in the direction p = 1 1 .
7 2 2
i. F x = --- x 1 6x 1 x 2 x 2
2
2 2
ii. F x = 5x 1 6x 1 x 2 + 5x 2 + 4x 1 + 4x 2
9 2 2
iii. F x = --- x 1 2x 1 x 2 + 3x 2 + 2x 1 x 2
2
1 2 2
iv. F x = --- 7x 1 + 12x 1 x 2 2x 2
2
2 2
v. F x = x 1 + x 1 x 2 + x 2 + 3x 1 + 3x 2
8-36
Exercises
1 2 1 2
vi. F x = --- x 1 3x 1 x 2 + --- x 2 4x 1 + 4x 2
2 2
1 2 2
vii. F x = --- x 1 2x 1 x 2 + 2x 2 + x 1 2x 2
2
3 2
viii. F x = --- x 1 + 2x 1 x 2 + 4x 1 + 4x 2
2
3 2 3 2
ix. F x = --- x 1 + 4x 1 x 2 + --- x 2 + 5x 1
2 2
2 1 2
x. F x = 2x 1 2x 1 x 2 + --- x 2 + x 1 + x 2
2
E8.4 For the following function,
4 1 2
F x = x --- x + 1 ,
2
i. find the stationary points,

ii. test the stationary points to find minimum and maximum points,
and
2+2
ans = iii. plot the function using MATLAB to verify your answers.
4
E8.5 Consider the following function of two variables:

4
F x = x 1 + x 2 12x 1 x 2 + x 1 + x 2 + 1 .
i. Verify that the function has three stationary points at
x = 0.6504 , x = 0.085 , x = 0.5655 .

1 2 3
0.6504 0.085 0.5655
ii. Test the stationary points to find any minima, maxima or saddle
points.
iii. Find the second-order Taylor series approximations for the function
at each of the stationary points.
2+2
ans = iv. Plot the function and the approximations using MATLAB.
4
E8.6 For the functions of Exercise E8.3:

i. find the stationary points,
ii. test the stationary points to find minima, maxima or saddle points,
iii. provide rough sketches of the contour plots, using the eigenvalues
8-37
and eigenvectors of the Hessian matrices, and

2+2
ans = iv. plot the functions using MATLAB to verify your answers.
4
E8.7 Consider the following quadratic function:
1 T
F x = --- x 1 3 x + 4 4 x + 2 .
2 3 1
i. Find the gradient and Hessian matrix for F x .

ii. Sketch the contour plot for F x .
T
iii. Find the directional derivative of F x at the point x 0 = 0 0 in
T
the direction p = 1 1 .
iv. Is your answer to part iii. consistent with your contour plot of part
ii.? Explain.
E8.8 Repeat Exercise E8.7 with the following quadratic function:
1 T
F x = --- x 3 2 x + 4 4 x + 2 .
2 2 0
E8.9 Consider the following function:
2 1 4
F x = 1 + x 1 + x 2 + --- x .
4 1
i. Find the quadratic approximation to F x about the point

T
x0 = 1 0
ii. Sketch the contour plot of the quadratic approximation in part i.
3 2 3
F x = --- x + 2x 1 x 2 + x 2 + 4x 1 + 4x 2 .
2 1

T
x0 = 1 0 .
ii. Locate the stationary point of the quadratic approximation you

found in part i.
8-38
Exercises
iii. Is the answer to part ii a minimum of F x ?
F x = x 1 x 2 x 1 + 2x 2 .
i. Locate any stationary points.

ii. For each answer to part i., determine, if possible, whether the sta-
tionary point is a minimum point, a maximum point, or a saddle
point.
iii. Find the directional derivative of the function at the point
T T
x 0 = 1 1 in the direction p = 1 1 .

2 2 3
F x = x 1 + 2x 1 x 2 + x 2 + x 1 x 2 .

T
x0 = 2 1 .
ii. Sketch the contour plot of the quadratic approximation.
E8.13 Recall the function in Problem P8.7. For that function there was no station-
ary point. It is possible to modify the function, by changing only the d vec-
tor, so that a stationary point will exist. Find a new nonzero d vector that
will create a weak minimum.
8-39
Objectives
9 Performance Optimization
Objectives 9-1
Steepest Descent 9-2
Stable Learning Rates 9-6
Minimizing Along a Line 9-8
Newtons Method 9-10
Epilogue 9-37
Exercises 9-39
Objectives
We initiated our discussion of performance optimization in Chapter 8.
There we introduced the Taylor series expansion as a tool for analyzing the
performance surface, and then used it to determine conditions that must be
satisfied by optimum points. In this chapter we will again use the Taylor
series expansion, in this case to develop algorithms to locate the optimum
points. We will discuss three different categories of optimization algorithm:
steepest descent, Newtons method and conjugate gradient. In Chapters
1014 we will apply all of these algorithms to the training of neural net-
works.
9-1
Theory and Examples

In the previous chapter we began our investigation of performance surfac-
es. Now we are in a position to develop algorithms to search the parameter
space and locate minimum points of the surface (find the optimum weights
and biases for a given neural network).
It is interesting to note that most of the algorithms presented in this chap-
ter were developed hundreds of years ago. The basic principles of optimiza-
tion were discovered during the 17th century, by such scientists and
mathematicians as Kepler, Fermat, Newton and Leibniz. From 1950 on,
these principles were rediscovered to be implemented on high speed (in
comparison to the pen and paper available to Newton) digital computers.
The success of these efforts stimulated significant research on new algo-
rithms, and the field of optimization theory became recognized as a major
branch of mathematics. Now neural network researchers have access to a
vast storehouse of optimization theory and practice that can be applied to
the training of neural networks.
The objective of this chapter, then, is to develop algorithms to optimize a
performance index F x . For our purposes the word optimize will mean
to find the value of x that minimizes F x . All of the optimization algo-
rithms we will discuss are iterative. We begin from some initial guess, x 0 ,
and then update our guess in stages according to an equation of the form
xk + 1 = xk + k pk , (9.1)
or
x k = x k + 1 x k = k p k , (9.2)
where the vector p k represents a search direction, and the positive scalar
k is the learning rate, which determines the length of the step.
The algorithms we will discuss in this chapter are distinguished by the
choice of the search direction, p k . We will discuss three different possibili-
ties. There are also a variety of ways to select the learning rate, k , and we
will discuss several of these.
Steepest Descent
When we update our guess of the optimum (minimum) point using Eq.
(9.1), we would like to have the function decrease at each iteration. In other
words,
F xk + 1 F xk . (9.3)
9-2
Steepest Descent
How can we choose a direction, p k , so that for sufficiently small learning

rate, k , we will move downhill in this way? Consider the first-order Tay-
lor series expansion (see Eq. (8.9)) of F x about the old guess x k :
T
F x k + 1 = F x k + x k F x k + g k x k , (9.4)
where g k is the gradient evaluated at the old guess x k :
g k F x . (9.5)
x = xk
For F x k + 1 to be less than F x k , the second term on the right-hand side

of Eq. (9.4) must be negative:
T T
g k x k = k g k p k 0 . (9.6)
We will select an k that is small, but greater than zero. This implies:
T
gk pk 0 . (9.7)
Descent Direction Any vector p k that satisfies this equation is called a descent direction. The
function must go down if we take a small enough step in this direction. This
brings up another question. What is the direction of steepest descent? (In
what direction will the function decrease most rapidly?) This will occur
when
T
gk pk (9.8)
is most negative. (We assume that the length of p k does not change, only
the direction.) This is an inner product between the gradient and the direc-
tion vector. It will be most negative when the direction vector is the nega-
tive of the gradient. (Review our discussion of directional derivatives on
page 8-6.) Therefore a vector that points in the steepest descent direction is
pk = gk . (9.9)
Steepest Descent Using this in the iteration of Eq. (9.1) produces the method of steepest de-
scent:
xk + 1 = xk k gk . (9.10)
For steepest descent there are two general methods for determining the
Learning Rate learning rate, k . One approach is to minimize the performance index F x
with respect to k at each iteration. In this case we are minimizing along
the line
xk k gk . (9.11)
9-3
The other method for selecting k is to use a fixed value (e.g., k = 0.02 ),
or to use variable, but predetermined, values (e.g., k = 1 k ). We will dis-
cuss the choice of k in more detail in the following examples.
2 Lets apply the steepest descent algorithm to the following function,
+2
2 2
F x = x 1 + 25x 2 , (9.12)
starting from the initial guess
x 0 = 0.5 . (9.13)
0.5
The first step is to find the gradient:

Fx
x1 2x 1
F x = = . (9.14)
50x 2
Fx
x2
If we evaluate the gradient at the initial guess we find
g 0 = F x = 1 . (9.15)
x = x0 25
Assume that we use a fixed learning rate of = 0.01 . The first iteration of
the steepest descent algorithm would be
x 1 = x 0 g 0 = 0.5 0.01 1 = 0.49 . (9.16)

0.5 25 0.25
The second iteration of steepest descent produces
x 2 = x 1 g 1 = 0.49 0.01 0.98 = 0.4802 . (9.17)

0.25 12.5 0.125
If we continue the iterations we obtain the trajectory illustrated in Figure

9.1.
9-4
Steepest Descent
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure 9.1 Trajectory for Steepest Descent with = 0.01
Note that the steepest descent trajectory, for small learning rate, follows a
path that is always orthogonal to the contour lines. This is because the gra-
dient is orthogonal to the contour lines. (See the discussion on page 8-6.)
How would a change in the learning rate change the performance of the al-
gorithm? If we increase the learning rate to = 0.035 , we obtain the tra-
jectory illustrated in Figure 9.2. Note that the trajectory now oscillates. If
we make the learning rate too large the algorithm will become unstable;
the oscillations will increase instead of decaying.
1
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure 9.2 Trajectory for Steepest Descent with = 0.035
We would like to make the learning rate large, since then we will be taking
large steps and would expect to converge faster. However, as we can see
from this example, if we make the learning rate too large the algorithm will
become unstable. Is there some way to predict the maximum allowable
learning rate? This is not possible for arbitrary functions, but for quadratic
functions we can set an upper limit.
9-5
Stable Learning Rates

Suppose that the performance index is a quadratic function:
1 T T
F x = --- x Ax + d x + c . (9.18)
2
From Eq. (8.38) the gradient of the quadratic function is
F x = Ax + d . (9.19)
If we now insert this expression into our expression for the steepest descent
algorithm (assuming a constant learning rate), we obtain
x k + 1 = x k g k = x k Ax k + d (9.20)
or
x k + 1 = I A x k d . (9.21)
This is a linear dynamic system, which will be stable if the eigenvalues of

the matrix I A are less than one in magnitude (see [Brog91]). We can
express the eigenvalues of this matrix in terms of the eigenvalues of the
Hessian matrix A . Let 1 2 n and z 1 z 2 z n be the eigenval-
ues and eigenvectors of the Hessian matrix. Then
I A z i = z i Az i = z i i z i = 1 i z i . (9.22)
Therefore the eigenvectors of I A are the same as the eigenvectors of

A , and the eigenvalues of I A are 1 i . Our condition for the sta-
bility of the steepest descent algorithm is then
1 i 1 . (9.23)
If we assume that the quadratic function has a strong minimum point, then
its eigenvalues must be positive numbers. Eq. (9.23) then reduces to
2
---- . (9.24)
i
Since this must be true for all the eigenvalues of the Hessian matrix we
have
2
----------- . (9.25)
max
The maximum stable learning rate is inversely proportional to the maxi-

mum curvature of the quadratic function. The curvature tells us how fast
the gradient is changing. If the gradient is changing too fast we may jump
9-6
Steepest Descent
past the minimum point so far that the gradient at the new location will be
larger in magnitude (but opposite direction) than the gradient at the old lo-
cation. This will cause the steps to increase in size at each iteration.
2 Lets apply this result to our previous example. The Hessian matrix for that
+2
quadratic function is
A = 2 0 . (9.26)
0 50
The eigenvalues and eigenvectors of A are
1 = 50 z = 0 .
1 = 2 z 1 = 2 2 (9.27)
0 1
Therefore the maximum allowable learning rate is
2 2
----------- = ------ = 0.04 . (9.28)
max 50
This result is illustrated experimentally in Figure 9.3, which shows the

steepest descent trajectories when the learning rate is just below
( = 0.039 ) and just above ( = 0.041 ), the maximum stable value.
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Figure 9.3 Trajectories for = 0.039 (left) and = 0.041 (right).
This example has illustrated several points. The learning rate is limited by
the largest eigenvalue (second derivative) of the Hessian matrix. The algo-
rithm tends to converge most quickly in the direction of the eigenvector cor-
responding to this largest eigenvalue, and we dont want to overshoot the
minimum point by too far in that direction. (Note that in our examples the
initial step is almost parallel to the x 2 axis, which is z 2 .) However, the al-
gorithm will tend to converge most slowly in the direction of the eigenvec-
9-7
tor that corresponds to the smallest eigenvalue ( z 1 for our example). In the
end it is the smallest eigenvalue, in combination with the learning rate,
that determines how quickly the algorithm will converge. When there is a
great difference in magnitude between the largest and smallest eigenval-
ues, the steepest descent algorithm will converge slowly.
To experiment with steepest descent on this quadratic function, use the Neu-
ral Network Design Demonstration Steepest Descent for a Quadratic
(nnd9sdq).
Minimizing Along a Line

Another approach for selecting the learning rate is to minimize the perfor-
mance index with respect to k at each iteration. In other words, choose k
to minimize
F xk + k pk . (9.29)
To do this for arbitrary functions requires a line search, which we will dis-
cuss in Chapter 12. For quadratic functions it is possible to perform the lin-
ear minimization analytically. The derivative of Eq. (9.29) with respect to
k , for quadratic F x , can be shown to be
d T T
F x k + k p k = F x p + k p k 2F x p . (9.30)
d k x = xk k x = xk k
If we set this derivative equal to zero and solve for k , we obtain

T
F x p T
gk pk
x = xk k
k = --------------------------------------------
T
- = -,
-----------------
T
(9.31)
p k 2F x pk pk Ak pk
x = xk
where A k is the Hessian matrix evaluated at the old guess x k :
A k 2F x . (9.32)
x = xk
(For quadratic functions the Hessian matrix is not a function of k .)

2 Lets apply steepest descent with line minimization to the following qua-
+2
dratic function:
1 T
F x = --- x 2 1 x , (9.33)
2 1 2
9-8
Steepest Descent
x0 = 0.8 . (9.34)
0.25
The gradient of this function is
2x 1 + x 2
F x = . (9.35)
x 1 + 2x 2
The search direction for steepest descent is the negative of the gradient.
For the first iteration this will be
p 0 = g 0 = F x = 1.35 . (9.36)
x = x0 0.3
From Eq. (9.31), the learning rate for the first iteration will be
1.35
1.35 0.3
0.3
0 = ---------------------------------------------------------------- = 0.413 . (9.37)
2 1 1.35
1.35 0.3
1 2 0.3
The first step of steepest descent will then produce
x1 = x0 0 g0 = 0.8 0.413 1.35 = 0.24 . (9.38)

0.25 0.3 0.37
The first five iterations of the algorithm are illustrated in Figure 9.4.
Note that the successive steps of the algorithm are orthogonal. Why does
this happen? First, when we minimize along a line we will always stop at
a point that is tangent to a contour line. Then, since the gradient is orthog-
onal to the contour line, the next step, which is along the negative of the
gradient, will be orthogonal to the previous step.
We can show this analytically by using the chain rule on Eq. (9.30):
d d T d
F xk + k pk = F x k + 1 = F x x + k pk
d k d k x = xk + 1 d k k
(9.39)
T T
= F x pk = gk + 1 pk .
x = xk + 1
9-9
Therefore at the minimum point, where this derivative is zero, the gradient
is orthogonal to the previous search direction. Since the next search direc-
tion is the negative of this gradient, the consecutive search directions must
be orthogonal. (Note that this result implies that when minimizing in any
direction, the gradient at the minimum point will be orthogonal to the
search direction, even if we are not using steepest descent. We will use this
result in our discussion of conjugate directions.)
Contour Plot
1
0.5
x2
-0.5
-1
-1 -0.5 0 0.5 1
x1
Figure 9.4 Steepest Descent with Minimization Along a Line

To experiment with steepest descent with minimization along a line, use the
Neural Network Design Demonstration Method Comparison (nnd9mc).
Later in this chapter we will find that we can improve performance if we
adjust the search directions, so that instead of being orthogonal they are
conjugate. (We will define this term later.) If conjugate directions are used
the function can be exactly minimized in at most n steps, where n is the
dimension of x . (There are certain types of quadratic functions that are
minimized in one step by the steepest descent algorithm. Can you think of
such a function? How is its Hessian matrix characterized?)
Newtons Method
The derivation of the steepest descent algorithm was based on the first-or-
der Taylor series expansion (Eq. (9.4)). Newtons method is based on the
second-order Taylor series:
T 1 T
F x k + 1 = F x k + x k F x k + g k x k + --- x k A k x k . (9.40)
2
The principle behind Newtons method is to locate the stationary point of

this quadratic approximation to F x . If we use Eq. (8.38) to take the gra-
dient of this quadratic function with respect to x k and set it equal to zero,
we find
9-10
Newtons Method
g k + A k x k = 0 . (9.41)
Solving for x k produces

1
x k = A k g k . (9.42)
Newtons Method Newtons method is then defined:

1
xk + 1 = xk Ak gk . (9.43)
2 To illustrate the operation of Newtons method, lets apply it to our previous
+2
example function of Eq. (9.12):
2 2
F x = x 1 + 25x 2 . (9.44)
The gradient and Hessian matrices are

F x
x1 2x 1
F x = = , 2F x = 2 0 . (9.45)
50x 2 0 50
F x
x2
If we start from the same initial guess
x 0 = 0.5 , (9.46)
0.5
the first step of Newtons method would be

1
x 1 = 0.5 2 0 1 = 0.5 0.5 = 0 . (9.47)
0.5 0 50 25 0.5 0.5 0
This method will always find the minimum of a quadratic function in one
step. This is because Newtons method is designed to approximate a func-
tion as quadratic and then locate the stationary point of the quadratic ap-
proximation. If the original function is quadratic (with a strong minimum)
it will be minimized in one step. The trajectory of Newtons method for this
problem is given in Figure 9.5.
If the function F x is not quadratic, then Newtons method will not gener-
ally converge in one step. In fact, we cannot be sure that it will converge at
all, since this will depend on the function and the initial guess.
9-11
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure 9.5 Trajectory for Newtons Method

2 Recall the function given by Eq. (8.18):
+2
4
F x = x 2 x 1 + 8x 1 x 2 x 1 + x 2 + 3 . (9.48)
We know from Chapter 8 (see Problem P8.5) that this function has three
stationary points:
x = 0.41878 , x = 0.134797 , x = 0.55358 .

1 2 3
(9.49)
0.41878 0.134797 0.55358
The first point is a strong local minimum, the second point is a saddle point,
and the third point is a strong global minimum.
If we apply Newtons method to this problem, starting from the initial
T
guess x 0 = 1.5 0 , our first iteration will be as shown in Figure 9.6. The
graph on the left-hand side of the figure is a contour plot of the original
function. On the right we see the quadratic approximation to the function
at the initial guess.
The function is not minimized in one step, which is not surprising since the
function is not quadratic. However, we do take a step toward the global
minimum, and if we continue for two more iterations the algorithm will
converge to within 0.01 of the global minimum. Newtons method converges
quickly in many applications because analytic functions can be accurately
approximated by quadratic functions in a small neighborhood of a strong
minimum. So as we move closer to the minimum point, Newtons method
will more accurately predict its location. In this case we can see that the
contour plot of the quadratic approximation is similar to the contour plot of
the original function near the initial guess.
9-12
Newtons Method
2 2
1 1
0 0
-1 -1
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
T
Figure 9.6 One Iteration of Newtons Method from x 0 = 1.5 0
In Figure 9.7 we see one iteration of Newtons method from the initial guess
T
x 0 = 1.5 0 . In this case we are converging to the local minimum. Clear-
ly Newtons method cannot distinguish between a local minimum and a glo-
bal minimum, since it approximates the function as a quadratic, and the
quadratic function can have only one minimum. Newtons method, like
steepest descent, relies on the local features of the surface (the first and
second derivatives). It cannot know the global character of the function.
2 2
1 1
0 0
-1 -1
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
T
Figure 9.7 One Iteration of Newtons Method from x 0 = 1.5 0
T
x 0 = 0.75 0.75 . Now we are converging toward the saddle point of the
function. Note that Newtons method locates the stationary point of the
quadratic approximation to the function at the current guess. It does not
distinguish between minima, maxima and saddle points. For this problem
the quadratic approximation has a saddle point (indefinite Hessian ma-
9-13
trix), which is near the saddle point of the original function. If we continue
the iterations, the algorithm does converge to the saddle point of F x .
2 2
1 1
0 0
-1 -1
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
T
Figure 9.8 One Iteration of Newtons Method from x 0 = 0.75 0.75
In each of the cases we have looked at so far the stationary point of the qua-
dratic approximation has been close to a corresponding stationary point of
F x . This is not always the case. In fact, Newtons method can produce
very unpredictable results.
T
x 0 = 1.15 0.75 . In this case the quadratic approximation predicts a sad-
dle point, however, the saddle point is located very close to the local mini-
mum of F x . If we continue the iterations, the algorithm will converge to
the local minimum. Notice that the initial guess was actually farther away
from the local minimum than it was for the previous case, in which the al-
gorithm converged to the saddle point.
2 2
1 1
0 0
-1 -1
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
T
Figure 9.9 One Iteration of Newtons Method from x 0 = 1.15 0.75
9-14
Conjugate Gradient
To experiment with Newtons method and steepest descent on this function,

use the Neural Network Design Demonstrations Newtons Method (nnd9nm)
and Steepest Descent (nnd9sd).
This is a good place to summarize some of the properties of Newtons meth-
od that we have observed.
While Newtons method usually produces faster convergence than steepest
descent, the behavior of Newtons method can be quite complex. In addition
to the problem of convergence to saddle points (which is very unlikely with
steepest descent), it is possible for the algorithm to oscillate or diverge.
Steepest descent is guaranteed to converge, if the learning rate is not too
large or if we perform a linear minimization at each stage.
In Chapter 12 we will discuss a variation of Newtons method that is well
suited to neural network training. It eliminates the divergence problem by
using steepest descent steps whenever divergence begins to occur.
Another problem with Newtons method is that it requires the computation
and storage of the Hessian matrix, as well as its inverse. If we compare
steepest descent, Eq. (9.10), with Newtons method, Eq. (9.43), we see that
their search directions will be identical when
1
Ak = Ak = I . (9.50)
This observation has lead to a class of optimization algorithms know as

1
quasi-Newton or one-step-secant methods. These methods replace A k
with a positive definite matrix, H k ,which is updated at each iteration with-
out matrix inversion. The algorithms are typically designed so that for qua-
1
dratic functions H k will converge to A . (The Hessian is constant for
quadratic functions.) See [Gill81], [Scal85] or [Batt92] for a discussion of
these methods.
Conjugate Gradient
Quadratic Termination Newtons method has a property called quadratic termination, which
means that it minimizes a quadratic function exactly in a finite number of
iterations. Unfortunately, it requires calculation and storage of the second
derivatives. When the number of parameters, n , is large, it may be imprac-
tical to compute all of the second derivatives. (Note that the gradient has
2
n elements, while the Hessian has n elements.) This is especially true
with neural networks, where practical applications can require several
hundred to many thousand weights. For these cases we would like to have
methods that require only first derivatives but still have quadratic termi-
nation.
Recall the performance of the steepest descent algorithm, with linear
searches at each iteration. The search directions at consecutive iterations
were orthogonal (see Figure 9.4). For quadratic functions with elliptical
9-15
contours this produces a zig-zag trajectory of short steps. Perhaps quadrat-

ic search directions are not the best choice. Is there a set of search direc-
tions that will guarantee quadratic termination? One possibility is
conjugate directions.
Suppose that we wish to locate the minimum of the following quadratic
function:
1 T T
F x = --- x Ax + d x + c . (9.51)
2
Conjugate A set of vectors p k is mutually conjugate with respect to a positive defi-
nite Hessian matrix A if and only if
k j. (9.52)
T
p k Ap j = 0
As with orthogonal vectors, there are an infinite number of mutually con-

jugate sets of vectors that span a given n -dimensional space. One set of
conjugate vectors consists of the eigenvectors of A . Let 1 2 n and
z 1 z 2 z n be the eigenvalues and eigenvectors of the Hessian matrix.
To see that the eigenvectors are conjugate, replace p k with z k in Eq. (9.52):
T T
z k Az j = j z k z j = 0 kj, (9.53)
where the last equality holds because the eigenvectors of a symmetric ma-
trix are mutually orthogonal. Therefore the eigenvectors are both conju-
gate and orthogonal. (Can you find a quadratic function where all
orthogonal vectors are also conjugate?)
It is not surprising that we can minimize a quadratic function exactly by
searching along the eigenvectors of the Hessian matrix, since they form the
principal axes of the function contours. (See the discussion on pages 8-13
through 8-19.) Unfortunately this is not of much practical help, since to
find the eigenvectors we must first find the Hessian matrix. We want to
find an algorithm that does not require the computation of second deriva-
tives.
It can be shown (see [Scal85] or [Gill81]) that if we make a sequence of ex-
act linear searches along any set of conjugate directions p 1 p 2 p n ,
then the exact minimum of any quadratic function, with n parameters, will
be reached in at most n searches. The question is How can we construct
these conjugate search directions? First, we want to restate the conjugacy
condition, which is given in Eq. (9.52), without use of the Hessian matrix.
Recall that for quadratic functions
F x = Ax + d , (9.54)
2F x = A . (9.55)
9-16
Conjugate Gradient
By combining these equations we find that the change in the gradient at

iteration k + 1 is
g k = g k + 1 g k = Ax k + 1 + d Ax k + d = A x k , (9.56)
where, from Eq. (9.2), we have
x k = x k + 1 x k = k p k , (9.57)
and k is chosen to minimize F x in the direction p k .

We can now restate the conjugacy conditions (Eq. (9.52)):
T T T
k p k Ap j = x k Ap j = g k p j = 0 k j. (9.58)
Note that we no longer need to know the Hessian matrix. We have restated
the conjugacy conditions in terms of the changes in the gradient at succes-
sive iterations of the algorithm. The search directions will be conjugate if
they are orthogonal to the changes in the gradient.
Note that the first search direction, p 0 , is arbitrary, and p 1 can be any vec-
tor that is orthogonal to g 0 . Therefore there are an infinite number of sets
of conjugate vectors. It is common to begin the search in the steepest de-
scent direction:
p0 = g0 . (9.59)
Then, at each iteration we need to construct a vector p k that is orthogonal

to g 0 g 1 g k 1 . It is a procedure similar to Gram-Schmidt orthog-
onalization, which we discussed in Chapter 5. It can be simplified (see
[Scal85]) to iterations of the form
pk = gk + k pk 1 . (9.60)
The scalars k can be chosen by several different methods, which produce

equivalent results for quadratic functions. The most common choices (see
[Scal85]) are
T
g k 1 g k
k = --------------------------
T
-, (9.61)
g k 1 p k 1
due to Hestenes and Stiefel,

T
gk gk
k = ----------------------
T
- (9.62)
gk 1 gk 1
9-17
due to Fletcher and Reeves, and

T
g k 1 g k
k = ----------------------
T
- (9.63)
gk 1 gk 1
due to Polak and Ribire.

Conjugate Gradient To summarize our discussion, the conjugate gradient method consists of
the following steps:
1. Select the first search direction to be the negative of the gradient, as in
Eq. (9.59).
2. Take a step according to Eq. (9.57), selecting the learning rate k to
minimize the function along the search direction. We will discuss gen-
eral linear minimization techniques in Chapter 12. For quadratic func-
tions we can use Eq. (9.31).
3. Select the next search direction according to Eq. (9.60), using Eq.
(9.61), Eq. (9.62), or Eq. (9.63) to calculate k .
4. If the algorithm has not converged, return to step 2.
2 To illustrate the performance of the algorithm, recall the example we used
+2
to demonstrate steepest descent with linear minimization:
1 T
F x = --- x 2 1 x , (9.64)
2 1 2
with initial guess
x0 = 0.8 . (9.65)
0.25
2x 1 + x 2
F x = . (9.66)
x 1 + 2x 2
As with steepest descent, the first search direction is the negative of the
gradient:
= 1.35 .
T
p 0 = g 0 = F x (9.67)
x = x0 0.3
9-18
Conjugate Gradient
From Eq. (9.31), the learning rate for the first iteration will be
1.35
1.35 0.3
0.3
0 = ---------------------------------------------------------------- = 0.413 . (9.68)
2 1 1.35
1.35 0.3
1 2 0.3
The first step of conjugate gradient is therefore:
x1 = x0 + 0 p0 = 0.8 + 0.413 1.35 = 0.24 , (9.69)

0.25 0.3 0.37
which is equivalent to the first step of steepest descent with minimization

along a line.
Now we need to find the second search direction from Eq. (9.60). This re-
quires the gradient at x 1 :
g 1 = F x = 2 1 0.24 = 0.11 . (9.70)

x = x1 1 2 0.37 0.5
We can now find 1 :
0.11
T 0.11 0.5
g1 g1 0.5 0.2621
1 = ----------
T
- = -------------------------------------------- = ---------------- = 0.137 , (9.71)
g0 g0 1.35
1.9125
1.35 0.3
0.3
using the method of Fletcher and Reeves (Eq. (9.62)). The second search di-
rection is then computed from Eq. (9.60):
p 1 = g 1 + 1 p 0 = 0.11 + 0.137 1.35 = 0.295 . (9.72)

0.5 0.3 0.459
From Eq. (9.31), the learning rate for the second iteration will be
0.295
0.11 0.5
0.459 0.262
1 = ------------------------------------------------------------------------ = ------------- = 0.807 . (9.73)
0.325
2 1 0.295
0.295 0.459
1 2 0.459
9-19
The second step of conjugate gradient is therefore
x2 = x1 + 1 p1 = 0.24 + 0.807 0.295 = 0 . (9.74)

0.37 0.459 0
As predicted, the algorithm converges exactly to the minimum in two iter-

ations (since this is a two-dimensional quadratic function), as illustrated in
Figure 9.10. Compare this result with the steepest descent algorithm, as
shown in Figure 9.4. The conjugate gradient algorithm adjusts the second
search direction so that it will pass through the minimum of the function
(center of the function contours), instead of using an orthogonal search di-
rection, as in steepest descent.
1
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure 9.10 Conjugate Gradient Algorithm

We will return to the conjugate gradient algorithm in Chapter 12. In that
chapter we will discuss how the algorithm should be adjusted for non-qua-
dratic functions.
To experiment with the conjugate gradient algorithm and compare it with
steepest descent, use the Neural Network Design Demonstration Method
Comparison (nnd9mc).
9-20
Summary of Results
Summary of Results
General Minimization Algorithm

xk + 1 = xk + k pk
or
x k = x k + 1 x k = k p k
Steepest Descent Algorithm

xk + 1 = xk k gk
Where g k F x
x = xk
Stable Learning Rate ( k = , constant)

2
-----------
max
1 2 n Eigenvalues of Hessian matrix A
Learning Rate to Minimize Along the Line x k + 1 = x k + k p k

T
gk pk
k = ---------------
T
- (For quadratic functions)
p k Ap k
After Minimizing Along the Line x k + 1 = x k + k p k

T
gk + 1 pk = 0
Newtons Method 9
1
xk + 1 = xk Ak gk
Where A k 2F x
x = xk
9-21
Conjugate Gradient Algorithm

x k = k p k
Learning rate k is chosen to minimize along the line x k + 1 = x k + k p k .
p0 = g0
pk = gk + k pk 1
T T T
g k 1 g k gk gk g k 1 g k
k = --------------------------
T
- or k = ----------------------
T
- or k = ----------------------
T
-
g k 1 p k 1 gk 1 gk 1 gk 1 gk 1
Where g k F x and g k = g k + 1 g k .
x = xk
9-22
Solved Problems
Solved Problems
P9.1 We want to find the minimum of the following function:
2 2
F x = 5x 1 6x 1 x 2 + 5x 2 + 4x 1 + 4x 2 .
i. Sketch a contour plot of this function.

ii. Sketch the trajectory of the steepest descent algorithm on
the contour plot of part (i) if the initial guess is
T
x 0 = 1 2.5 . Assume a very small learning rate is used.
iii. What is the maximum stable learning rate?
i. To sketch the contour plot we first need to find the Hessian matrix. For
quadratic functions we can do this by putting the function into the stan-
dard form (see Eq. (8.35)):
1 T 1 T
F x = --- x Ax + d x + c = --- x 10 6 x + 4 4 x .
T
2 2 6 10
From Eq. (8.39) the Hessian matrix is
2F x = A = 10 6 .
6 10
The eigenvalues and eigenvectors of this matrix are
1 = 4 , z 1 = 1 , 2 = 16 , z 2 = 1 .
1 1
From the discussion on quadratic functions in Chapter 8 (see page 8-15) we

know that the function contours are elliptical. The maximum curvature of
F x is in the direction of z 2 , since 2 is larger than 1 , and the minimum
curvature is in the direction of z 1 (the long axis of the ellipses).
Next we need to find the center of the contours (the stationary point). This
occurs when the gradient is equal to zero. From Eq. (8.38) we find
F x = Ax + d = 10 6 x + 4 = 0 .
6 10 4 0
Therefore
9-23
1
x = 10 6 4 = 1 .
6 10 4 1
The contours will be elliptical, centered at x , with long axis in the direc-
tion of z 1 . The contour plot is shown in Figure P9.1.
ii. We know that the gradient is always orthogonal to the contour line,
therefore the steepest descent trajectory, if we take small enough steps,
will follow a path that is orthogonal to each contour line it intersects. We
can therefore trace the trajectory without performing any computations.
The result is shown in Figure P9.1.
0
-1
-2
-3
-3 -2 -1 0
Figure P9.1 Contour Plot and Steep. Desc. Trajectory for Problem P9.1
iii. From Eq. (9.25) we know that the maximum stable learning rate for a
quadratic function is determined by the maximum eigenvalue of the Hes-
sian matrix:
2
----------- .
max
The maximum eigenvalue for this problem is 2 = 16 , therefore for stabil-

ity
2
------ = 0.125 .
16
This result is verified experimentally in Figure P9.2, which shows the

steepest descent trajectories when the learning rate is just below
( = 0.12 ) and just above ( = 0.13 ) the maximum stable value.
9-24
Solved Problems
0 0
-1 -1
-2 -2
-3 -3
-3 -2 -1 0 -3 -2 -1 0
Figure P9.2 Trajectories for = 0.12 (left) and = 0.13 (right)
P9.2 Consider again the quadratic function of Problem P9.1. Take two
steps of the steepest descent algorithm, minimizing along a line at
each step. Use the following initial condition:
T
x0 = 0 2 .
In Problem P9.1 we found the gradient of the function to be
F x = Ax + d = 10 6 x + 4 .
6 10 4
If we evaluate this at x 0 , we find
g 0 = F x 0 = Ax 0 + d = 10 6 0 + 4 = 16 .
6 10 2 4 16
Therefore the first search direction is
p 0 = g 0 = 16 .
16
To minimize along a line, for a quadratic function, we can use Eq. (9.31):
16
T 16 16
g0 p0 16 512
0 = ---------------
T
- = -------------------------------------------------------- = ------------ = 0.0625 .
p 0 Ap 0 10 6 16
8192
16 16
6 10 16
9-25
Therefore the first iteration of steepest descent will be
x1 = x0 0 g0 = 0 0.0625 16 = 1 .
2 16 1
To begin the second iteration we need to find the gradient at x 1 :
g 1 = F x 1 = Ax 1 + d = 10 6 1 + 4 = 0 .
6 10 1 4 0
Therefore we have reached a stationary point; the algorithm has con-

verged. From Problem P9.1 we know that x 1 is indeed the minimum point
of this quadratic function. The trajectory is shown in Figure P9.3.
0
-1
-2
-3
-3 -2 -1 0
Figure P9.3 Steepest Descent with Linear Minimization for Problem P9.2
This is an unusual case, where the steepest descent algorithm located the
minimum in one iteration. Notice that this occurred because the initial
guess was located in the direction of one of the eigenvectors of the Hessian
matrix, with respect to the minimum point. For those cases where every di-
rection is an eigenvector, the steepest descent algorithm will always locate
the minimum in one iteration. What would this imply about the eigenval-
ues of the Hessian matrix?
P9.3 Recall Problem P8.6, in which we derived a performance index for

a linear neural network. The network, which is displayed again in
Figure P9.4, was to be trained for the following input/output pairs:
p 1 = 2 t 1 = 0.5 p 2 = 1 t 2 = 0
The performance index for the network was defined to be

2 2
F x = t1 a1 x + t2 a2 x ,
9-26
Solved Problems
which was displayed in Figure P8.8.

i. Use the steepest descent algorithm to locate the optimal pa-
T
rameters for this network (recall that x = w b ), starting
T
from the initial guess x 0 = 1 1 . Use a learning rate of
= 0.05 .
ii. What is the maximum stable learning rate?
Input Linear Neuron
w n a
p
b
1
a = purelin (wp + b)
Figure P9.4 Linear Network for Problems P9.3 and P8.6

i. In Problem P8.6 we found that the performance index could be written
in quadratic form:
1 T T
F x = --- x Ax + d x + c ,
2
where
c = t t = 0.5 0 0.5 = 0.25 ,

T
d = 2 G t = 2 2 1 0.5 = 2 ,
T
1 1 0 1
A = 2G G = 10 2 .
T
2 4
The gradient at x 0 is
g 0 = F x 0 = Ax 0 + d = 10 2 1 + 2 = 10 .
2 4 1 1 5
9-27
The first iteration of steepest descent will be
x 1 = x 0 g 0 = 1 0.05 10 = 0.5 .
1 5 0.75
The second iteration will be
x 2 = x 1 g 1 = 0.5 0.05 4.5 = 0.275 .

0.75 3 0.6
The remaining iterations are displayed in Figure P9.5. The algorithm con-
T
verges to the minimum point x = 0.167 0.167 . Therefore the optimal
value for both the weight and the bias of this network is 0.167 .
-1
-2
-2 -1 0 1 2
Figure P9.5 Steepest Descent Trajectory for Problem P9.3 with = 0.05
Note that in order to train this network we needed to know all of the input/
output pairs. We then performed iterations of the steepest descent algo-
rithm until convergence was achieved. In Chapter 10 we will introduce an
adaptive algorithm, based on steepest descent, for training linear net-
works. With this adaptive algorithm the network parameters are updated
after each input/output pair is presented. We will show how this allows the
network to adapt to a changing environment.
ii. The maximum eigenvalue of the Hessian matrix for this problem is
1 = 10.6 (see Problem P8.6), therefore for stability
2
---------- = 0.1887 .
10.6
9-28
Solved Problems
P9.4 Consider the function

2 2
x 1 x 1 + 2x 2 + 4
F x = e .
Take one iteration of Newtons method from the initial guess

T
x 0 = 1 2 . How close is this result to the minimum point of F x ?
Explain.
The first step is to find the gradient and the Hessian matrix. The gradient
is given by

F x
x1 2 2
x 1 x 1 + 2x 2 + 4 2x 1 1
F x = = e ,
4x 2
F x
x2
and the Hessian matrix is given by
2 2

Fx Fx
x1
2 x 1 x 2
2F x =
2 2

Fx Fx
x 2 x 1 x2
2
2
2 2
x 1 x 1 + 2x 2 + 4 4x 1 4x 1 + 3 2x 1 1 4x 2
= e
2
2x 1 1 4x 2 16x 2 + 4
If we evaluate these at the initial guess we find
6
g 0 = F x = 0.163 10 ,
x = x0 6
1.302 10
and
7 7
A 0 = 2F x = 0.049 10 0.130 10 .
x = x0 7 7
0.130 10 1.107 10
Therefore the first iteration of Newtons method, from Eq. (9.43), will be
9-29
1
7 7 6
g 0 = 1 0.049 10 0.130 10 0.163 10 0.971
1
x1 = x0 A0 =
2 7 7 6 1.886
0.130 10 1.107 10 1.302 10
How close is this to the true minimum point of F x ? First, note that the
exponent of F x is a quadratic function:
1 T 1 T
x 1 x 1 + 2x 2 + 4 = --- x Ax + d x + c = --- x 2 0 x + 1 0 x + 4 .
2 2 T
2 2 0 4
The minimum point of F x will be the same as the minimum point of the
exponent, which is
1
x = A d = 2 0 1 = 0.5 .
1
0 4 0 0
Therefore Newtons method has taken only a very small step toward the
true minimum point. This is because F x cannot be accurately approxi-
T
mated by a quadratic function in the neighborhood of x 0 = 1 2 .
For this problem Newtons method will converge to the true minimum
point, but it will take many iterations. The trajectory for Newtons method
is illustrated in Figure P9.6.
-1
-2
-2 -1 0 1 2
Figure P9.6 Newtons Method Trajectory for Problem P9.4
9-30
Solved Problems
P9.5 Compare the performance of Newtons method and steepest de-

scent on the following function:
1 T
F x = --- x 1 1 x .
2 1 1
Start from the initial guess
x0 = 1 .
0
Recall that this function is an example of a stationary valley (see Eq. (8.59)
and Figure 8.9). The gradient is
F x = Ax + d = 1 1 x
1 1
2F x = A = 1 1 .
1 1
Newtons method is given by
1
xk + 1 = xk Ak gk .
Note, however, that we cannot actually perform this algorithm, because the
Hessian matrix is singular. We know from our discussion of this function
in Chapter 8 that this function does not have a strong minimum, but it does
have a weak minimum along the line x 1 = x 2 .
What about steepest descent? If we start from the initial guess, with learn-
ing rate = 0.1 , the first two iterations will be
x 1 = x 0 g 0 = 1 0.1 1 = 0.9 ,
0 1 0.1
x 2 = x 1 g 1 = 0.9 0.1 0.8 = 0.82 .

0.1 0.8 0.18
9-31
The complete trajectory is shown in Figure P9.7. This is a case where the
steepest descent algorithm performs better than Newtons method. Steep-
est descent converges to a minimum point (weak minimum), while New-
tons method fails to converge. In Chapter 12 we will discuss a technique
that combines steepest descent with Newtons method, to overcome the
problem of singular (or almost singular) Hessian matrices.
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure P9.7 Steepest Descent Trajectory for Problem P9.5 with = 0.1
P9.6 Consider the following function:

3 2 2
F x = x1 + x1 x2 x1 x2
i. Perform one iteration of Newtons method from the initial

T
guess x 0 = 1 1 .
ii. Find the second-order Taylor series expansion of F x about

x 0 . Is this quadratic function minimized at the point x 1
found in part (i)? Explain.
i. The gradient of F x is

Fx 2 2
x1 3x 1 + x 2 2x 1 x 2
F x = = ,
2
Fx x 1 2x 1 x 2
x2
9-32
Solved Problems
2
6x 1 2x 2 1 4x 1 x 2
2F x = .
2
1 4x 1 x 2 2x 1
If we evaluate these at the initial guess we find
g 0 = F x = 2 ,
x = x0 1
A 0 = 2F x = 4 3 .
x = x0 3 2
The first iteration of Newtons method is then

1
x1 = x0 A0 g0 = 1 4 3 2 = 0.5882 .
1
1 3 2 1 1.1176
ii. From Eq. (9.40), the second-order Taylor series expansion of F x

about x 0 is
T 1 T
F x = F x 0 + x 0 F x 0 + g 0 x 0 + --- x 0 A 0 x 0 .
2
If we substitute the values for x 0 , g 0 and A 0 , we find

T
1
F x 1 + 2 1 x 1 + --- x 1 4 3 x 1 .
1 2 1 3 2 1
This can be reduced to
1 T
F x 2 + 1 4 x + --- x 4 3 x .
2 3 2
This function has a stationary point at x 1 . The question is whether or not

the stationary point is a strong minimum. This can be determined from the
eigenvalues of the Hessian matrix. If both eigenvalues are positive, it is a
strong minimum. If both eigenvalues are negative, it is a strong maximum.
If the two eigenvalues have opposite signs, it is a saddle point. In this case
the eigenvalues of A 0 are
1 = 5.24 and 2 = 3.24 .
9-33
Therefore the quadratic approximation to F x at x 0 is not minimized at

x 1 , since it is a saddle point. Figure P9.8 displays the contour plots of F x
and its quadratic approximation.
This sort of problem was also illustrated in Figure 9.8 and Figure 9.9. New-
tons method does locate the stationary point of the quadratic approxima-
tion of the function at the current guess. It does not distinguish between
minima, maxima and saddle points.
2 2
1 1
0 0
-1 -1
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
T
Figure P9.8 One Iteration of Newtons Method from x 0 = 1 1
P9.7 Repeat Problem P9.3 (i) using the conjugate gradient algorithm.
Recall that the function to be minimized was
1 T
F x = 0.25 + 2 1 x + --- x 10 2 x .
2 2 4
The gradient at x 0 is
g 0 = F x 0 = Ax 0 + d = 10 2 1 + 2 = 10 .
2 4 1 1 5
The first search direction is then
p 0 = g 0 = 10 .
5
To minimize along a line, for a quadratic function, we can use Eq. (9.31):
9-34
Solved Problems
10
T 10 5
g0 p0 5 125
0 = ---------------
T
- = ----------------------------------------------------- = ------------ = 0.0962 .
p 0 Ap 0 10 2 10
1300
10 5
2 4 5
Therefore the first iteration of conjugate gradient will be
x 1 = x 0 + 0 p 0 = 1 + 0.0962 10 = 0.038 .
1 5 0.519
Now we need to find the second search direction from Eq. (9.60). This re-
quires the gradient at x 1 :
g 1 = F x = 10 2 0.038 + 2 = 0.577 .
x = x1 2 4 0.519 1 1.154
We can now find 1 :
0.577
T 10.577 3.846
g 0 g 1 1.154 1.665
1 = --------------
T
= ----------------------------------------------------------------- = ------------- = 0.0133 ,
g0 g0 10
125
10 5
5
using the method of Polak and Ribire (Eq. (9.63)). (The other two methods
for computing 1 will produce the same results for a quadratic function.
You may want to try them.) The second search direction is then computed
from Eq. (9.60):
p1 = g1 + 1 p0 = 0.577 + 0.0133 10 = 0.444 .

1.154 5 1.220
From Eq. (9.31), the learning rate for the second iteration will be
0.444
0.577 1.154
1.220 1.664
1 = --------------------------------------------------------------------------- = ---------------- = 0.2889 .
5.758
10 2 0.444
0.444 1.220
2 4 1.220
The second step of conjugate gradient is therefore
9-35
x 2 = x 1 + 1 p 1 = 0.038 + 0.2889 0.444 = 0.1667 .

0.519 1.220 0.1667
As expected, the minimum is reached in two iterations. The trajectory is il-

lustrated in Figure P9.9.
2
-1
-2
-2 -1 0 1 2
Figure P9.9 Conjugate Gradient Trajectory for Problem P9.7
P9.8 Show that conjugate vectors are independent.

Suppose that we have a set of vectors, p 0 p 1 p n 1 , which are conju-
gate with respect to the Hessian matrix A . If these vectors are dependent,
then, from Eq. (5.4), it must be true that
n1
aj pj = 0,
j=0
for some set of constants a 0 a 1 a n 1 , at least one of which is nonzero.

T
If we multiply both sides of this equation by p k A , we obtain
n1 n1
aj pj a j pk Apj
T T T
pk A = = a k p k Ap k = 0 ,
j=0 j=0
where the second equality comes from the definition of conjugate vectors in
Eq. (9.52). If A is positive definite (a unique strong minimum exists), then
T
p k Ap k must be strictly positive. This implies that a k must be zero for all k .
Therefore conjugate directions must be independent.
9-36
Epilogue
Epilogue
In this chapter we have introduced three different optimization algorithms:
steepest descent, Newtons method and conjugate gradient. The basis for
these algorithms is the Taylor series expansion. Steepest descent is derived
by using a first-order expansion, whereas Newtons method and conjugate
gradient are designed for second-order (quadratic) functions.
Steepest descent has the advantage that it is very simple, requiring calcu-
lation only of the gradient. It is also guaranteed to converge to a stationary
point if the learning rate is small enough. The disadvantage of steepest de-
scent is that training times are generally longer than for other algorithms.
This is especially true when the eigenvalues of the Hessian matrix, for qua-
dratic functions, have a wide range of magnitudes.
Newtons method is generally much faster than steepest descent. For qua-
dratic functions it will locate a stationary point in one iteration. One disad-
vantage is that it requires calculation and storage of the Hessian matrix,
as well as its inverse. In addition, the convergence properties of Newtons
method are quite complex. In Chapter 12 we will introduce a modification
of Newtons method that overcomes some of the disadvantages of the stan-
dard algorithm.
The conjugate gradient algorithm is something of a compromise between
steepest descent and Newtons method. It will locate the minimum of a qua-
dratic function in a finite number of iterations, but it does not require cal-
culation and storage of the Hessian matrix. It is well suited to problems
with large numbers of parameters, where it is impractical to compute and
store the Hessian.
In later chapters we will apply each of these optimization algorithms to the
training of neural networks. In Chapter 10 we will demonstrate how an ap-
proximate steepest descent algorithm, Widrow-Hoff learning, can be used
to train linear networks. In Chapter 11 we generalize Widrow-Hoff learn-
ing to train multilayer networks. In Chapter 12 the conjugate gradient al-
gorithm, and a variation of Newtons method, are used to speed up the
training of multilayer networks.
9-37
Further Reading
[Batt92] R. Battiti, First and Second Order Methods for Learning:
Between Steepest Descent and Newtons Method, Neural
Computation, Vol. 4, No. 2, pp. 141-166, 1992.
This article reviews the latest developments in uncon-
strained optimization using first and second derivatives.
The techniques discussed are those that are most suitable
for neural network applications.
[Gill81] P. E. Gill, W. Murray and M. H. Wright, Practical Optimi-
zation, New York: Academic Press, 1981.
As the title implies, this text emphasizes the practical im-
plementation of optimization algorithms. It provides moti-
vation for the optimization methods, as well as details of
implementation that affect algorithm performance.
York: McGraw-Hill, 1972.
This is a comprehensive text on nonlinear optimization. It
covers both constrained and unconstrained optimization
problems. The text is very complete, with many examples
worked out in detail.
A very readable text describing the major optimization al-
gorithms, this text emphasizes methods of optimization
rather than existence theorems and proofs of convergence.
Algorithms are presented with intuitive explanations,
along with illustrative figures and examples. Pseudo-code
is presented for most algorithms.
9-38
Exercises
Exercises
E9.1 In Problem P9.1 we found the maximum stable learning rate for the steep-
est descent algorithm when applied to a particular quadratic function. Will
the algorithm always diverge when a larger learning rate is used, or are
there any conditions for which the algorithm will still converge?
E9.2 We want to find the minimum of the following function:
1 T
F x = --- x 6 2 x + 1 1 x .
2 2 6
i. Sketch a contour plot of this function.

ii. Sketch the trajectory of the steepest descent algorithm on the con-
T
tour plot of part (i), if the initial guess is x 0 = 0 0 . Assume a very
small learning rate is used.
iii. Perform two iterations of steepest descent with learning rate

= 0.1 .
iv. What is the maximum stable learning rate?

v. What is the maximum stable learning rate for the initial guess giv-
en in part (ii)? (See Exercise E9.1.)
2+2
ans = vi. Write a MATLAB M-file to implement the steepest descent algo-
4
rithm for this problem, and use it to check your answers to parts (i).
through (v).
E9.3 For the quadratic function
2 2
F x = x 1 + 2x 2 ,
i. Find the minimum of the function along the line
x = 1 + 1 .
1 2
ii. Verify that the gradient of F x at the minimum point from part (i)
is orthogonal to the line along which the minimization occurred.
9-39
E9.4 For the functions given in Exercise E8.3 perform two iterations of the
2+2
steepest descent algorithm with linear minimization, starting from the ini-
T
ans = tial guess x 0 = 1 1 . Write MATLAB M-files to check your answer.
4

2 2
F x = 1 + x 1 + x 2 5 1 + 3x 1 2x 2 .
i. Perform one iteration of Newtons method, starting from the initial

T
guess x 0 = 10 10 .
T
ii. Repeat part (i), starting from the initial guess x 0 = 2 2 .
iii. Find the minimum of the function, and compare with your results
from the previous two parts.
E9.6 Consider the following quadratic function
1 T
F x = --- x 3 2 x + 4 4 x
2 2 0
i. Sketch the contour plot for F x . Show all work.

ii. Take one iteration of Newtons method from the initial guess
T
x0 = 0 0 .
iii. In part (ii), did you reach the minimum of F x ? Explain.
E9.7 Consider the function

4 2
F x = x1 + x2 + 2 x2 1
i. Find the second-order Taylor series approximation of this function

T
about the point x 0 = 1 1 .
ii. Is this point a minimum point? Does it satisfy the first and second
order conditions?
iii. Perform one iteration of Newton's method from the initial guess
T
x 0 = 0.5 0 .
9-40
Exercises
1 T
F x = --- x 7 9 x + 16 8 x
2 9 17
i. Sketch the contour plot for this function.

ii. Take one step of Newtons method from the initial guess
T
x0 = 2 2 .
iii. Did you reach the minimum of the function after the Newton step of
part (ii)? Explain.
iv. From the initial guess in part ii, trace the path of steepest descent,
with very small learning rate, on your contour plot from part (i). Ex-
plain how you determined the path. Will steepest descent eventual-
ly converge to the same result you found in part (ii)? Explain.
2 1 4
F x = 1 + x 1 + x 2 + --- x .
4 1

T
x0 = 2 2 .
ii. Sketch the contour plot of the quadratic approximation in part i.

iii. Perform one iteration of Newtons method on the function F x from
the initial condition x 0 given in part (i). Sketch the path from x 0 to
x 1 on your contour plot from part (ii).
iv. Is the x 1 in part iii. a strong minimum of the quadratic approxima-
tion? Is it a strong minimum of the original function F x ? Explain.
v. Will Newtons method always converge to a strong minimum of
F x , given enough iterations? Will it always converge to a strong
minimum of the quadratic approximation of F x ? Explain your an-
swers in detail.
2+2
ans = E9.10 Recall the function presented in Exercise E8.5. Write MATLAB M-files to
4
implement the steepest descent algorithm and Newtons method for that
function. Test the performance of the algorithms for various initial guesses.
E9.11 Repeat Exercise E9.4 using the conjugate gradient algorithm. Use each of
the three methods (Eq. (9.61)Eq. (9.63)) at least once.
9-41
E9.12 Prove or disprove the following statement:

If p 1 is conjugate to p 2 and p 2 is conjugate to p 3 ,
then p 1 is conjugate to p 3 .
9-42
Objectives
Objectives 10-1
ADALINE Network 10-2
Single ADALINE 10-3
Mean Square Error 10-4
LMS Algorithm 10-7
Analysis of Convergence 10-9
Adaptive Filtering 10-13
Adaptive Noise Cancellation 10-15
Echo Cancellation 10-21
Epilogue 10-40
Exercises 10-42
Objectives
In the previous two chapters we laid the foundation for performance learn-
ing, in which a network is trained to optimize its performance. In this chap-
ter we apply the principles of performance learning to a single-layer linear
neural network.
Widrow-Hoff learning is an approximate steepest descent algorithm, in
which the performance index is mean square error. This algorithm is im-
portant to our discussion for two reasons. First, it is widely used today in
many signal processing applications, several of which we will discuss in
this chapter. In addition, it is the precursor to the backpropagation algo-
rithm for multilayer networks, which is presented in Chapter 11
10-1
Theory and Examples

Bernard Widrow began working in neural networks in the late 1950s, at
about the same time that Frank Rosenblatt developed the perceptron
learning rule. In 1960 Widrow, and his graduate student Marcian Hoff, in-
troduced the ADALINE (ADAptive LInear NEuron) network, and a learn-
ing rule which they called the LMS (Least Mean Square) algorithm
[WiHo60].
Their ADALINE network is very similar to the perceptron, except that its
transfer function is linear, instead of hard-limiting. Both the ADALINE
and the perceptron suffer from the same inherent limitation: they can only
solve linearly separable problems (recall our discussion in Chapters 3 and
4). The LMS algorithm, however, is more powerful than the perceptron
learning rule. While the perceptron rule is guaranteed to converge to a so-
lution that correctly categorizes the training patterns, the resulting net-
work can be sensitive to noise, since patterns often lie close to the decision
boundaries. The LMS algorithm minimizes mean square error, and there-
fore tries to move the decision boundaries as far from the training patterns
as possible.
The LMS algorithm has found many more practical uses than the percep-
tron learning rule. This is especially true in the area of digital signal pro-
cessing. For example, most long distance phone lines use ADALINE
networks for echo cancellation. We will discuss these applications in detail
later in the chapter.
Because of the great success of the LMS algorithm in signal processing ap-
plications, and because of the lack of success in adapting the algorithm to
multilayer networks, Widrow stopped work on neural networks in the early
1960s and began to work full time on adaptive signal processing. He re-
turned to the neural network field in the 1980s and began research on the
use of neural networks in adaptive control, using temporal backpropaga-
tion, a descendant of his original LMS algorithm.
ADALINE Network
The ADALINE network is shown in Figure 10.1. Notice that it has the
same basic structure as the perceptron network we discussed in Chapter 4.
The only difference is that it has a linear transfer function.
10-2
ADALINE Network
Input Linear Neuron
p a
Rx1 W Sx1
SxR n
Sx1
1 b
Sx1
R S
Figure 10.1 ADALINE Network

a = purelin Wp + b = Wp + b . (10.1)
Recall from our discussion of the perceptron network that the ith element
of the network output vector can be written
T T
a i = purelin n i = purelin iw p + b i = iw p + b i , (10.2)
where iw is made up of the elements of the ith row of W :
w i 1
w i 2
iw =

w i R
. (10.3)
Single ADALINE
To simplify our discussion, lets consider a single ADALINE with two in-
puts. The diagram for this network is shown in Figure 10.2.
T T
a = purelin n = purelin 1w p + b = 1w p + b
(10.4)
T
= 1w p + b = w 1 1 p 1 + w 1 2 p 2 + b .
10-3
p1 w1,1
n a
p2

w1,2 b
1
Figure 10.2 Two-Input Linear Neuron

You may recall from Chapter 4 that the perceptron has a decision bound-
ary, which is determined by the input vectors for which the net input n is
zero. Now, does the ADALINE also have such a boundary? Clearly it does.
T
If we set n = 0 then 1w p + b = 0 specifies such a line, as shown in Figure
10.3.
p2
a<0 a>0
-b/w1,2
1w
1 w Tp + b = 0
p1
-b/w1,1
Figure 10.3 Decision Boundary for Two-Input ADALINE

The neuron output is greater than 0 in the gray area. In the white area the
output is less than zero. Now, what does this imply about the ADALINE?
It says that the ADALINE can be used to classify objects into two catego-
ries. However, it can do so only if the objects are linearly separable. Thus,
in this respect, the ADALINE has the same limitation as the perceptron.
Mean Square Error

Now that we have examined the characteristics of the ADALINE network,
we are ready to begin our development of the LMS algorithm. As with the
perceptron rule, the LMS algorithm is an example of supervised training,
in which the learning rule is provided with a set of examples of proper
network behavior:
10-4
Mean Square Error
{p 1, t 1} { p 2, t 2} {p Q, tQ} , (10.5)
where p q is an input to the network, and t q is the corresponding target

output. As each input is applied to the network, the network output is com-
pared to the target.
The LMS algorithm will adjust the weights and biases of the ADALINE in
order to minimize the mean square error, where the error is the difference
between the target output and the network output. In this section we want
to discuss this performance index. We will consider first the single-neuron
case.
To simply our development, we will lump all of the parameters we are
adjusting, including the bias, into one vector:
x = 1w . (10.6)
b
Similarly, we include the bias input 1 as a component of the input vector
z = p . (10.7)
1
Now the network output, which we usually write in the form
T
a = 1w p + b , (10.8)
can be written as
T
a = x z. (10.9)
This allows us to conveniently write out an expression for the ADALINE

Mean Square Error network mean square error:
2 2 2
F x = E e = E t a = E t xTz , (10.10)
where the expectation is taken over all sets of input/target pairs. (Here we
use E to denote expected value. We use a generalized definition of ex-
pectation, which becomes a time-average for deterministic signals. See
[WiSt85].) We can expand this expression as follows:
2
F x = E t 2tx T z + x T zz x
T 10
2 T
= E t 2x T E tz + x T E zz x . (10.11)
10-5
This can be written in the following convenient form:

T T
F x = c 2x h + x Rx , (10.12)
where
2 T
c = E t , h = E tz and R = E zz . (10.13)
Here the vector h gives the cross-correlation between the input vector and
Correlation Matrix its associated target, while R is the input correlation matrix. The diagonal
elements of this matrix are equal to the mean square values of the elements
of the input vectors.
Take a close look at Eq. (10.12), and compare it with the general form of the
quadratic function given in Eq. (8.35) and repeated here:
T 1 T
F x = c + d x + --- x Ax . (10.14)
2
We can see that the mean square error performance index for the ADA-
LINE network is a quadratic function, where
d = 2h and A = 2R . (10.15)
This is a very important result, because we know from Chapter 8 that the
characteristics of the quadratic function depend primarily on the Hessian
matrix A . For example, if the eigenvalues of the Hessian are all positive,
then the function will have one unique global minimum.
In this case the Hessian matrix is twice the correlation matrix R , and it
can be shown that all correlation matrices are either positive definite or
positive semidefinite, which means that they can never have negative
eigenvalues. We are left with two possibilities. If the correlation matrix has
only positive eigenvalues, the performance index will have one unique glo-
bal minimum (see Figure 8.7). If the correlation matrix has some zero
eigenvalues, the performance index will either have a weak minimum (see
Figure 8.9) or no minimum (see Problem P8.7), depending on the vector
d = 2h .
Now lets locate the stationary point of the performance index. From our
previous discussion of quadratic functions we know that the gradient is
F x = c + d x + --- x Ax = d + Ax = 2h + 2Rx .
T 1 T
(10.16)
2
The stationary point of F x can be found by setting the gradient equal to

zero:
2h + 2Rx = 0 . (10.17)
10-6
LMS Algorithm
Therefore, if the correlation matrix is positive definite there will be a

unique stationary point, which will be a strong minimum:
x = R h .
1
(10.18)
It is worth noting here that the existence of a unique solution depends only
on the correlation matrix R . Therefore the characteristics of the input vec-
tors determine whether or not a unique solution exists.
LMS Algorithm
Now that we have analyzed our performance index, the next step is to de-
sign an algorithm to locate the minimum point. If we could calculate the
statistical quantities h and R , we could find the minimum point directly
from Eq. (10.18). If we did not want to calculate the inverse of R , we could
use the steepest descent algorithm, with the gradient calculated from Eq.
(10.16). In general, however, it is not desirable or convenient to calculate h
and R . For this reason we will use an approximate steepest descent algo-
rithm, in which we use an estimated gradient.
The key insight of Widrow and Hoff was that they could estimate the mean
square error F x by
F x = t k a k = e k ,
2 2
(10.19)
where the expectation of the squared error has been replaced by the
squared error at iteration k . Then, at each iteration we have a gradient es-
timate of the form:
2
F x = e k . (10.20)
Stochastic Gradient This is sometimes referred to as the stochastic gradient. When this is used
in a gradient descent algorithm, it is referred to as on-line or incremental
learning, since the weights are updated as each input is presented to the
network.
2
The first R elements of e k are derivatives with respect to the network
weights, while the R + 1 st element is the derivative with respect to the bi-
as. Thus we have
2
2 e k e k
e k j = --------------- = 2e k ------------- for j = 1 2 R , (10.21)
w 1 j w 1 j
and
2
2 e k e k
e k R + 1 = --------------- = 2e k ------------- . (10.22)
b b
10-7
Now consider the partial derivative terms at the ends of these equations.
First evaluate the partial derivative of e k with respect to the weight w 1 j :
e k t k a k T
------------- = ---------------------------------- = t k 1w p k + b
w 1 j w 1 j w 1 j
R

= t k w 1 i p i k + b
w 1 j
i=1 (10.23)
where p i k is the ith element of the input vector at the kth iteration. This
simplifies to
e k
------------- = p j k . (10.24)
w 1 j
In a similar way we can obtain the final element of the gradient:
e k
------------- = 1 . (10.25)
b
Note that p j k and 1 are the elements of the input vector z , so the gradi-
ent of the squared error at iteration k can be written
2
F x = e k = 2e k z k . (10.26)
Now we can see the beauty of approximating the mean square error by the
single error at iteration k , as in Eq. (10.19). To calculate this approximate
gradient we need only multiply the error times the input.
This approximation to F x can now be used in the steepest descent algo-
rithm. From Eq. (9.10) the steepest descent algorithm, with constant learn-
ing rate, is
x k + 1 = x k F x . (10.27)
x = xk

If we substitute F x , from Eq. (10.26), for F x we find
x k + 1 = x k + 2e k z k , (10.28)
or
1w k + 1 = 1w k + 2e k p k , (10.29)
and
b k + 1 = b k + 2e k . (10.30)
10-8
Analysis of Convergence
These last two equations make up the least mean square (LMS) algorithm.
This is also referred to as the delta rule or the Widrow-Hoff learning algo-
rithm.
The preceding results can be modified to handle the case where we have
multiple outputs, and therefore multiple neurons, as in Figure 10.1. To up-
date the ith row of the weight matrix use
iw k + 1 = iw k + 2e i k p k , (10.31)
where e i k is the ith element of the error at iteration k . To update the ith
element of the bias we use
b i k + 1 = b i k + 2e i k . (10.32)
LMS Algorithm The LMS algorithm can be written conveniently in matrix notation:
T
W k + 1 = W k + 2e k p k , (10.33)
and
b k + 1 = b k + 2e k . (10.34)
Note that the error e and the bias b are now vectors.
The stability of the steepest descent algorithm was investigated in Chapter
9. There we found that the maximum stable learning rate for quadratic
functions is 2 max , where max is the largest eigenvalue of the Hessian
matrix. Now we want to investigate the convergence of the LMS algorithm,
which is approximate steepest descent. We will find that the result is the
same.
To begin, note that in the LMS algorithm, Eq. (10.28), x k is a function only
of z k 1 z k 2 z 0 . If we assume that successive input vectors are
statistically independent, then x k is independent of z k . We will show in
the following development that for stationary input processes meeting this
condition, the expected value of the weight vector will converge to
x = R h .
1
(10.35)
2
This is the minimum mean square error E ek solution, as we saw in Eq.
(10.18).
Recall the LMS algorithm (Eq. (10.28)):
x k + 1 = x k + 2e k z k . (10.36)
10-9
Now take the expectation of both sides:
E x k + 1 = E x k + 2E e k z k . (10.37)
Substitute t k x kT z k for the error to give

T
E x k + 1 = E x k + 2 E t k z k E x k z k z k . (10.38)
T T
Finally, substitute z k x k for x k z k and rearrange terms to give
T
E x k + 1 = E x k + 2 E t k z k E z k z k x k . (10.39)
Since x k is independent of z k :
E x k + 1 = E x k + 2 h RE x k . (10.40)
This can be written as
E x k + 1 = I 2R E x k + 2h . (10.41)
This dynamic system will be stable if all of the eigenvalues of I 2R fall

inside the unit circle (see [Brog91]). Recall from Chapter 9 that the eigen-
values of I 2R will be 1 2 i , where the i are the eigenvalues of R .
Therefore, the system will be stable if
1 2 i 1 . (10.42)
Since i 0 , 1 2 i is always less than 1. The condition on stability is

therefore
1 i for all i , (10.43)
or
0 1 max . (10.44)
Note that this condition is equivalent to the condition we derived in Chap-

ter 9 for the steepest descent algorithm, although in that case we were us-
ing the eigenvalues of the Hessian matrix A . Now we are using the
eigenvalues of the input correlation matrix R . (Recall that A = 2R .)
If this condition on stability is satisfied, the steady state solution is
E x ss = I 2R E x ss + 2h , (10.45)
or
E x ss = R h = x .
1
(10.46)
10-10
Thus the LMS solution, obtained by applying one input vector at a time, is
the same as the minimum mean square error solution of Eq. (10.18).
2 To test the ADALINE network and the LMS algorithm consider again the
+2
apple/orange recognition problem originally discussed in Chapter 3. For
simplicity we will assume that the ADALINE network has a zero bias.
The LMS weight update algorithm of Eq. (10.29) will be used to calculate
the new weights at each step in the network training:
T
W k + 1 = W k + 2e k p k . (10.47)
First lets compute the maximum stable learning rate . We can get such
a value by finding the eigenvalues of the input correlation matrix. Recall
that the orange and apple vectors and their associated targets are
1 1

p1 = 1 t1 = 1 p2 = 1 t2 = 1 . (10.48)

1 1
If we assume that the input vectors are generated randomly with equal
probability, we can compute the input correlation matrix:
T 1 T 1 T
R = E pp = --- p 1 p 1 + --- p 2 p 2
2 2
1 1 1 0 1
1 1
= --- 1 1 1 1 + --- 1 1 1 1 = 0 1 0 . (10.49)
2 2
1 1 1 0 1
The eigenvalues of R are
1 = 1.0 2 = 0.0 3 = 2.0 . (10.50)
Thus, the maximum stable learning rate is
1 1
----------- = ------- = 0.5 . (10.51)
max 2.0
To be conservative we will pick = 0.2 . (Note that in practical applica-

tions it might not be practical to calculate R , and could be selected by
trial and error. Other techniques for choosing are given in [WiSt85].)
We will start, arbitrarily, with all the weights set to zero, and then will ap-
ply inputs p 1 , p 2 , p 1 , p 2 , etc., in that order, calculating the new weights
after each input is presented. (The presentation of the weights in alternat-
10-11
ing order is not necessary. A random sequence would be fine.) Presenting

p 1 , the orange, and using its target of -1 we get
1
a 0 = W 0 p 0 = W 0 p 1 = 0 0 0 1 = 0 , (10.52)
1
and
e 0 = t 0 a 0 = t1 a 0 = 1 0 = 1 . (10.53)
Now we can calculate the new weight matrix:

T
W 1 = W 0 + 2e 0 p 0
T
1
= 0 0 0 + 2 0.2 1 1 = 0.4 0.4 0.4 .
(10.54)
1
According to plan, we will next present the apple, p 2 , and its target of 1:
1
a 1 = W 1 p 1 = W 1 p 2 = 0.4 0.4 0.4 1 = 0.4 , (10.55)
1
and so the error is
e 1 = t 1 a 1 = t 2 a 1 = 1 0.4 = 1.4 . (10.56)
Now we calculate the new weights:

T
W 2 = W 1 + 2e 1 p 1
T
1
= 0.4 0.4 0.4 + 2 0.2 1.4 1 = 0.16 0.96 0.16 .
1
(10.57)
Next we present the orange again:
1
a 2 = W 2 p 2 = W 2 p 1 = 0.16 0.96 0.16 1 = 0.64 . (10.58)
1
10-12
Adaptive Filtering
The error is
e 2 = t 2 a 2 = t 1 a 2 = 1 0.64 = 0.36 . (10.59)
The new weights are

T
W 3 = W 2 + 2e 2 p 2 = 0.016 1.1040 0.0160 . (10.60)
If we continue this procedure, the algorithm converges to
W = 0 1 0 . (10.61)
Compare this result with the result of the perceptron learning rule in
Chapter 4. You will notice that the ADALINE has produced the same deci-
sion boundary that we designed in Chapter 3 for the apple/orange problem.
This boundary falls halfway between the two reference patterns. The per-
ceptron rule did not produce such a boundary. This is because the percep-
tron rule stops as soon as the patterns are correctly classified, even though
some patterns may be close to the boundaries. The LMS algorithm mini-
mizes the mean square error. Therefore it tries to move the decision bound-
aries as far from the reference patterns as possible.
Adaptive Filtering
As we mentioned at the beginning of this chapter, the ADALINE network
has the same major limitation as the perceptron network; it can only solve
linearly separable problems. In spite of this, the ADALINE has been much
more widely used than the perceptron network. In fact, it is safe to say that
it is one of the most widely used neural networks in practical applications.
One of the major application areas of the ADALINE has been adaptive fil-
tering, where it is still used extensively. In this section we will demonstrate
an adaptive filtering example.
In order to use the ADALINE network as an adaptive filter, we need to in-
Tapped Delay Line troduce a new building block, the tapped delay line. A tapped delay line
with R outputs is shown in Figure 10.4.
The input signal enters from the left. At the output of the tapped delay line
we have an R -dimensional vector, consisting of the input signal at the cur-
rent time and at delays of from 1 to R 1 time steps.
10-13
y(k) p1(k) = y(k)
D
p2(k) = y(k - 1)
D
pR(k) = y(k - R + 1)
Figure 10.4 Tapped Delay Line

If we combine a tapped delay line with an ADALINE network, we can cre-
Adaptive Filter ate an adaptive filter, as is shown in Figure 10.5. The output of the filter is
given by
R
a k = purelin Wp + b = w 1 i y k i + 1 + b . (10.62)
i=1
Inputs ADALINE
y(k)
w1,1
D
w1,2
n(k) a(k)
D
SxR
b
1
D w1,R
a(k) = purelin (Wp(k) + b)
Figure 10.5 Adaptive Filter ADALINE
10-14
Adaptive Filtering
If you are familiar with digital signal processing, you will recognize the net-
work of Figure 10.5 as a finite impulse response (FIR) filter [WiSt85]. It is
beyond the scope of this text to review the field of digital signal processing,
but we can demonstrate the usefulness of this adaptive filter through a
simple, but practical, example.
Adaptive Noise Cancellation

2 An adaptive filter can be used in a variety of novel ways. In the following
+2
example we will use it for noise cancellation. Take some time to look at this
example, for it is a little different from what you might expect. For in-
stance, the output error that the network tries to minimize is actually an
approximation to the signal we are trying to recover!
Lets suppose that a doctor, in trying to review the electroencephalogram
(EEG) of a distracted graduate student, finds that the signal he would like
to see has been contaminated by a 60-Hz noise source. He is examining the
patient on-line and wants to view the best signal that can be obtained. Fig-
ure 10.6 shows how an adaptive filter can be used to remove the contami-
nating signal.
EEG Signal Contaminated Restored Signal

(random) s t Signal e
+
- "Error"
Contaminating
Noise Adaptively Filtered
Noise to Cancel
m
Contamination
Noise Path
Filter
Graduate
Student
v a
Adaptive
Filter
60-Hz
Noise Source
Adaptive Filter Adjusts to Minimize Error (and in doing
this removes 60-Hz noise from contaminated signal)
Figure 10.6 Noise Cancellation System

As shown, a sample of the original 60-Hz signal is fed to an adaptive filter,
whose elements are adjusted so as to minimize the error e . The desired
output of the filter is the contaminated EEG signal t . The adaptive filter
will do its best to reproduce this contaminated signal, but it only knows
about the original noise source, v . Thus, it can only reproduce the part of t
that is linearly correlated with v , which is m . In effect, the adaptive filter
will attempt to mimic the noise path filter, so that the output of the filter
10-15
a will be close to the contaminating noise m . In this way the error e will
be close to the original uncontaminated EEG signal s .
In this simple case of a single sine wave noise source, a neuron with two
weights and no bias is sufficient to implement the filter. The inputs to the
filter are the current and previous values of the noise source. Such a two-
input filter can attenuate and phase-shift the noise v in the desired way.
The filter is shown in Figure 10.7.
Inputs ADALINE
v(k) w1,1
n(k) a(k)
D w1,2
SxR
a(k) = w1,1 v(k) + w1,2 v(k - 1)
Figure 10.7 Adaptive Filter for Noise Cancellation

We can apply the mathematical relationships developed in the previous
sections of this chapter to analyze this system. In order to do so, we will
first need to find the input correlation matrix R and the input/target cross-
correlation vector h :
T
R = zz and h = E tz . (10.63)
In our case the input vector is given by the current and previous values of
the noise source:
z k = vk , (10.64)
v k 1
while the target is the sum of the current signal and filtered noise:
t k = s k + mk . (10.65)
Now expand the expressions for R and h to give
2
R = Ev k E v k v k 1 , (10.66)
2
E v k 1 v k E v k 1
10-16
Adaptive Filtering
and
h = E s k + m k v k . (10.67)
E s k + m k v k 1
To obtain specific values for these two quantities we must define the noise
signal v , the EEG signal s and the filtered noise m . For this exercise we
will assume: the EEG signal is a white (uncorrelated from one time step to
the next) random signal uniformly distributed between the values -0.2 and
+0.2, the noise source (60-Hz sine wave sampled at 180 Hz) is given by
v k = 1.2 sin --------- ,

2k
(10.68)
3
and the filtered noise that contaminates the EEG is the noise source atten-
uated by a factor of 10 and shifted in phase by 2 :
2k
m k = 0.12 sin --------- + --- . (10.69)
3 2
Now calculate the elements of the input correlation matrix R :

3
2
21 sin --------
2k

2 2
E v k = 1.2 --- - = 1.2 0.5 = 0.72 , (10.70)
3 3
k=1
2 2
E v k 1 = E v k = 0.72 , (10.71)
3
2 k 1
1.2 sin --------- 1.2 sin -----------------------
1 2k
E v k v k 1 = ---
3 3 3
k=1
= 1.2 0.5 cos ------ = 0.36

2 2
3
(10.72)
(where we have used some trigonometric identities).
Thus R is
R = 0.72 0.36 . (10.73)

0.36 0.72
The terms of h can be found in a similar manner. We will consider the top
term in Eq. (10.67) first:
E s k + m k v k = E s k v k + E m k v k . (10.74)
10-17
Here the first term on the right is zero because s k and v k are indepen-
dent and zero mean. The second term is also zero:
3
2k
0.12 sin --------- + --- 1.2 sin --------- = 0
1 2k
E m k v k = --- (10.75)
3 3 2 3
k=1
Thus, the first element of h is zero.

Next consider the second element of h :
E s k + m k v k 1 = E s k v k 1
+ E m k v k 1 . (10.76)
As with the first element of h , the first term on the right is zero because
s k and v k 1 are independent and zero mean. The second term is eval-
uated as follows:
3
2k 2 k 1
0.12 sin --------- + --- 1.2 sin -----------------------
1
E m k v k 1 = ---
3 3 2 3
k=1
= 0.0624 . (10.77)
Thus, h is
h = 0 . (10.78)
0.0624
The minimum mean square error solution for the weights is given by Eq.
(10.18):
1
x = R h = 0.72 0.36 0 = 0.0578 .
1
(10.79)
0.36 0.72 0.0624 0.1156
Now, what kind of error will we have at the minimum solution? To find this
error recall Eq. (10.12):
T T
F x = c 2x h + x Rx . (10.80)
We have just found x , h and R , so we only need to find c :

2 2
c = E t k= E s k + m k
2 2
= E s k + 2E s k m k + E m k . (10.81)
10-18
Adaptive Filtering
The middle term is zero because s k and m k are independent and zero
mean. The first term, the mean squared value of the random signal, can be
calculated as follows:
0.2
1 1 0.2

2 2 3
E s k = ------- s ds = ---------------s = 0.0133 . (10.82)
0.4 3 0.4 0.2
0.2
The mean square value of the filtered noise is

3 2
2
0.12 sin -----3- + --2- = 0.0072 ,
2 1
E m k = --- (10.83)
3
k=1
so that
c = 0.0133 + 0.0072 = 0.0205 . (10.84)
Substituting x , h and R into Eq. (10.80), we find that the minimum mean
square error is
F x = 0.0205 2 0.0072 + 0.0072 = 0.0133 . (10.85)
The minimum mean square error is the same as the mean square value of
the EEG signal. This is what we expected, since the error of this adaptive
noise canceller is in fact the reconstructed EEG signal.
Figure 10.8 illustrates the trajectory of the LMS algorithm in the weight
space with learning rate = 0.1 . The system weights w 1 1 and w 1 2 in this
simulation were initialized arbitrarily to 0 and -2, respectively. You can
see from this figure that the LMS trajectory looks like a noisy version of
steepest descent.
2
w 1 2 0
-1
-2
-2 -1 0 1 2
w 1 1
Figure 10.8 LMS Trajectory for = 0.1
10-19
Note that the contours in this figure reflect the fact that the eigenvalues
and eigenvectors of the Hessian matrix ( A = 2R ) are
1 = 2.16 , z 1 = 0.7071 , 2 = 0.72 , z 2 = 0.7071 . (10.86)

0.7071 0.7071
(Refer back to our discussion in Chapter 8 on the eigensystem of the Hes-

sian matrix.)
If the learning rate is decreased, the LMS trajectory is smoother than that
shown in Figure 10.8, but the learning proceeds more slowly. If the learn-
ing rate is increased, the trajectory is more jagged and oscillatory. In fact,
as noted earlier in this chapter, if the learning rate is increased too much
the system does not converge at all. The maximum stable learning rate is
2 2.16 = 0.926 .
In order to judge the performance of our noise canceller, consider Figure
10.9. This figure illustrates how the filter adapts to cancel the noise. The
top graph shows the restored and original EEG signals. At first the re-
stored signal is a poor approximation of the original EEG signal. It takes
about 0.2 second (with = 0.1 ) for the filter to adjust to give a reasonable
restored signal. The mean square difference between the original and re-
stored signal over the last half of the experiment was 0.002. This compares
favorably with the signal mean square value of 0.0133. The difference be-
tween the original and restored signal is shown in the lower graph.
Original and Restored EEG Signals

2 Original and Restored EEG Signals
-1
-2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
2 EEG Signal Minus Restored Signal
-1
-2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time
Figure 10.9 Adaptive Filter Cancellation of Contaminating Noise
10-20
Adaptive Filtering
You might wonder why the error does not go to zero. This is because the
LMS algorithm is approximate steepest descent; it uses an estimate of the
gradient, not the true gradient, to update the weights. The estimate of the
gradient is a noisy version of the true gradient. This will cause the weights
to continue to change slightly, even after the mean square error is at the
minimum point. You can see this effect in Figure 10.8.
To experiment with the use of this adaptive noise cancellation filter, use the
MATLAB Neural Network Design Demonstration Adaptive Noise Cancel-
lation (nnd10nc). A more complex noise source and actual EEG data are used
in the Demonstration Electroencephalogram Noise Cancellation (nnd10eeg).
Echo Cancellation
Another very important practical application of adaptive noise cancellation
is echo cancellation. Echoes are common in long distance telephone lines
because of impedance mismatch at the hybrid device that forms the junc-
tion between the long distance line and the customers local line. You may
have experienced this effect on international telephone calls.
Figure 10.10 illustrates how an adaptive noise cancellation filter can be
used to reduce these echoes [WiWi85]. At the end of the long distance line
the incoming signal is sent to an adaptive filter, as well as to the hybrid de-
vice. The target output of the filter is the output of the hybrid. The filter
thus tries to cancel the part of the hybrid output that is correlated with the
input signal the echo.
+
Transmission
Line
-
Adaptive Adaptive
Phone Hybrid Hybrid Phone
Filter Filter
-
Transmission
Line +
Figure 10.10 Echo Cancellation System
10-21
Summary of Results
ADALINE
Input Linear Neuron
p a
Rx1 W Sx1
SxR n
Sx1
1 b
Sx1
R S
Mean Square Error

2 2 2
F x = E e = E t a = E t xTz
T T
F x = c 2x h + x Rx ,
2 T
c = E t , h = E tz and R = E zz
Unique minimum, if it exists, is x = R h .

1
Where x = 1w and z = p .
b 1
LMS Algorithm
T
W k + 1 = W k + 2e k p k
b k + 1 = b k + 2e k
Convergence Point
x = R h
1
10-22
Summary of Results
Stable Learning Rate

0 1 max where max is the maximum eigenvalue of R
Tapped Delay Line

y(k) p1(k) = y(k)
D
p2(k) = y(k - 1)
D
pR(k) = y(k - R + 1)
Adaptive Filter ADALINE

Inputs ADALINE
y(k)
w1,1
D
w1,2
n(k) a(k)
D
SxR
b
1
D w1,R

R
10
a k = purelin Wp + b = w 1 i y k i + 1 + b
i=1
10-23
Solved Problems
P10.1 Consider the ADALINE filter in Figure P10.1.
Inputs ADALINE
y(k)
w1,1
D w1,2 n(k) a(k)

SxR
D w1,3
Figure P10.1 ADALINE Filter

Suppose that
w 1 1 = 2 w 1 2 = 1 w 1 3 = 3 ,
and the input sequence is
y k = 0 0 0 5 4 0 0 0
where y 0 = 5 , y 1 = 4 , etc.
i. What is the filter output just prior to k = 0 ?
ii. What is the filter output from k = 0 to k = 5 ?
iii. How long does y 0 contribute to the output?
i. Just prior to k = 0 three zeros have entered the filter, and the output is
zero.
ii. At k = 0 the digit 5 has entered the filter, and it will be multiplied by
w 1 1 , which has the value 2, so that a 0 = 10 . This can be viewed as the
matrix operation:
10-24
Solved Problems
y0 5
a 0 = Wp 0 = w 1 1 w 1 2 w 1 3 y 1 = 2 1 3 0 = 10 .
y 2 0
Similarly, one can calculate the next outputs as
4
a 1 = Wp 1 = 2 1 3 5 = 13
0
0
a 2 = Wp 2 = 2 1 3 4 = 19
5
0 0
a 3 = Wp 3 = 2 1 3 0 = 12 , a 4 = Wp 4 = 2 1 3 0 = 0 .
4 0
All remaining outputs will be zero.

iii. The effects of y 0 last from k = 0 through k = 2 , so it will have an in-
fluence for three time intervals. This corresponds to the length of the im-
pulse response of this filter.
P10.2 Suppose that we want to design an ADALINE network to distin-

guish between various categories of input vectors. Let us first try
the categories listed below:
T T
Category I: p1 = 1 1 and p 2 = 1 1
T
Category II: p3 = 2 2 .
i. Can an ADALINE network be designed to make such a dis-

tinction?
ii. If the answer to part (i) is yes, what set of weights and bias
might be used?
Next consider a different set of categories.
10
T T
Category III: p1 = 1 1 and p 2 = 1 1
10-25
T
Category IV: p3 = 1 0 .
iii. Can an ADALINE network be designed to make such a dis-

tinction?
iv. If the answer to part (iii) is yes, what set of weights and bias
might be used?
i. The input vectors are plotted in Figure P10.2.
p3
I p1 II
p2
Figure P10.2 Input Vectors for Problem P10.1 (i)

The blue line in this figure is a decision boundary that separates the two
categories successfully. Since they are linearly separable, an ADALINE
network will do the job.
ii. The decision boundary passes through the points 3 0 and 0 3 . We
know these points to be the intercepts b w 1 1 and b w 1 2 . Thus, a solu-
tion
b = 3 , w 1 1 = 1 , w 1 2 = 1 ,
is satisfactory. Note that if the output of the ADALINE is positive or zero

the input vector is classified as Category I, and if the output is negative the
input vector is classified as Category II. This solution also provides for er-
ror, since the decision boundary bisects the line between p 1 and p 3 .
iii. The input vectors to be distinguished are shown in Figure P10.3. The
vectors in the figure are not linearly separable, so an ADALINE network
cannot distinguish between them.
iv. As noted in part (iii), an ADALINE cannot do the job, so there are no
values for the weights and bias that are satisfactory.
10-26
Solved Problems
p1
p3
p2
Figure P10.3 Input Vectors for Problem P10.1 (iii)
P10.3 Suppose that we have the following input/target pairs:
1 t = 1 , p = 1 t = 1 .
p1 = 1 2 2
1 1
These patterns occur with equal probability, and they are used to
train an ADALINE network with no bias. What does the mean
square error performance surface look like?
First we need to calculate the various terms of the quadratic function. Re-
call from Eq. (10.11) that the performance index can be written as
T T
F x = c 2x h + x Rx .
Therefore we need to calculate c , h and R .

The probability of each input occurring is 0.5, so the probability of each tar-
get is also 0.5. Thus, the expected value of the square of the targets is
2 2 2
c = E t = 1 0.5 + 1 0.5 = 1 .
In a similar way, the cross-correlation between the input and the target can
be calculated:
h = E tz = 0.5 1 1 + 0.5 1 1 = 0 .
1 1 1
Finally, the input correlation matrix R is
10-27
T T T
R = E zz = p 1 p 1 0.5 + p 2 p 2 0.5
= 0.5 1 1 1 + 1 1 1 = 1 0
1 1 0 1
Therefore the mean square error performance index is

T T
F x = c 2x h + x Rx
w 1 1
= 1 2 w 1 1 w 1 2 0 + w 1 1 w 1 2 1 0
1 0 1 w 1 2
2 2
= 1 2w 1 2 + w 1 1 + w 1 2
The Hessian matrix of F x , which is equal to 2R , has both eigenvalues at

2. Therefore the contours of the performance surface will be circular. To
find the center of the contours (the minimum point), we need to solve Eq.
(10.18):
1
x = R h = 1 0 0 = 0 .
1
0 1 1 1
Thus we have a minimum at w 1 1 = 0 , w 1 2 = 1 . The resulting mean

square error performance surface is shown in Figure P10.4.
4
w 1 2 1
-1
-2
-3 -2 -1 0 1 2 3
w 1 1
Figure P10.4 Contour Plot of F x for Problem P10.3
10-28
Solved Problems
P10.4 Consider the system of Problem P10.3 again. Train the network us-
ing the LMS algorithm, with the initial guess set to zero and a
learning rate = 0.25 . Apply each reference pattern only once
during training. Draw the decision boundary at each stage.
Assume the input vector p 1 is presented first. The output, error and new
weights are calculated as follows:
a 0 = purelin 0 0 1 = 0 ,
1
e 0 = t0 a0 = 1 0= 1 ,
a>0
W 1 = W 0 + 2e 0 p 0 = 0 0 + 2 --- 1 1 1 = 1
T 1 1
4 --- --- .
2 2
1w
The decision boundary associated with these weights is shown to the left.
Now apply the second input vector:
a<0
1 1
a 1 = purelin 1
--- --- = 0 ,
2 2 1
e 1 = t 1 a 1 = 1 0 = 1 ,
W 2 = W 1 + 2e 1 p 1 = 1
T 1 1- 1
1w
--- --- + 2 --
4 1 1 = 0 1 .
2 2
a>0
The decision boundary associated with these weights is shown to the left.
a<0 This boundary shows real promise. It is exactly halfway between the input
vectors. You might verify for yourself that each input vector, when applied,
yields its correct associated target. (What set of weights would be optimal
if the targets associated with the two input vectors were exchanged?)
P10.5 Now consider the convergence of the system of Problems P10.3 and
P10.4. What is the maximum stable learning rate for the LMS algo-
rithm?
The LMS convergence is determined by the learning rate , which should
2+2
not exceed the reciprocal of the largest eigenvalue of R . We can determine
ans = this limit by finding these eigenvalues using MATLAB.
4
10-29
[V,D] = eig (R)

V =
1 0
0 1
D=
1 0
0 1
The diagonal terms of matrix D give the eigenvalues, 1 and 1, while the col-
umns of V show the eigenvectors. Note, incidentally, that the eigenvectors
have the same direction as those shown in Figure P10.4.
The largest eigenvalue, max = 1 , sets the upper limit on the learning rate
at
1 max = 1 1 = 1 .
The suggested learning rate in the previous problem was 0.25, and you
found (perhaps) that the LMS algorithm converged quickly. What do you
suppose happens when the learning rate is 1.0 or larger?
P10.6 Consider the adaptive filter ADALINE shown in Figure P10.5. The
purpose of this filter is to predict the next value of the input signal
from the two previous values. Suppose that the input signal is a
stationary random process, with autocorrelation function given by
C y n = E y k y k + n
Cy 0 = 3 , Cy 1 = 1 , Cy 2 = 1 .
i. Sketch the contour plot of the performance index (mean

square error).
ii. What is the maximum stable value of the learning rate ( )
for the LMS algorithm?
iii. Assume that a very small value is used for . Sketch the path
of the weights for the LMS algorithm, starting with initial
T
guess W 0 = 0.75 0 . Explain your procedure for sketch-
ing the path.
10-30
Solved Problems
Inputs ADALINE
t(k) = y(k)
y(k) D w1,1
n(k) a(k)
+
D w1,2
SxR
- e(k)
a(k) = w1,1 y(k - 1) + w1,2 y(k - 2)
Figure P10.5 Adaptive Predictor

i. To sketch the contour plot we first need to find the performance index
and the eigenvalues and eigenvectors of the Hessian matrix. First note that
the input vector is given by
z k = pk = yk 1 .
yk 2
Now consider the performance index. Recall from Eq. (10.12) that
T T
F x = c 2x h + x Rx .
We can calculate the constants in the performance index as shown below:

2 2
c = E t k = E y k = Cy 0 = 3 ,
2
T
R = E zz = E y k 1 y k 1 y k 2
2
y k 1 y k 2 y k 2
Cy 0 Cy 1 3 1
= =
Cy 1 Cy 0 1 3
Cy 1
h = E t z = E y k y k 1 = = 1 .
y k y k 2 Cy 2 1
The optimal weights are
10-31
1
x = R h =
1 3 1 1 = 3 8 1 8 1 = 1 2 .
1 3 1 4 8 3 8 1 1 2
The Hessian matrix is
2
F x = A = 2R = 6 2 .
2 6
Now we can get the eigenvalues:
6 2 2
= 12 + 32 = 8 4 .
A I =
2 6
Thus,
1 = 4 2 = 8 .
To find the eigenvectors we use
A I v = 0 .
For 1 = 4 ,
2 2 v = 0 v1 = 1 ,
1
2 2 1
and for 2 = 8 ,
2 2 v = 0 v2 = 1 .
2
2 2 1
Therefore the contours of F x will be elliptical, with the long axis of each
ellipse along the first eigenvector, since the first eigenvalue has the small-
est magnitude. The ellipses will be centered at x . The contour plot is
shown in Figure P10.6.
10-32
Solved Problems
w 1 2 0
-1
-2
-2 -1 0 1 2
w 1 1
Figure P10.6 Error Contour for Problem P10.6

You might check your sketch by writing a MATLAB M-file to plot the con-
tours.
ii. The maximum stable learning rate is the reciprocal of the maximum
eigenvalue of R , which is the same as twice the reciprocal of the largest
2
eigenvalue of the Hessian matrix F x = A :
2 max = 2 8 = 0.25 .
iii. The LMS algorithm is approximate steepest descent, so the trajectory

for small learning rates will move perpendicular to the contour lines, as
Contour Plot
1
0.5
w 1 2 0
x2
-0.5
-1
-1 -0.5 0 0.5 1
w 1x1 1
Figure P10.7 LMS Weight Trajectory
10-33
P10.7 The pilot of an airplane is talking into a microphone in his cockpit.

The sound received by the air traffic controller in the tower is gar-
bled because the pilots voice signal has been contaminated by en-
gine noise that reaches his microphone. Can you suggest an
adaptive ADALINE filter that might help reduce the noise in the
signal received by the control tower? Explain your system.
The engine noise that has been inadvertently added to the microphone in-
put can be minimized by using the adaptive filtering system shown in Fig-
ure P10.8. A sample of the engine noise is supplied to an adaptive filter
through a microphone in the cockpit. The desired output of the filter is the
contaminated signal coming from the pilots microphone. The filter at-
tempts to reduce the error signal to a minimum. It can do this only by
subtracting the component of the contaminated signal that is linearly cor-
related with the engine noise (and presumably uncorrelated with the pilots
voice). The result is that a clear voice signal is sent to the control tower, in
spite of the fact that the engine noise got into the pilots microphone along
with his voice signal. (See [WiSt85] for discussion of similar noise cancella-
tion systems.)
Contaminated Restored Signal

s t Signal e
Pilot Voice + "Error"
Signal -
Adaptively Filtered
Contaminating m Noise to Cancel
Noise Contamination
Noise Path
Filter
v a
Adaptive
Filter
Airplane Engine
Noise Source
Figure P10.8 Filtering Engine Noise from Pilots Voice Signal
P10.8 This is a classification problem like that described in Problems

P4.3 and P4.5, except that here we will use an ADALINE network
and the LMS learning rule rather than the perceptron learning
rule. First we will describe the problem.
We have a classification problem with four classes of input vector.
The four classes are
10-34
Solved Problems

class 1: p 1 = 1 p 2 = 1 , class 2: p 3 = 2 p 4 = 2 ,
1 2 1 0
1 p = 2 .
class 3: p 5 = 1 p 6 = 2 , class 4: p7 = 8
2 1 1 2
Train an ADALINE network to solve this problem using the LMS

learning rule. Assume that each pattern occurs with probability
18.
Lets begin by displaying the input vectors, as in Figure P10.9. The light
circles indicate class 1 vectors, the light squares indicate class 2 vec-
tors, the dark circles indicate class 3 vectors, and the dark squares in-
dicate class 4 vectors. These input vectors can be plotted as shown in
Figure P10.9.
1
3
Figure P10.9 Input Vectors for Problem P10.8

We will use target vectors similar to the ones we introduced in Problem
P4.3, except that we will replace any targets of 0 by targets of -1. (The per-
ceptron could only output 0 or 1.) Thus, the training set will be:
1 t = 1 p = 1 t = 1 p = 2 t = 1
p1 = 1 2 2 3 3
1 1 2 1 1 1
2 t = 1 p = 1 t = 1 p = 2 t = 1
p4 = 4 5 5 6 6
0 1 2 1 1 1
1 t = 1 p = 2 t = 1
p7 = 7 8 8
1 1 2 1
10-35
Also, we will begin as in Problem P4.5 with the following initial weights
and biases:
W 0 = 1 0 , b 0 = 1 .
0 1 1
Now we are almost ready to train an ADALINE network using the LMS
rule. We will use a learning rate of = 0.04 , and we will present the input
vectors in order according to their subscripts. The first iteration is

a 0 = purelin (W 0 p 0 + b 0 ) = purelin 1 0 1 + 1 = 2
0 1 1 1 2
e 0 = t 0 a 0 = 1 2 = 3
1 2 3
T
W 1 = W 0 + 2e 0 p 0
= 1 0 + 2 0.04 3 1 1 = 0.76 0.24

0 1 3 0.24 0.76
b 1 = b 0 + 2e 0 = 1 + 2 0.04 3 = 0.76 .
1 3 0.76
The second iteration is
a 1 = purelin (W 1 p 1 + b 1 )

= purelin 0.76 0.24 1 + 0.76 = 1.04
0.24 0.76 2 0.76 2.04
e 1 = t 1 a 1 = 1 1.04 = 2.04
1 2.04 3.04
T
W 2 = W 1 + 2e 1 p 1
= 0.76 0.24 + 2 0.04 2.04 0.5968 0.5664

1 2 =
0.24 0.76 3.04 0.4832 0.2736
10-36
Solved Problems
b 2 = b 1 + 2e 1 = 0.76 + 2 0.04 2.04 = 0.5968 .

0.76 3.04 0.5168
If we continue until the weights converge we find
W = 0.5948 0.0523 , b = 0.0131 .

0.1667 0.6667 0.1667
The resulting decision boundaries are shown in Figure P10.10. Compare

this result with the final decision boundaries created by the perceptron
learning rule in Problem P4.5 (Figure P4.7). The perceptron rule stops
training when all the patterns are classified correctly. The LMS algorithm
moves the boundaries as far from the patterns as possible.
1
3
Figure P10.10 Final Decision Boundaries for Problem P10.8
P10.9 Repeat the work of Widrow and Hoff on a pattern recognition

problem from their classic 1960 paper [WiHo60]. They wanted to
design a recognition system that would classify the six patterns
Patterns
T G F
Targets 60 0 -60
Figure P10.11 Patterns and Their Classification Targets
10-37
These patterns represent the letters T, G and F, in an original form

on the top and in a shifted form on the bottom. The targets for
these letters (in their original and shifted forms) are +60, 0 and -60,
respectively. (The values of 60, 0 and -60 were nice for use on the
face of a meter that Widrow and Hoff used to display their network
output.) The objective is to train a network so that it will classify
the six patterns into the appropriate T, G or F groups.
The blue squares in the letters will be assigned the value +1, and the white
squares will be assigned the value -1. First we convert each of the letters
into a single 16-element vector. We choose to do this by starting at the up-
per left corner, going down the left column, then going down the second col-
umn, etc. For example, the vector corresponding to the unshifted letter T is
T
p1 = 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
We have such an input vector for each of the six letters.

The ADALINE network that we will use is shown in Figure P10.12.
Input Linear Neuron
p a
16 x 1
W 1x1
1 x 16
n
1x1
1 b
1x1
16 1
Figure P10.12 Adaptive Pattern Classifier
(Widrow and Hoff built their own machine to realize this ADALINE. Ac-
cording to them, it was about the size of a lunch pail.)
Now we will present the six vectors to the network in a random sequence
and adjust the weights of the network after each presentation using the
LMS algorithm with a learning rate of = 0.03 . After each adjustment of
weights, all six vectors will be presented to the network to generate their
outputs and corresponding errors. The sum of the squares of the errors will
be examined as a measure of the quality of the network.
Figure P10.13 illustrates the convergence of the network. The network is
trained to recognize these six characters in about 60 presentations, or
roughly 10 for each of the possible input vectors.
10-38
Solved Problems
The results shown in Figure P10.13 are quite like those obtained and pub-
lished by Widrow and Hoff some 35 years ago. Widrow and Hoff did good
science. One can indeed duplicate their work, even decades later (without
a lunch pail).
4
x 10 Convergence of Error
2.5
Summ Sq. Error

1.5
0.5
0
0 10 20 30 40 50 60 70 80 90 100
Time Steps
Figure P10.13 Error Convergence with Learning Rate of 0.03

To experiment with this character recognition problem, use the MATLAB
Neural Network Design Demonstration Linear Pattern Classification
(nnd10lc). Notice the sensitivity of the network to noise in the input pattern.
10-39
Epilogue
In this chapter we have presented the ADALINE neural network and the
LMS learning rule. The ADALINE network is very similar to the percep-
tron network of Chapter 4, and it has the same fundamental limitation: it
can only classify linearly separable patterns. In spite of this limitation on
the network, the LMS algorithm is in fact more powerful than the percep-
tron learning rule. Because it minimizes mean square error, the algorithm
is able to create decision boundaries that are more robust to noise than
those of the perceptron learning rule.
The ADALINE network and the LMS algorithm have found many practical
applications. Even though they were first presented in the late 1950s, they
are still very much in use in adaptive filtering applications. For example,
echo cancellers using the LMS algorithm are currently employed on many
long distance telephone lines. (Chapter 14 provides more extensive cover-
age of dynamic networks, which are widely used for filtering, prediction
and control.)
In addition to its importance as a practical solution to many adaptive fil-
tering problems, the LMS algorithm is also important because it is the fore-
runner of the backpropagation algorithm, which we will discuss in
Chapters 11 through 14. Like the LMS algorithm, backpropagation is an
approximate steepest descent algorithm that minimizes mean square er-
ror. The only difference between the two algorithms is in the manner in
which the derivatives are calculated. Backpropagation is a generalization
of the LMS algorithm that can be used for multilayer networks. These more
complex networks are not limited to linearly separable problems. They can
solve arbitrary classification problems.
10-40
Further Reading
Further Reading
[AnRo89] J. A. Anderson, E. Rosenfeld, Neurocomputing: Founda-
tions of Research, Cambridge, MA: MIT Press, 1989.
Neurocomputing is a fundamental reference book. It con-
tains over forty of the most important neurocomputing
writings. Each paper is accompanied by an introduction
that summarizes its results and gives a perspective on the
position of the paper in the history of the field.
[StDo84] W. D. Stanley, G. R. Dougherty, R. Dougherty, Digital Sig-
nal Processing, Reston VA: Reston, 1984
[WiHo60] B. Widrow, M. E. Hoff, Adaptive switching circuits, 1960
IRE WESCON Convention Record, New York: IRE Part 4,
pp. 96104.
This seminal paper describes an adaptive perceptron-like
network that can learn quickly and accurately. The authors
assumed that the system had inputs, a desired output clas-
sification for each input, and that the system could calcu-
late the error between the actual and desired output. The
weights are adjusted, using a gradient descent method, so
as to minimize the mean square error. (Least mean square
error or LMS algorithm.)
This paper is reprinted in [AnRo88].
[WiSt 85] B. Widrow and S. D. Stearns, Adaptive Signal Processing,
Englewood Cliffs, NJ: Prentice-Hall, 1985.
This informative book describes the theory and application
of adaptive signal processing. The authors include a review
of the mathematical background that is needed, give de-
tails on their adaptive algorithms, and then discuss practi-
cal information about many applications.
[WiWi 88] B. Widrow and R. Winter, Neural nets for adaptive filter-
ing and adaptive pattern recognition, IEEE Computer
Magazine, March 1988, pp. 2539.
This is a particularly readable paper that summarizes ap-
plications of adaptive multilayer neural networks. The net-
works are applied to system modeling, statistical
prediction, echo cancellation, inverse modeling and pattern 10
recognition.
10-41
Exercises
E10.1 An adaptive filter ADALINE is shown in Figure E10.1. Suppose that the
weights of the network are given by
w 1 1 = 1 , w 1 2 = 4 , w 1 3 = 2 ,
and the input to the filter is
y k = 0 0 0 1 1 2 0 0 .
Find the response a k of the filter.
Inputs ADALINE
y(k)
w1,1
D w1,2 n(k) a(k)

SxR
D w1,3
Figure E10.1 Adaptive Filter ADALINE for Exercise E10.1
E10.2 In Figure E10.2 two classes of patterns are given.

i. Use the LMS algorithm to train an ADALINE network to distin-
guish between class I and class II patterns (we want the network to
identify horizontal and vertical lines).
ii. Can you explain why the ADALINE network might have difficulty
with this problem?
Class I Class II
Figure E10.2 Pattern Classification Problem for Exercise E10.2
10-42
Exercises
E10.3 Suppose that we have the following two reference patterns and their tar-
gets:
1 t = 1 , p = 1 t = 1 .
p1 = 1 2 2
1 1
In Problem P10.3 these input vectors to an ADALINE were assumed to oc-

cur with equal probability. Now suppose that the probability of vector p 1
is 0.75 and that the probability of vector p 2 is 0.25. Does this change of
probabilities change the mean square error surface? If yes, what does the
surface look like now? What is the maximum stable learning rate?
E10.4 In this exercise we will modify the reference pattern p 2 from Problem
P10.3:
1 t = 1 , p = 1 t = 1 .
p1 = 1 2 2
1 1
i. Assume that the patterns occur with equal probability. Find the
mean square error and sketch the contour plot.
ii. Find the maximum stable learning rate.
2+2
ans = iii. Write a MATLAB M-file to implement the LMS algorithm for this
4 problem. Take 40 steps of the algorithm for a stable learning rate.
Use the zero vector as the initial guess. Sketch the trajectory on the
contour plot.
iv. Take 40 steps of the algorithm after setting the initial values of both
parameters to 1. Sketch the final decision boundary.
v. Compare the final parameters from parts (iii) and (iv). Explain your
results.
E10.5 We again use the reference patterns and targets from Problem P10.3, and
assume that they occur with equal probability. This time we want to train
an ADALINE network with a bias. We now have three parameters to find:
w 1 1 , w 1 2 and b .
i. Find the mean square error and the maximum stable learning rate.
2+2
ans = ii. Write a MATLAB M-file to implement the LMS algorithm for this
4 problem. Take 40 steps of the algorithm for a stable learning rate.
Use the zero vector as the initial guess. Sketch the final decision
boundary.
iii. Take 40 steps of the algorithm after setting the initial values of all
parameters to 1. Sketch the final decision boundary.
iv. Compare the final parameters and the decision boundaries from
parts (iii) and (iv). Explain your results.
10-43
E10.6 We have two categories of vectors. Category I consists of
1 1
.
1 2
Category II consists of
0
4 .
1 1
We want to train a single-neuron ADALINE network without a bias to rec-

ognize these categories (t = 1 for Category I and t = -1 for Category II). As-
sume that each pattern occurs with equal probability.
i. Draw the network diagram.
ii. Take four steps of the LMS algorithm, using the zero vector as the
initial guess. (one pass through the four vectors above - present each
vector once). Use a learning rate of 0.1.
iii. What are the optimal weights?
iv. Sketch the optimal decision boundary.
v. How do you think the boundary would change if the network were
allowed to have a bias? If the boundary would change, indicate the
approximate new position on your sketch of part iv. You do not need
to perform any calculations here - just explain your reasoning.
gets:
3 t = 6 t = 6 t =
p1 = 1 75 , p2 = 2 75 , p3 = 3 75 .
6 3 3
Each pattern is equally likely.

i. Draw the network diagram for an ADALINE network with no bias
that could be trained on these patterns.
ii. We want to train the ADALINE network with no bias using these
patterns. Sketch the contour plot of the mean square error perfor-
mance index.
iii. Find the maximum stable learning rate for the LMS algorithm.
10-44
Exercises
iv. Sketch the trajectory of the LMS algorithm on your contour plot. As-
sume a very small learning rate, and start with all weights equal to
zero. This does not require any calculations.
E10.8 Suppose that we have the following two reference patterns and their tar-
gets:
1 t = 2 t =
p1 = 1 1 , p2 = 2 1 .
2 1
The probability of vector p 1 is 0.5 and the probability of vector p 2 is 0.5.We

want to train an ADALINE network without a bias on this data set.
i. Sketch the contour plot of the mean square error performance in-
dex.
ii. Sketch the optimal decision boundary.
iii. Find the maximum stable learning rate.
iv. Sketch the trajectory of the LMS algorithm on your contour plot. As-
sume a very small learning rate, and start with initial weights
W0 = 0 1 .
E10.9 We have the following input/target pairs:
4 t = 5 , p = 2 t = 2 , p = 4 t = 9 .
p1 = 1 2 2 3 3
2 4 4
The first two pair each occurs with probability of 0.25, and the third pair
occurs with probability 0.5. We want to train a single-neuron ADALINE
network without a bias to perform the desired mapping.
ii. What is the maximum stable learning rate?
iii. Perform one iteration of the LMS algorithm. Apply the input p 1 and
use a learning
T
rate of = 0.1 . Start from the initial weights
x0 = 0 0 .
10
10-45
E10.10 Repeat E10.9 for the following input/target pairs:
2 t = 1 , p = 4 t = 1 , p = 4 t = 1 .
p1 = 1 2 2 3 3
4 4 2
The first two pair each occurs with probability of 0.25, and the third pair
occurs with probability 0.5. We want to train a single-neuron ADALINE
network without a bias to perform the desired mapping.
E10.11 We want to train a single-neuron ADALINE network without a bias, using

the following training set, which categorizes vectors into two classes. Each
pattern occurs with equal probability.
1 t = 1 p = 2 t = 1 p = 0 t = 1 p = 1 t = 1
p1 = 1 2 2 3 3 4 4
2 1 1 0

ii. Take one step of the LMS algorithm (present p 1 only) starting from
the initial weight W 0 = 0 0 . Use a learning rate of 0.1.
iii. What are the optimal weights? Show all calculations.

iv. Sketch the optimal decision boundary.
v. How do you think the boundary would change if the network were
allowed to have a bias? Indicate the approximate new position on
your sketch of part iv.
vi. What is the maximum stable learning rate for the LMS algorithm?
vii. Sketch the contour plot of the mean square error performance sur-
face.
viii. On your contour plot of part vii, sketch the path of the LMS algo-
rithm for a very small learning rate (e.g., 0.001) starting from the
initial condition W 0 = 2 0 . This does not require any calcula-
tions, but explain how you obtained your answer.
gets:
2 t = 4 t = 2 t =
p1 = 1 26 , p2 = 2 26 , p3 = 3 26 .
4 2 2
10-46
Exercises
The probability of vector p 1 is 0.25, the probability of vector p 2 is 0.25 and

the probability of vector p 3 is 0.5.
i. Draw the network diagram for an ADALINE network with no bias
that could be trained on these patterns.
ii. Sketch the contour plot of the mean square error performance in-
dex.
iii. Show the optimal decision boundary (for the weights that minimize
mean square error) and verify that it separates the patterns into the
appropriate categories.
iv. Find the maximum stable learning rate for the LMS algorithm. If
the target values are changed from 26 and -26 to 2 and -2, how
would this change the maximum stable learning rate?
v. Perform one iteration of the LMS algorithm, starting with all
weights equal to zero, and presenting input vector p 1 . Use a learn-
ing rate of = 0.5 .
vi. Sketch the trajectory of the LMS algorithm on your contour plot. As-
sume a very small learning rate, and start with all weights equal to
zero.
E10.13 Consider the adaptive predictor in Figure E10.3.
Inputs ADALINE
t(k) = y(k)
y(k) D w1,1
n(k) a(k)
+
D w1,2
SxR
- e(k)
a(k) = w1,1 y(k - 1) + w1,2 y(k - 2)
Figure E10.3 Adaptive Predictor for Exercise E10.13

Assume that y k is a stationary process with autocorrelation function
Cy n = E y k y k + n .
i. Write an expression for the mean square error in terms of C y n .

ii. Give a specific expression for the mean square error when
10-47
y k = sin ------ .
k
5
iii. Find the eigenvalues and eigenvectors of the Hessian matrix for the
mean square error. Locate the minimum point and sketch a rough
contour plot.
iv. Find the maximum stable learning rate for the LMS algorithm.
v. Take three steps of the LMS algorithm by hand, using a stable
learning rate. Use the zero vector as the initial guess.
2+2
ans = vi. Write a MATLAB M-file to implement the LMS algorithm for this
4 problem. Take 40 steps of the algorithm for a stable learning rate
and sketch the trajectory on the contour plot. Use the zero vector as
the initial guess. Verify that the algorithm is converging to the op-
timal point.
vii. Verify experimentally that the algorithm is unstable for learning
rates greater than that found in part (iv).
E10.14 Repeat Problem P10.9, but use the numerals 1, 2 and 4, instead of the
letters T, G and F. Test the trained network on each reference pattern
and on noisy patterns. Discuss the sensitivity of the network. (Use the Neu-
ral Network Design Demonstration Linear Pattern Classification (nnd10lc).)
10-48
Objectives
11 Backpropagation
Objectives 11-1
Multilayer Perceptrons 11-2
The Backpropagation Algorithm 11-7
Performance Index 11-8
Chain Rule 11-9
Backpropagating the Sensitivities 11-11
Summary 11-13
Example 11-14
Batch vs. Incremental Training 11-17
Using Backpropagation 11-18
Convergence 11-20
Generalization 11-22
Epilogue 11-41
Exercises 11-44
Objectives
In this chapter we continue our discussion of performance learning, which
we began in Chapter 8, by presenting a generalization of the LMS algo-
rithm of Chapter 10. This generalization, called backpropagation, can be
used to train multilayer networks. As with the LMS learning law, back-
propagation is an approximate steepest descent algorithm, in which the
performance index is mean square error. The difference between the LMS
algorithm and backpropagation is only in the way in which the derivatives
are calculated. For a single-layer linear network the error is an explicit lin-
ear function of the network weights, and its derivatives with respect to the
weights can be easily computed. In multilayer networks with nonlinear
transfer functions, the relationship between the network weights and the
error is more complex. In order to calculate the derivatives, we need to use
the chain rule of calculus. In fact, this chapter is in large part a demonstra-
tion of how to use the chain rule.
11-1
11 Backpropagation
Theory and Examples

The perceptron learning rule of Frank Rosenblatt and the LMS algorithm
of Bernard Widrow and Marcian Hoff were designed to train single-layer
perceptron-like networks. As we have discussed in previous chapters, these
single-layer networks suffer from the disadvantage that they are only able
to solve linearly separable classification problems. Both Rosenblatt and
Widrow were aware of these limitations and proposed multilayer networks
that could overcome them, but they were not able to generalize their algo-
rithms to train these more powerful networks.
Apparently the first description of an algorithm to train multilayer net-
works was contained in the thesis of Paul Werbos in 1974 [Werbo74]. This
thesis presented the algorithm in the context of general networks, with
neural networks as a special case, and was not disseminated in the neural
network community. It was not until the mid 1980s that the backpropaga-
tion algorithm was rediscovered and widely publicized. It was rediscovered
independently by David Rumelhart, Geoffrey Hinton and Ronald Williams
[RuHi86], David Parker [Park85], and Yann Le Cun [LeCu85]. The algo-
rithm was popularized by its inclusion in the book Parallel Distributed Pro-
cessing [RuMc86], which described the work of the Parallel Distributed
Processing Group led by psychologists David Rumelhart and James Mc-
Clelland. The publication of this book spurred a torrent of research in neu-
ral networks. The multilayer perceptron, trained by the backpropagation
algorithm, is currently the most widely used neural network.
In this chapter we will first investigate the capabilities of multilayer net-
works and then present the backpropagation algorithm.
Multilayer Perceptrons
We first introduced the notation for multilayer networks in Chapter 2. For
ease of reference we have reproduced the diagram of the three-layer per-
ceptron in Figure 11.1. Note that we have simply cascaded three percep-
tron networks. The output of the first network is the input to the second
network, and the output of the second network is the input to the third net-
work. Each layer may have a different number of neurons, and even a dif-
ferent transfer function. Recall from Chapter 2 that we are using
superscripts to identify the layer number. Thus, the weight matrix for the
1
first layer is written as W and the weight matrix for the second layer is
2
written W .
To identify the structure of a multilayer network, we will sometimes use
the following shorthand notation, where the number of inputs is followed
by the number of neurons in each layer:
1 2 3
RS S S . (11.1)
11-2
Inputs First Layer Second Layer Third Layer
w 11,1 n11 a11 w 21,1 n21 a21 w 31,1 n31 a31

p1
f1 f2 f3
b1
1 b1
2 b31
p2 1 1 1
n12 a12 n22 a22 n32 a32
f 1
f 2
f 3
p3 b2
1 b2
2 b2
3
1 1 1
pR n1S 1 a1S 1 n2S 2 a2S 2 n3S 3 a3S 3
w 1 1
S ,R f1
w 2S 2 , S 1
f2
w 3S 3, S 2
f3
b1S 1 b2S 2 b3S 3
1 1 1

a3 = f 3 (W3f 2 (W2f 1 (W1p + b1) + b2) + b3)
Figure 11.1 Three-Layer Network

Lets now investigate the capabilities of these multilayer perceptron net-
works. First we will look at the use of multilayer networks for pattern clas-
sification, and then we will discuss their application to function
approximation.
Pattern Classification
2 To illustrate the capabilities of the multilayer perceptron for pattern clas-
+2
sification, consider the classic exclusive-or (XOR) problem. The input/tar-
get pairs for the XOR gate are
0 t = 0 p = 0 t = 1 p = 1 t = 1 p = 1 t = 0 .
p1 = 1 2 2 3 3 4 4
p2 p4 0 1 0 1
This problem, which is illustrated graphically in the figure to the left, was
p p used by Minsky and Papert in 1969 to demonstrate the limitations of the
1 3
single-layer perceptron. Because the two categories are not linearly sepa-
rable, a single-layer perceptron cannot perform the classification.
A two-layer network can solve the XOR problem. In fact, there are many
different multilayer solutions. One solution is to use two neurons in the
first layer to create two decision boundaries. The first boundary separates
p 1 from the other patterns, and the second boundary separates p 4 . Then
the second layer is used to combine the two boundaries together using an
11
11-3
11 Backpropagation
AND operation. The decision boundaries for each first-layer neuron are
shown in Figure 11.2.
1 w1
2w1
Layer 1/Neuron 1 Layer 1/Neuron 2
Figure 11.2 Decision Boundaries for XOR Network

The resulting two-layer, 2-2-1 network is shown in Figure 11.3. The overall
decision regions for this network are shown in the figure in the left margin.
The shaded region indicates those inputs that will produce a network out-
put of 1.
Inputs Individual Decisions AND Operation
2 n11 a11
p1
2
1 n21 a21
1
-1

-1 n2
1 a2
1
1 -1.5
p2
-1 1
1.5
1
Figure 11.3 Two-Layer XOR Network

See Problems P11.1 and P11.2 for more on the use of multilayer networks
for pattern classification.
Function Approximation
Up to this point in the text we have viewed neural networks mainly in the
context of pattern classification. It is also instructive to view networks as
function approximators. In control systems, for example, the objective is to
find an appropriate feedback function that maps from measured outputs to
control inputs. In adaptive filtering (Chapter 10) the objective is to find a
function that maps from delayed values of an input signal to an appropri-
ate output signal. The following example will illustrate the flexibility of the
multilayer perceptron for implementing functions.
11-4
Consider the two-layer, 1-2-1 network shown in Figure 11.4. For this exam-
2
+2 ple the transfer function for the first layer is log-sigmoid and the transfer
function for the second layer is linear. In other words,
1 -
f n = ---------------
1 2
n
and f n = n . (11.2)
1+e
Input Log-Sigmoid Layer Linear Layer
w11,1 n11 a11

w21,1
b11 n2 a2
p
1
n12 a12 b2
w12,1 w21,2
1
b12
1
a1 = logsig (W1p + b1) a2 = purelin (W2a1 + b2)

Figure 11.4 Example Function Approximation Network
Suppose that the nominal values of the weights and biases for this network
are
1 1 1 1
w 1 1 = 10 , w 2 1 = 10 , b 1 = 10 , b 2 = 10 ,
2 2 2
w 1 1 = 1 , w 1 2 = 1 , b = 0 .
The network response for these parameters is shown in Figure 11.5, which
2
plots the network output a as the input p is varied over the range 2 2 .
Notice that the response consists of two steps, one for each of the log-sig-
moid neurons in the first layer. By adjusting the network parameters we
can change the shape and location of each step, as we will see in the follow-
ing discussion.
The centers of the steps occur where the net input to a neuron in the first
layer is zero:
1
b1 10
n 1 = w 1 1 p + b 1 = 0 p = ---------
- = --------- = 1 ,
1 1 1
1
(11.3)
w 1 1 10
11-5
11 Backpropagation
1
b2 10
n 2 = w 2 1 p + b 2 = 0 p = ---------
- = ------ = 1 .
1 1 1
1
(11.4)
w 2 1 10
The steepness of each step can be adjusted by changing the network

weights.
3
2
a 1
-1
-2 -1 0 1 2
Figure 11.5 Nominal Response of Network of Figure 11.4

Figure 11.6 illustrates the effects of parameter changes on the network re-
sponse. The blue curve is the nominal response. The other curves corre-
spond to the network response when one parameter at a time is varied over
the following ranges:
2 2 1 2
1 w 1 1 1 , 1 w 1 2 1 , 0 b 2 20 , 1 b 1 . (11.5)
Figure 11.6 (a) shows how the network biases in the first (hidden) layer can
be used to locate the position of the steps. Figure 11.6 (b) illustrates how
the weights determine the slope of the steps. The bias in the second (out-
put) layer shifts the entire network response up or down, as can be seen in
Figure 11.6 (d).
From this example we can see how flexible the multilayer network is. It
would appear that we could use such networks to approximate almost any
function, if we had a sufficient number of neurons in the hidden layer. In
fact, it has been shown that two-layer networks, with sigmoid transfer
functions in the hidden layer and linear transfer functions in the output
layer, can approximate virtually any function of interest to any degree of
accuracy, provided sufficiently many hidden units are available (see
[HoSt89]).
To experiment with the response of this two-layer network, use the MAT-
LAB Neural Network Design Demonstration Network Function (nnd11nf).
11-6
The Backpropagation Algorithm
3 3
1 2
b2 w 1 1
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
(a) (b)
3 3
2 2
w 1 2 b
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
(c) (d)
Figure 11.6 Effect of Parameter Changes on Network Response

Now that we have some idea of the power of multilayer perceptron net-
works for pattern recognition and function approximation, the next step is
to develop an algorithm to train such networks.

It will simplify our development of the backpropagation algorithm if we use
the abbreviated notation for the multilayer network, which we introduced
in Chapter 2. The three-layer network in abbreviated notation is shown in
Figure 11.7.
As we discussed earlier, for multilayer networks the output of one layer be-
comes the input to the following layer. The equations that describe this op-
eration are
m+1 m+1 m+1 m m+1
a = f W a +b for m = 0 1 M 1 , (11.6)
where M is the number of layers in the network. The neurons in the first
layer receive external inputs:
0
a = p, (11.7)
which provides the starting point for Eq. (11.6). The outputs of the neurons
in the last layer are considered the network outputs:
M
a = a . (11.8)
11-7
11 Backpropagation
p a1 a2 a3
Rx1
W1 S1 x 1
W2 S2 x 1
W3 S3 x 1
n1 n2 n3
S1 x R
S1 x 1
f1 S2 x S1
S2 x 1
f2 S3 x S2
S3 x 1
f3
1 b1 1 b2 1 b3
S1 x 1 S2 x 1 S3 x 1
R S1 S2 S3

a3 = f 3 (W3 f 2 (W2f 1 (W1p + b1) + b2) + b3)
Figure 11.7 Three-Layer Network, Abbreviated Notation
Performance Index
The backpropagation algorithm for multilayer networks is a generalization
of the LMS algorithm of Chapter 10, and both algorithms use the same per-
formance index: mean square error. The algorithm is provided with a set of
examples of proper network behavior:
{p 1, t 1} { p 2, t 2} {p Q, tQ} , (11.9)
where p q is an input to the network, and t q is the corresponding target out-

put. As each input is applied to the network, the network output is com-
pared to the target. The algorithm should adjust the network parameters
in order to minimize the mean square error:
2 2
Fx= Ee = E t a . (11.10)
where x is the vector of network weights and biases (as in Chapter 10). If
the network has multiple outputs this generalizes to
T T
Fx= Ee e = Et a t a . (11.11)
As with the LMS algorithm, we will approximate the mean square error by
T T
F x = t k a k t k a k = e k e k , (11.12)
where the expectation of the squared error has been replaced by the
squared error at iteration k .
The steepest descent algorithm for the approximate mean square error
(stochastic gradient descent) is
11-8

F -
w i j k + 1 = w i j k ----------
m m
m
, (11.13)
w i j

F
b i k + 1 = b i k --------m- ,
m m
(11.14)
b i
where is the learning rate.

So far, this development is identical to that for the LMS algorithm. Now we
come to the difficult part the computation of the partial derivatives.
Chain Rule
For a single-layer linear network (the ADALINE) these partial derivatives
are conveniently computed using Eq. (10.33) and Eq. (10.34). For the mul-
tilayer network the error is not an explicit function of the weights in the
hidden layers, therefore these derivatives are not computed so easily.
Because the error is an indirect function of the weights in the hidden lay-
ers, we will use the chain rule of calculus to calculate the derivatives. To
review the chain rule, suppose that we have a function f that is an explicit
function only of the variable n . We want to take the derivative of f with
respect to a third variable w . The chain rule is then:
df n w df n dn w
---------------------- = ------------ --------------- . (11.15)
dw dn dw
2 For example, if
+2
n 2w
f n = e and n = 2w , so that f n w = e , (11.16)
then
f n w -
d--------------------- df n dn w
= ------------ --------------- = e 2 .
n
(11.17)
dw dn dw
We will use this concept to find the derivatives in Eq. (11.13) and Eq.
(11.14):
m
n i
F F- ----------
----------
- = -------- m- , (11.18)
m m
w i j n i w i j
m
F F n i
--------- = --------m- --------m- . (11.19)
m
b i n i b i
11-9
11 Backpropagation
The second term in each of these equations can be easily computed, since
the net input to layer m is an explicit function of the weights and bias in
that layer:
m1
S
w i j a j
m m m1 m
ni = + bi . (11.20)
j=1
Therefore
m m
n i 1 n i
----------
m
- = am
j , --------m- = 1 . (11.21)
w i j b i
If we now define

F
s i --------m- ,
m
(11.22)
n i

Sensitivity (the sensitivity of F to changes in the ith element of the net input at layer
m ), then Eq. (11.18) and Eq. (11.19) can be simplified to

F
----------
m
- = sm m1
i aj , (11.23)
w i j

F
--------m- = s m
i . (11.24)
b i
We can now express the approximate steepest descent algorithm as

m m m m1
w i j k + 1 = w i j k s i a j , (11.25)
m m m
b i k + 1 = b i k s i . (11.26)
In matrix form this becomes:
m m m m1 T
W k + 1 = W k s a , (11.27)
m m m
b k + 1 = b k s , (11.28)
where
11-10

F
--------m-
n 1

F
F ---------
s --------m- =
m m
n 2 . (11.29)
n

F
---------
m
-
n m
S
(Note the close relationship between this algorithm and the LMS algorithm
of Eq. (10.33) and Eq. (10.34)).
Backpropagating the Sensitivities

m
It now remains for us to compute the sensitivities s , which requires an-
other application of the chain rule. It is this process that gives us the term
backpropagation, because it describes a recurrence relationship in which
the sensitivity at layer m is computed from the sensitivity at layer m + 1 .
To derive the recurrence relationship for the sensitivities, we will use the
following Jacobian matrix:
m+1 m+1 m+1

n 1 n 1 n 1
---------------
m
--------------
m
- --------------
m
-
n 1 n 2 n S m
m+1 m+1 m+1
n 2 n 2 n 2
n
m+1 --------------- --------------
- --------------
-
---------------
m
- n m
n 2
m
n S m
m . (11.30)
n 1

m+1 m+1 m+1

n S m + 1 n S m + 1 n S m + 1
---------------
m
--------------
m
- --------------
m
-
n 1 n 2 n S m
Next we want to find an expression for this matrix. Consider the i, j ele-
ment of the matrix:
11-11
11 Backpropagation
m
S
w i l a l + b i
m+1 m m+1
m+1 m
n i l = 1 m + 1 a j
--------------
- = --------------------------------------------------------
- = w i j --------
m
-
m m
n j n j n j
m m
m + 1 f n j + 1m m
= w i j ------------------
m
- = wm
i j f n j ,
(11.31)
n j
where
m m
m m f n j
f n j = ------------------
m
-. (11.32)
n j
Therefore the Jacobian matrix can be written

m+1
n
- = W m + 1 F n m ,
m
---------------
m
(11.33)
n
where
m m
f n1 0 0
m m
F n = 0 f n2 0
m m
. (11.34)

m m
0 0 f nSm
We can now write out the recurrence relation for the sensitivity by using
the chain rule in matrix form:
n m + 1 F
T
F F
m n m W m + 1 T ---------------
s
m
= --------
- = ---------------
- ---------------
- = F -
n n
m m m+1 m+1
n n
= F (n ) W
m m m+1 T m+1 (11.35)
s .
Now we can see where the backpropagation algorithm derives its name.
The sensitivities are propagated backward through the network from the
last layer to the first layer:
M M1 2 1
s s s s . (11.36)
11-12
At this point it is worth emphasizing that the backpropagation algorithm

uses the same approximate steepest descent technique that we used in the
LMS algorithm. The only complication is that in order to compute the gra-
dient we need to first backpropagate the sensitivities. The beauty of back-
propagation is that we have a very efficient implementation of the chain
rule.
We still have one more step to make in order to complete the backpropaga-
M
tion algorithm. We need the starting point, s , for the recurrence relation
of Eq. (11.35). This is obtained at the final layer:
M
S
tj aj
2
T
a
F t a t a
= --------M- = ------------------------------------ - = 2 t i a i --------Mi- .
- = -------------------------------
M j=1
si M M
(11.37)
n i n i n i n i
Now, since
M M M
a a f n i
- = f n M
M
--------Mi- = --------iM- = -------------------
M i , (11.38)
n i n i n i
we can write
M M M
si = 2 ti ai f ni . (11.39)
This can be expressed in matrix form as
= 2F (n ) t a .
M M M
s (11.40)
Summary
Lets summarize the backpropagation algorithm. The first step is to propa-
gate the input forward through the network:
0
a = p, (11.41)
m+1 m+1 m+1 m m+1
a = f W a +b for m = 0 1 M 1 , (11.42)
M
a = a . (11.43)
The next step is to propagate the sensitivities backward through the net-
work:
= 2F (n ) t a ,
M M M
s (11.44)
11-13
11 Backpropagation
= F (n ) W
m m m m+1 T m+1
s s , for m = M 1 2 1 . (11.45)
Finally, the weights and biases are updated using the approximate steep-
est descent rule:
m m m m1 T
W k + 1 = W k s a , (11.46)
m m m
b k + 1 = b k s . (11.47)
Example
2 To illustrate the backpropagation algorithm, lets choose a network and ap-
+2
ply it to a particular problem. To begin, we will use the 1-2-1 network that
we discussed earlier in this chapter. For convenience we have reproduced
the network in Figure 11.8.
Next we want to define a problem for the network to solve. Suppose that we
want to use the network to approximate the function

g p = 1 + sin --- p for 2 p 2 . (11.48)
4
To obtain our training set we will evaluate this function at several values
of p .
w11,1 n11 a11

w21,1
b11 n2 a2
p
1
n12 a12 b2
w12,1 w21,2
1
b12
1
Figure 11.8 Example Function Approximation Network

Before we begin the backpropagation algorithm we need to choose some ini-
tial values for the network weights and biases. Generally these are chosen
to be small random values. In the next chapter we will discuss some rea-
sons for this. For now lets choose the values
11-14
Example
W 0 = 0.27 , b 0 = 0.48 , W 0 = 0.09 0.17 , b 0 = 0.48 .

1 1 2 2
0.41 0.13
The response of the network for these initial values is illustrated in Figure
11.9, along with the sine function we wish to approximate.
2.5
1.5
2
a 1
0.5
0.5
1
2 1.5 1 0.5 0 0.5 1 1.5 2
Figure 11.9 Initial Network Response

Next, we need to select a training set {p 1, t 1} {p 2, t 2} {p Q, tQ} . In this
case, we will sample the function at 21 points in the range [-2,2] at equally
spaced intervals of 0.2. The training points are indicated by the circles in
Figure 11.9.
Now we are ready to start the algorithm. The training points can be pre-
sented in any order, but they are often chosen randomly. For our initial in-
put we will choose p = 1 , which is the 16th training point:
0
a = p = 1.
The output of the first layer is then

a = f W a + b = logsig 0.27 1 + 0.48 = logsig 0.75
1 1 1 0 1
0.41 0.13 0.54
1
------------------
0.75
-
= 1 + e = 0.321 .
1
-------------------
0.368
0.54
1+e
The second layer output is
11-15
11 Backpropagation
a = f W a + b = purelin ( 0.09 0.17 0.321 + 0.48 ) = 0.446 .

2 2 2 1 2
0.368
The error would then be
2
e = t a = 1 + sin --- p a = 1 + sin --- 1 0.446 = 1.261 .
4 4

The next stage of the algorithm is to backpropagate the sensitivities. Be-

fore we begin the backpropagation, recall that we will need the derivatives
1 2
of the transfer functions, f (n) and f (n) . For the first layer
n
1 d ---------------
1 -
= -----------------------2 = 1 ---------------
e 1 1
f n = - ---------------
- = 1 a1 a1 .
d n 1 + e n 1 + e
n 1+e
n
1+e
n
For the second layer we have
f n = d n = 1 .
2
dn
We can now perform the backpropagation. The starting point is found at

the second layer, using Eq. (11.44):
s = 2F (n ) t a = 2 f 2 n 2 1.261 = 2 1 1.261 = 2.522 .

2 2 2
The first layer sensitivity is then computed by backpropagating the sensi-

tivity from the second layer, using Eq. (11.45):
1 1
1 a1 a1
s = F (n ) W s =
1 1 1 2 T 2 0 0.09
1 1
2.522
0 1 a 2 a 2 0.17
= 1 0.321 0.321 0 0.09

2.522
0 1 0.368 0.368 0.17
= 0.218 0 0.227 = 0.0495 .

0 0.233 0.429 0.0997
The final stage of the algorithm is to update the weights. For simplicity, we
will use a learning rate = 0.1 . (In Chapter 12 the choice of learning rate
will be discussed in more detail.) From Eq. (11.46) and Eq. (11.47) we have
11-16
Batch vs. Incremental Training
2 2 2 1 T
W 1 = W 0 s a = 0.09 0.17 0.1 2.522 0.321 0.368
= 0.171 0.0772 ,
2 2 2
b 1 = b 0 s = 0.48 0.1 2.522 = 0.732 ,
1 0 T
W 1 = W 0 s a = 0.27 0.1 0.0495 1 = 0.265 ,
1 1
0.41 0.0997 0.420
b 1 = b 0 s = 0.48 0.1 0.0495 = 0.475 .

1 1 1
0.13 0.0997 0.140
This completes the first iteration of the backpropagation algorithm. We

next proceed to randomly choose another input from the training set and
perform another iteration of the algorithm. We continue to iterate until the
difference between the network response and the target function reaches
some acceptable level. (Note that this will generally take many passes
through the entire training set.) We will discuss convergence criteria in
more detail in Chapter 12.
To experiment with the backpropagation calculation for this two-layer net-
work, use the MATLAB Neural Network Design Demonstration Backprop-
agation Calculation (nnd11bc).
Batch vs. Incremental Training

The algorithm described above is the stochastic gradient descent algo-
Incremental Training rithm, which involves on-line or incremental training, in which the net-
work weights and biases are updated after each input is presented (as with
Batch Training the LMS algorithm of Chapter 10). It is also possible to perform batch train-
ing, in which the complete gradient is computed (after all inputs are ap-
plied to the network) before the weights and biases are updated. For
example, if each input occurs with equal probability, the mean square error
performance index can be written
Q
1
F x = E e e = E t a t a = ---- tq aq
T T T
tq aq . (11.49)
Q
q=1
The total gradient of this performance index is
11
11-17
11 Backpropagation
Q Q
1 1
F x = ---- t q a q t q a q = ---- t q a q t q a q . (11.50)
T T
Q
q=1 Q
q=1
Therefore, the total gradient of the mean square error is the mean of the
gradients of the individual squared errors. Therefore, to implement a batch
version of the backpropagation algorithm, we would step through Eq.
(11.41) through Eq. (11.45) for all of the inputs in the training set. Then,
the individual gradients would be averaged to get the total gradient. The
update equations for the batch steepest descent algorithm would then be
Q
m1 T
W k + 1 = W k ---- sq aq
m m m
, (11.51)
Q
q=1
Q

b k + 1 = b k ---- sq .
m m m
(11.52)
Q
q=1
Using Backpropagation
In this section we will present some issues relating to the practical imple-
mentation of backpropagation. We will discuss the choice of network archi-
tecture, and problems with network convergence and generalization. (We
will discuss implementation issues again in Chapter 12, which investigates
procedures for improving the algorithm.)
Choice of Network Architecture

As we discussed earlier in this chapter, multilayer networks can be used to
approximate almost any function, if we have enough neurons in the hidden
layers. However, we cannot say, in general, how many layers or how many
neurons are necessary for adequate performance. In this section we want
to use a few examples to provide some insight into this problem.
2 For our first example lets assume that we want to approximate the follow-
+2
ing functions:
g p = 1 + sin ----- p for 2 p 2 ,

i
(11.53)
4
where i takes on the values 1, 2, 4 and 8. As i is increased, the function

becomes more complex, because we will have more periods of the sine wave
over the interval 2 p 2 . It will be more difficult for a neural network
with a fixed number of neurons in the hidden layers to approximate g p
as i is increased.
11-18
For this first example we will use a 1-3-1 network, where the transfer func-
tion for the first layer is log-sigmoid and the transfer function for the sec-
ond layer is linear. Recall from our example on page 11-5 that this type of
two-layer network can produce a response that is a sum of three log-sig-
moid functions (or as many log-sigmoids as there are neurons in the hidden
layer). Clearly there is a limit to how complex a function this network can
implement. Figure 11.10 illustrates the response of the network after it has
been trained to approximate g p for i = 1 2 4 8 . The final network re-
sponses are shown by the blue lines.
3 3
i = 1 i = 2
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
3 3
i = 4 i = 8
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
Figure 11.10 Function Approximation Using a 1-3-1 Network

We can see that for i = 4 the 1-3-1 network reaches its maximum capabil-
ity. When i 4 the network is not capable of producing an accurate approx-
imation of g p . In the bottom right graph of Figure 11.10 we can see how
the 1-3-1 network attempts to approximate g p for i = 8 . The mean
square error between the network response and g p is minimized, but the
network response is only able to match a small part of the function.
2 In the next example we will approach the problem from a slightly different
+2
perspective. This time we will pick one function g p and then use larger
and larger networks until we are able to accurately represent the function.
For g p we will use
g p = 1 + sin ------ p for 2 p 2 .

6
(11.54)
4
To approximate this function we will use two-layer networks, where the

transfer function for the first layer is log-sigmoid and the transfer function
1
for the second layer is linear (1- S -1 networks). As we discussed earlier in
11-19
11 Backpropagation
1
this chapter, the response of this network is a superposition of S sigmoid
functions.
Figure 11.11 illustrates the network response as the number of neurons in
the first layer (hidden layer) is increased. Unless there are at least five neu-
rons in the hidden layer the network cannot accurately represent g p .
3 3
1-2-1 1-3-1
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
3 3
1-4-1 1-5-1
2 2
1 1
0 0
-1 -1
-2 -1 0 1 2 -2 -1 0 1 2
Figure 11.11 Effect of Increasing the Number of Hidden Neurons

1
To summarize these results, a 1- S -1 network, with sigmoid neurons in the
hidden layer and linear neurons in the output layer, can produce a re-
1
sponse that is a superposition of S sigmoid functions. If we want to ap-
proximate a function that has a large number of inflection points, we will
need to have a large number of neurons in the hidden layer.
Use the MATLAB Neural Network Design Demonstration Function Ap-
proximation (nnd11fa) to develop more insight into the capability of a two-
layer network.
Convergence
In the previous section we presented some examples in which the network
response did not give an accurate approximation to the desired function,
even though the backpropagation algorithm produced network parameters
that minimized mean square error. This occurred because the capabilities
of the network were inherently limited by the number of hidden neurons it
contained. In this section we will provide an example in which the network
is capable of approximating the function, but the learning algorithm does
not produce network parameters that produce an accurate approximation.
In the next chapter we will discuss this problem in more detail and explain
why it occurs. For now we simply want to illustrate the problem.
11-20
The function that we want the network to approximate is

2
+2
g p = 1 + sin p for 2 p 2 . (11.55)
To approximate this function we will use a 1-3-1 network, where the trans-
fer function for the first layer is log-sigmoid and the transfer function for
the second layer is linear.
Figure 11.12 illustrates a case where the learning algorithm converges to
a solution that minimizes mean square error. The thin blue lines represent
intermediate iterations, and the thick blue line represents the final solu-
tion, when the algorithm has converged. (The numbers next to each curve
indicate the sequence of iterations, where 0 represents the initial condition
and 5 represents the final solution. The numbers do not correspond to the
iteration number. There were many iterations for which no curve is repre-
sented. The numbers simply indicate an ordering.)
2 5
1 3
2
4
0
0
-1
-2 -1 0 1 2
Figure 11.12 Convergence to a Global Minimum

Figure 11.13 illustrates a case where the learning algorithm converges to
a solution that does not minimize mean square error. The thick blue line
(marked with a 5) represents the network response at the final iteration.
The gradient of the mean square error is zero at the final iteration, there-
fore we have a local minima, but we know that a better solution exists, as
evidenced by Figure 11.12. The only difference between this result and the
result shown in Figure 11.12 is the initial condition. From one initial con-
dition the algorithm converged to a global minimum point, while from an-
other initial condition the algorithm converged to a local minimum point.
11-21
11 Backpropagation
5
1 3
4
0
0
1
-1
-2 -1 0 1 2
Figure 11.13 Convergence to a Local Minimum

Note that this result could not have occurred with the LMS algorithm. The
mean square error performance index for the ADALINE network is a qua-
dratic function with a single minimum point (under most conditions).
Therefore the LMS algorithm is guaranteed to converge to the global min-
imum as long as the learning rate is small enough. The mean square error
for the multilayer network is generally much more complex and has many
local minima (as we will see in the next chapter). When the backpropaga-
tion algorithm converges we cannot be sure that we have an optimum so-
lution. It is best to try several different initial conditions in order to ensure
that an optimum solution has been obtained.
Generalization
In most cases the multilayer network is trained with a finite number of ex-
amples of proper network behavior:
{p 1, t 1} { p 2, t 2} {p Q, tQ} . (11.56)
This training set is normally representative of a much larger class of pos-

sible input/output pairs. It is important that the network successfully gen-
eralize what it has learned to the total population.
2 For example, suppose that the training set is obtained by sampling the fol-
+2
lowing function:

g p = 1 + sin --- p , (11.57)
4
at the points p = 2 1.6 1.2 1.6 2 . (There are a total of 11 input/tar-

get pairs.) In Figure 11.14 we see the response of a 1-2-1 network that has
11-22
been trained on this data. The black line represents g p , the blue line rep-
resents the network response, and the + symbols indicate the training set.
3
-1
-2 -1 0 1 2
Figure 11.14 1-2-1 Network Approximation of g p
We can see that the network response is an accurate representation of

g p . If we were to find the response of the network at a value of p that was
not contained in the training set (e.g., p = 0.2 ), the network would still
produce an output close to g p . This network generalizes well.
Now consider Figure 11.15, which shows the response of a 1-9-1 network
that has been trained on the same data set. Note that the network response
accurately models g p at all of the training points. However, if we com-
pute the network response at a value of p not contained in the training set
(e.g., p = 0.2 ) the network might produce an output far from the true re-
sponse g p . This network does not generalize well.
3
-1
-2 -1 0 1 2
Figure 11.15 1-9-1 Network Approximation of g p
The 1-9-1 network has too much flexibility for this problem; it has a total
of 28 adjustable parameters (18 weights and 10 biases), and yet there are
only 11 data points in the training set. The 1-2-1 network has only 7 param-
11-23
11 Backpropagation
eters and is therefore much more restricted in the types of functions that it
can implement.
For a network to be able to generalize, it should have fewer parameters than
there are data points in the training set. In neural networks, as in all mod-
eling problems, we want to use the simplest network that can adequately
represent the training set. Dont use a bigger network when a smaller net-
work will work (a concept often referred to as Ockhams Razor).
An alternative to using the simplest network is to stop the training before
the network overfits. A reference to this procedure and other techniques to
improve generalization are given in Chapter 13.
To experiment with generalization in neural networks, use the MATLAB
Neural Network Design Demonstration Generalization (nnd11gn).
11-24
Summary of Results
Summary of Results
Multilayer Network
p a1 a2 a3
Rx1
W1 S1 x 1
W2 S2 x 1
W3 S3 x 1
n1 n2 n3
S1 x R
S1 x 1
f 1 S2 x S1
S2 x 1
f 2 S3 x S2
S3 x 1
f 3
1 b1 1 b2 1 b3
S1 x 1 S2 x 1 S3 x 1
R S1 S2 S3

a3 = f 3 (W3 f 2 (W2f 1 (W1p + b1) + b2) + b3)
Backpropagation Algorithm
Performance Index
T T
Fx= Ee e = Et a t a
Approximate Performance Index

T T
F x = e k e k = t k a k t k a k
Sensitivity

F
--------m-
n 1

F
F --------m-
s --------m- =
m
n 2
n

F
---------
m
-
n S m
11
11-25
11 Backpropagation
Forward Propagation
0
a = p,
m+1 m+1 m+1 m m+1
a = f W a +b for m = 0 1 M 1 ,
M
a = a .
Backward Propagation
= 2F (n ) t a ,
M M M
s
= F (n ) W
m m m m+1 T m+1
s s , for m = M 1 2 1 ,
where
f n 1
m m
0 0
m
F n =
m
m m 0 f n2 0
,

f n S m
m m
0 0
m m
m m f n j
f n j = ------------------
m
-.
n j
Weight Update (Approximate Steepest Descent)

m m m m1 T
W k + 1 = W k s a ,
m m m
b k + 1 = b k s .
11-26
Solved Problems
Solved Problems
P11.1 Consider the two classes of patterns that are shown in Figure
P11.1. Class I represents vertical lines and Class II represents hor-
izontal lines.
Class I
Class II
Figure P11.1 Pattern Classes for Problem P11.1

i. Are these categories linearly separable?
ii. Design a multilayer network to distinguish these categories.
i. Lets begin by converting the patterns to vectors by scanning each 2X2
grid one column at a time. Each white square will be represented by a -1
and each blue square by a 1. The vertical lines (Class I patterns) then be-
come
1 1
p1 = 1 and p 2 = 1 ,
1 1
1 1
and the horizontal lines (Class II patterns) become
1 1
p 3 = 1 and p 4 = 1 .
1 1
1 1
In order for these categories to be linearly separable we must be able to

place a hyperplane between the two categories. This means there must be
a weight matrix W and a bias b such that
Wp 1 + b 0 , Wp 2 + b 0 , Wp 3 + b 0 , Wp 4 + b 0 .
These conditions can be converted to
11-27
11 Backpropagation
1
1 =
w 1 1 w 1 2 w 1 3 w 1 4 w 1 1 + w 1 2 w 1 3 w 1 4 0 ,
1
1
w 1 1 w 1 2 + w 1 3 + w 1 4 0 ,
w 1 1 w 1 2 + w 1 3 w 1 4 0 ,
w 1 1 + w 1 2 w 1 3 + w 1 4 0 .
The first two conditions reduce to
w 1 1 + w 1 2 w 1 3 + w 1 4 and w 1 3 + w 1 4 w 1 1 + w 1 2 ,
which are contradictory. The final two conditions reduce to
w 1 1 + w 1 3 w 1 2 + w 1 4 and w 1 2 + w 1 4 w 1 1 + w 1 3 ,
which are also contradictory. Therefore there is no hyperplane that can

separate these two categories.
ii. There are many different multilayer networks that could solve this prob-
lem. We will design a network by first noting that for the Class I vectors
either the first two elements or the last two elements will be 1. The Class
II vectors have alternating 1 and -1 patterns. This leads to the network
Inputs AND Operations OR Operation
p1 2
n11 a11
p2
2 n21 a21
2
1
-1

p3 2 n12 a12 2 1
1
p4 2 -1
1
Figure P11.2 Network to Categorize Horizontal and Vertical Lines
11-28
Solved Problems
The first neuron in the first layer tests the first two elements of the input
vector. If they are both 1 it outputs a 1, otherwise it outputs a -1. The
second neuron in the first layer tests the last two elements of the input vec-
tor in the same way. Both of the neurons in the first layer perform AND op-
erations. The second layer of the network tests whether either of the
outputs of the first layer are 1. It performs an OR operation. In this way,
the network will output a 1 if either the first two elements or the last two
elements of the input vector are both 1.
P11.2 Figure P11.3 illustrates a classification problem, where Class I vec-

tors are represented by light circles, and Class II vectors are rep-
resented by dark circles. These categories are not linearly
separable. Design a multilayer network to correctly classify these
categories.
Figure P11.3 Classification Problem

We will solve this problem with a procedure that can be used for arbitrary
classification problems. It requires a three-layer network, with hard-limit-
ing neurons in each layer. In the first layer we create a set of linear decision
boundaries that separate every Class I vector from every Class II vector.
For this problem we used 11 such boundaries. They are shown in Figure
P11.4.
5
2 3 6
1 4 7
8
9
10
11
Figure P11.4 First Layer Decision Boundaries

Each row of the weight matrix in the first layer corresponds to one decision
boundary. The weight matrix and bias vector for the first layer are
1 T
W = 1 1 1 1 1 1 1 1 1 1 1 ,
1 1 1 1 1 1 1 1 1 1 1
11-29
11 Backpropagation
1 T
b = 2 3 0.5 0.5 1.75 2.25 3.25 3.75 6.25 5.75 4.75 .
(Review Chapters 3, 4 and 10 for procedures for calculating the appropriate

weight matrix and bias for a given decision boundary.) Now we can com-
bine the outputs of the 11 first layer neurons into groups with a second lay-
er of AND neurons, such as those we used in the first layer of the network
in Problem P11.1. The second layer weight matrix and bias are
1 1 1 1 0 0 0 0 0 0 0 3
W = 0 0 0 0 1 1 0 0 1 0 1 , b = 3 .
2 T
0 0 0 0 1 0 0 1 1 1 0 3
0 0 0 0 0 0 1 1 1 0 1 3
The four decision boundaries for the second layer are shown in Figure
P11.5. For example, the neuron 2 decision boundary is obtained by combin-
ing the boundaries 5, 6, 9 and 11 from layer 1. This can be seen by looking
2
at row 2 of W .
2
3
1
Figure P11.5 Second Layer Decision Regions

In the third layer of the network we will combine together the four decision
regions of the second layer into one decision region using an OR operation,
just as in the last layer of the network in Problem P11.1. The weight matrix
and bias for the third layer are
3 3
W = 1 1 1 1 ,b = 3 .
The complete network is shown in Figure P11.6.

The procedure that we used to develop this network can be used to solve
classification problems with arbitrary decision boundaries as long as we
have enough neurons in the hidden layers. The idea is to use the first layer
to create a number of linear boundaries, which can be combined by using
AND neurons in the second layer and OR neurons in the third layer. The
decision regions of the second layer are convex, but the final decision
boundaries created by the third layer can have arbitrary shapes.
11-30
Solved Problems
Input Initial Decisions AND Operations OR Operation
p a1 a2 a3
2x1
W1 11 x 1
W2 4x1
W3 1x1
11 x 2
n1 4 x 11
n2 1x4
n3
11 x 1 4x1 1x1
1 b 1 1 b 2 1 b 3
11 x 1 4x1 1x1
2 11 4 1
a1 = hardlims (W1p + b1) a2 = hardlims (W2a1 + b2) a3 = hardlims (W3a2 + b3)
Figure P11.6 Network for Problem P11.2

The final network decision regions are given in Figure P11.7. Any vector in
the shaded areas will produce a network output of 1, which corresponds to
Class II. All other vectors will produce a network output of -1, which cor-
responds to Class I.
Figure P11.7 Final Decision Regions
P11.3 Show that a multilayer network with linear transfer functions is

equivalent to a single-layer linear network.
For a multilayer linear network the forward equations would be
1 1 1
a = W p+b
2 2 1 2 2 1 2 1 2
a = W a + b = W W p + W b + b ,
3 3 2 3 3 2 1 3 2 1 3 2 3
a = W a + b = W W W p + W W b + W b + b .
If we continue this process we can see that for an M-layer linear network,
the equivalent single-layer linear network would have the following weight
matrix and bias vector
M M1 2 1
W = W W W W ,
11-31
11 Backpropagation
W b + + b .
M M1 2 1 M M1 3 2 M
b = W W W b + W W
P11.4 The purpose of this problem is to illustrate the use of the chain
rule. Consider the following dynamic system:
y k + 1 = fyk .
We want to choose the initial condition y 0 so that at some final

time k = K the system output y K will be as close as possible to
some target output t . We will minimize the performance index
2
F y 0 = t yK
using steepest descent, so we need the gradient

Fy0 .
y0
Find a procedure for computing this using the chain rule.

The gradient is
2
t y K
F y0 = = 2t yK y K .
y0 y0 y 0
The key term is
yK
,
y0
which cannot be computed directly, since y K is not an explicit function of

y 0 . Lets define an intermediate term

rk y k .
y 0
Then we can use the chain rule:
y k + 1 y k y k + 1
r k + 1 = y k + 1 = ---------------------- ------------- = ---------------------- r k .
y 0 y k y 0 y k
From the system dynamics we know
----------------------
yk + 1 f y k
= -------------------- = f y k .
y k y k
Therefore the recursive equation for the computation of r k is
11-32
Solved Problems

r k + 1 = f y k r k .
This is initialized at k = 0 :
y 0
r 0 = ------------- = 1 .
y 0
The total procedure for computing the gradient is then
r0 = 1 ,

r k + 1 = f y k r k , for k = 0 1 K 1 ,

F y 0 = 2 t y K r K .
y0
P11.5 Consider the two-layer network shown in Figure P11.8. The initial
weights and biases are set to
1 1 2 2
w = 1 , b = 1 , w = 2 , b = 1 .
An input/target pair is given to be
( p = 1 , t = 1 ) .
2
i. Find the squared error e as an explicit function of all
weights and biases.
2 1
ii. Using part (i) find e w at the initial weights and biases.
iii. Repeat part (ii) using backpropagation and compare results.
Inputs Log-Sigmoid Layer Linear Layer
w1 n1 a1 w2 n2 a2
p
b1 b2
1 1
a 1 = logsig (w 1 p + b 1) a 2 = purelin (w 2 a 1 + b 2)
Figure P11.8 Two-Layer Network for Problem P11.5

i. The squared error is given by
11-33
11 Backpropagation
2 2 2 1
2
e = t a = t w ---------------------------------------------------------
- + b 2 .
2
1 + exp w p + b
1 1
ii. The derivative is
e
2
e 2 1
------------
1
= 2e --------1- = 2e w -----------------------------------------------------------2- exp w p + b p
1 1
w w 1 + exp w p + b 1 1

To evaluate this at the initial weights and biases we find
1 1
a = ---------------------------------------------------------
- = --------------------------------------------------------- = 0.8808
1
1 + exp w p + b
1 1 1 + exp 1 1 + 1
2 2 1 2
a = w a + b = 2 0.8808 + 1 = 0.7616
2
e = t a = 1 0.7616 = 1.7616
e
2
2 1
------------
1
= 2e w -----------------------------------------------------------2- exp w p + b p
1 1
w 1 + exp w p + b 1 1

1
= 2 1.7616 2 ----------------------------------------------------------2- exp 1 1 + 1 1
1 + exp 1 1 + 1
= 3.5232 0.2707 -------------------2- = 0.7398 .

1
1.289
iii. To backpropagate the sensitivities we use Eq. (11.44) and Eq. (11.45):
s = 2F (n ) t a = 2 1 1 0.7616 = 3.5232 ,
2 2 2
s = F (n ) W s = a 1 a 2 s
1 1 1 2 T 2 1 1 2
= 0.8808 1 0.8808 2 3.5232 = 0.7398 .

2 1
From Eq. (11.23) we can compute e w :
2
e
------------
1
1 0 1
= s a = s p = 0.7398 1 = 0.7398 .
w
This agrees with our result from part (ii).
11-34
Solved Problems
P11.6 Earlier in this chapter we showed that if the neuron transfer func-
tion is log-sigmoid,
1
a = f n = ---------------
n
-,
1+e
then the derivative can be conveniently computed by

fn = a1 a .
Find a convenient way to compute the derivative for the hyperbol-

ic tangent sigmoid:
n n
e e -
a = f n = tansig n = -----------------
n n
.
e +e
Computing the derivative directly we find
d -----------------
e e -
n n n n n n
df n e e - n n e + e
------------------------
f n = ------------ = n = e e + -----------------
-
dn d n e + e n n
e + e
n 2 n
e +e
n
n n 2
e e
= 1 -------------------------2 = 1 a .
2
n n
e + e
P11.7 For the network shown in Figure P11.9 the initial weights and bi-
ases are chosen to be
1 1 2 2
w 0 = 1 , b 0 = 1 , w 0 = 2 , b 0 = 1 .
( p = 1 , t = 1 ) .
Perform one iteration of backpropagation with = 1 .
11-35
11 Backpropagation
Inputs Tan-Sigmoid Layer Tan-Sigmoid Layer
w1 n1 a1 w2 n2 a2
p
b1 b2
1 1
a 1 = tansig (w 1 p + b 1) a 2 = tansig (w 2 a 1 + b 2)
Figure P11.9 Two-Layer Tan-Sigmoid Network

The first step is to propagate the input through the network.
1 1 1
n = w p + b = 1 1 + 1 = 2
1 1
exp n exp n exp 2 exp 2
a = tansig n = ------------------------------------------------
- = -------------------------------------------- = 0.964
1 1
1
exp n + exp n
1 exp 2 + exp 2
2 2 1 2
n = w a + b = 2 0.964 + 1 = 0.928
2 2
exp n exp n exp 0.928 exp 0.928
a = tansig n = ------------------------------------------------
- = -----------------------------------------------------------------
2 2
2
exp n + exp n
2 exp 0.928 + exp 0.928
= 0.7297
2
e = t a = 1 0.7297 = 1.7297
Now we backpropagate the sensitivities using Eq. (11.44) and Eq. (11.45).
s = 2F (n ) t a = 2 1 a e = 2 1 0.7297 1.7297
2 2 2 2 2 2
= 1.6175
s = F (n ) W s = 1 a w s = 1 0.964 2 1.6175
1 1 1 2 T 2 1 2 2 2 2
= 0.2285
Finally, the weights and biases are updated using Eq. (11.46) and Eq.
(11.47):
2 2 2 1 T
w 1 = w 0 s a = 2 1 1.6175 0.964 = 0.4407 ,
11-36
Solved Problems
1 1 1 0 T
w 1 = w 0 s a = 1 1 0.2285 1 = 0.7715 ,
2 2 2
b 1 = b 0 s = 1 1 1.6175 = 2.6175 ,
1 1 1
b 1 = b 0 s = 1 1 0.2285 = 0.7715 .
P11.8 In Figure P11.10 we have a network that is a slight modification to

the standard two-layer feedforward network. It has a connection
from the input directly to the second layer. Derive the backpropa-
gation algorithm for this network.
Input Layer 1 Layer 2
W2,1
p
a1 a2
W1 S1 x 1
W2 S2 x 1
Rx1 n1 n2
S1 x R
S1 x 1
f1 S2 x S1
S2 x 1
f2
1 b1 1 b2
S1 x 1 S2 x 1
R S1 S2
Figure P11.10 Network with Bypass Connection

We begin with the forward equations:
1 1 1
n = W p+b ,
1 1 1 1 1 1
a = f (n ) = f (W p + b ) ,
2 2 1 2 1 2
n = W a +W p+b ,
2 2 2 2 2 1 2 1 2
a = f (n ) = f (W a + W p+b ).
The backpropagation equations for the sensitivities will not change from
those for a standard two-layer network. The sensitivities are the deriva-
tives of the squared error with respect to the net inputs; these derivatives
dont change, since we are simply adding a term to the net input.
Next we need the elements of the gradient for the weight update equations.
For the standard weights and biases we have
11-37
11 Backpropagation
n i
m
F F
----------
m
- = --------
m
- ----------
m
- = sm m1
i aj ,
w i j n i w i j
m
F F n
--------m- = --------m- --------im- = s m
i .
b i n i b i
1 1 2 2
Therefore the update equations for W , b , W and b do not change. We
2 1
do need an additional equation for W :
n i
2
n i
2
F F
------------
- = -------
- 2 1 = s i 2 1- .
------------
- 2
------------
2 1 2
w i j n i w i j w i j
To find the derivative on the right-hand side of this equation note that
1
S R
2 1
w i j p j + bi .
2 2 1 2
ni = w i j a j +
j=1 j=1
Therefore
n i
2
F
------------
- = p j and ------------- = s 2i p j .
2 1 2 1
w i j w i j
The update equations can thus be written in matrix form as:
m m m m1 T
W k + 1 = W k s a , m = 1 2 ,
m m m
b k + 1 = b k s , m = 1 2 .
2 1 2 1 2 0 T 2 1 2 T
W k + 1 = W k s a = W k s p .
The main point of this problem is that the backpropagation concept can be
used on networks more general than the standard multilayer feedforward
network.
P11.9 Find an algorithm, based on the backpropagation concept, that

can be used to update the weights w 1 and w 2 in the recurrent net-
work shown in Figure P11.11.
11-38
Solved Problems
Inputs Linear Recurrent Layer
w2 a(k)
n(k) a(k + 1)
D
w1
p(k)
a(0)
a(k + 1) = purelin (w1 p(k) + w2 a(k))
Figure P11.11 Linear Recurrent Network

The first step is to define our performance index. As with the multilayer
networks, we will use squared error:
2 2
Fx = tk ak = e k .
For our weight updates we will use the steepest descent algorithm:
F x .
w i =
wi
These derivatives can be computed as follows:
a k
t k a k = 2 t k a k ------------- .
2
Fx =
wi wi w i
Therefore, the key terms we need to compute are
a k
------------- .
w i
To compute these terms we first need to write out the network equation:
a k + 1 = purelin w 1 p k + w 2 a k = w 1 p k + w 2 a k .
Next we take the derivative of both sides of this equation with respect to
the network weights:
a k + 1 a k
----------------------- = p k + w 2 ------------- ,
w 1 w 1
a k + 1 - a k
---------------------- = a k + w 2 ------------- .
w 2 w 2
11
11-39
11 Backpropagation
(Note that we had to take account of the fact that a k is itself a function
of w 1 and w 2 .) These two recursive equations are then used to compute the
derivatives needed for the steepest descent weight update. The equations
are initialized with
a 0 a 0
-------------- = 0 , -------------- = 0 ,
w 1 w 2
since the initial condition is not a function of the weight.

To illustrate the process, lets say that a 0 = 0 . The first network update
would be
a 1 = w1 p 0 + w2 a 0 = w1 p 0 .
The first derivatives would be computed:
a 1 a 0 a 1 a 0
-------------- = p 0 + w 2 -------------- = p 0 , -------------- = a 0 + w 2 -------------- = 0 .
w 1 w 1 w 2 w 2
The first weight updates would be
F x a 1
w i = = 2 t 1 a 1 --------------
wi w i
w 1 = 2 t 1 a 1 p 0
w 2 = 2 t 1 a 1 0 = 0 .
This algorithm is a type of dynamic backpropagation, in which the gradient

is computed by means of a difference equation.
P11.10 Show that backpropagation reduces to the LMS algorithm for a

single-layer linear network (ADALINE).
The sensitivity calculation for a single-layer linear network would be:
s = 2F (n ) t a = 2 I t a = 2e ,
1 1 1
The weight update (Eq. (11.46) and Eq. (11.47)) would be
1 1 1 0 T 1 T 1 T
W k + 1 = W k s a = W k 2e p = W k + 2ep
1 1 1 1 1
b k + 1 = b k s = b k 2e = b k + 2e .
This is identical to the LMS algorithm of Chapter 10.
11-40
Epilogue
Epilogue
In this chapter we have presented the multilayer perceptron network and
the backpropagation learning rule. The multilayer network is a powerful
extension of the single-layer perceptron network. Whereas the single-layer
network is only able to classify linearly separable patterns, the multilayer
network can be used for arbitrary classification problems. In addition, mul-
tilayer networks can be used as universal function approximators. It has
been shown that a two-layer network, with sigmoid-type transfer functions
in the hidden layer, can approximate any practical function, given enough
neurons in the hidden layer.
The backpropagation algorithm is an extension of the LMS algorithm that
can be used to train multilayer networks. Both LMS and backpropagation
are approximate steepest descent algorithms that minimize squared error.
The only difference between them is in the way in which the gradient is cal-
culated. The backpropagation algorithm uses the chain rule in order to
compute the derivatives of the squared error with respect to the weights
and biases in the hidden layers. It is called backpropagation because the
derivatives are computed first at the last layer of the network, and then
propagated backward through the network, using the chain rule, to com-
pute the derivatives in the hidden layers.
One of the major problems with backpropagation has been the long train-
ing times. It is not feasible to use the basic backpropagation algorithm on
practical problems, because it can take weeks to train a network, even on
a large computer. Since backpropagation was first popularized, there has
been considerable work on methods to accelerate the convergence of the al-
gorithm. In Chapter 12 we will discuss the reasons for the slow conver-
gence of backpropagation and will present several techniques for
improving the performance of the algorithm.
Another key problem in training multilayer networks is overfitting. The
network may memorize the data in the training set, but fail to generalize
to new situations. In Chapter 13 we will describe in detail training proce-
dures that can be used to produce networks with excellent generalization.
This chapter has focused mainly on the theoretical development of the
backpropagation learning rule for training multilayer networks. Practical
aspects of training networks with this method are discussed in Chapter 22.
Real-world case studies that demonstrate how to train and validate multi-
layer networks are provided in Chapter 23 (function approximation), Chap-
ter 24 (probability estimation) and Chapter 25 (pattern recognition).
11-41
11 Backpropagation
Further Reading
[HoSt89] K. M. Hornik, M. Stinchcombe and H. White, Multilayer
feedforward networks are universal approximators, Neu-
ral Networks, vol. 2, no. 5, pp. 359366, 1989.
This paper proves that multilayer feedforward networks
with arbitrary squashing functions can approximate any
Borel integrable function from one finite dimensional space
to another finite dimensional space.
[LeCu85] Y. Le Cun, Une procedure dapprentissage pour reseau a
seuil assymetrique, Cognitiva, vol. 85, pp. 599604, 1985.
Yann Le Cun discovered the backpropagation algorithm at
about the same time as Parker and Rumelhart, Hinton and
Williams. This paper describes his algorithm.
[Park85] D. B. Parker, Learning-logic: Casting the cortex of the hu-
man brain in silicon, Technical Report TR-47, Center for
Computational Research in Economics and Management
Science, MIT, Cambridge, MA, 1985.
David Parker independently derived the backpropagation
algorithm at about the same time as Le Cun and Rumel-
hart, Hinton and Williams. This report describes his algo-
rithm.
[RuHi86] D. E. Rumelhart, G. E. Hinton and R. J. Williams, Learn-
ing representations by back-propagating errors, Nature,
vol. 323, pp. 533536, 1986.
This paper contains the most widely publicized description
of the backpropagation algorithm.
Cognition, vol. 1, Cambridge, MA: MIT Press, 1986.
This book was one of the two key influences in the resur-
gence of interest in the neural networks field during the
1980s. Among other topics, it presents the backpropagation
algorithm for training multilayer neural networks.
[Werbo74] P. J. Werbos, Beyond regression: New tools for prediction
and analysis in the behavioral sciences, Ph.D. Thesis,
Harvard University, Cambridge, MA, 1974.
This Ph.D. thesis contains what appears to be the first de-
scription of the backpropagation algorithm (although that
11-42
Further Reading
name is not used). The algorithm is described here in the

context of general networks, with neural networks as a spe-
cial case. Backpropagation did not become widely known
until it was rediscovered in the mid 1980s by Rumelhart,
Hinton and Williams [RuHi86], David Parker [Park85] and
Yann Le Cun [LeCu85].
11
11-43
11 Backpropagation
Exercises
E11.1 Design multilayer networks to perform the classifications illustrated in
Figure E11.1. The network should output a 1 whenever the input vector is
in the shaded region (or on the boundary) and a -1 otherwise. Draw the net-
work diagram in abbreviated notation and show the weight matrices and
bias vectors.
i. ii.
iii. iv.
Figure E11.1 Pattern Classification Tasks
E11.2 Choose the weights and biases for the 1-2-1 network shown in Figure 11.4
so that the network response passes through the points indicated by the
blue circles in Figure E11.2.
Use the MATLAB Neural Network Design Demonstration Two-Layer Net-
work Function (nnd11nf) to check your result.
11-44
Exercises
4 4
3 i. 3 ii.
2 2
t 1 t 1
0 0
1 1
2 2
2 1 0 1 2 2 1 0 1 2
p p
4 4
3 iii. 3 iv.
2 2
t 1 t 1
0 0
1 1
2 2
2 1 0 1 2 2 1 0 1 2
p p
Figure E11.2 Function Approximation Tasks
E11.3 Find a single-layer network that has the same input/output characteristic
as the network in Figure E11.3.
Inputs Linear Layer 1 Linear Layer 2
n 11 a 11

-2
p1
-1 1 n 21 a 21
1
-0.5

1 n 12 a 12 1 0.5
p2
3 1
-0.5
1
Figure E11.3 Two-Layer Linear Network
11-45
11 Backpropagation
E11.4 Use the chain rule to find the derivative f w in the following cases:
2
i. f n = sin n , n w = w .
ii. f n = tanh n , n w = 5w .
iii. f n = exp n , n w = cos w .
iv. f n = logsig n , n w = exp w .
E11.5 Consider again the backpropagation example that begins on page 11-14.
2
i. Find the squared error e as an explicit function of all weights and
biases.
2 1
ii. Using part (i), find e w 1 1 at the initial weights and biases.
iii. Compare the results of part (ii) with the backpropagation results de-
scribed in the text.
E11.6 For the network shown in Figure E11.4 the initial weights and biases are
chosen to be
1 1 2 2
w 0 = 1 , b 0 = 2 , w 0 = 1 , b 0 = 1 .
The network transfer functions are
1
f n = n , f n = --- ,
1 2 2
n
and an input/target pair is given to be
p = 1 t = 1 .
Inputs Layer 1 Layer 2
w1 n1 a1 w2 n2 a2
p f1 f2
b1 b2
1 1
a 1 = f 1(w 1 p + b 1) a 2 = f 2 (w 2 a 1 + b 2)
Figure E11.4 Two-Layer Network for Exercise E11.6
11-46
Exercises
E11.7 Consider the two-layer network in Figure E11.5.
Inputs Linear Layer Linear Layer
1 1
1
w 1,1 n1 a1
S w21,1
b11 n11 a2
p
1 S
1
n2 a12 1
b1
1
w 2,1
S 2
w 1,2 1
b12
1
1 1 1 2 2 1 2
a = purelin(W p + b ) a = purelin(W a + b )
with the following input and target: p 1 = 1 t 1 = 2 . The initial weights

and biases are given by
1
W 0 = 1 , W2 0 = 1 2 , b2 0 =
1 1 , b 0 = 3
1 1
i. Apply the input to the network and make one pass forward through
the network to compute the output and the error.
ii. Compute the sensitivities by backpropagating through the network.
2 1
iii. Compute the derivative e w 1 1 using the results of part ii.
(Very little calculation is required here.)
E11.8 For the network shown in Figure E11.6 the neuron transfer function is
1 2
f n = n ,
and an input/target pair is given to be
1 t = 8 .
p =
1 2
11-47
11 Backpropagation
Inputs Layer 1
1 n 11 a 11
p1
-1
f1
1
1
1 n 12 a 12
p2
1 f1
1
1
Figure E11.6 Single-Layer Network for Exercise E11.8
E11.9 We want to train the network in Figure E11.7 using the standard back-
propagation algorithm (approximate steepest descent).
Inputs Square Law Neuron
p1 w1,1
n a
S f(n) = n2
p2 w1,2 b
1
2
a = (Wp + b)
Figure E11.7 Square Law Neuron
The following input and target are given:
1 t=
p = 0
1
The initial weights and bias are
W 0 = 1 1 , b 0 = 1 .
i. Propagate the input forward through the network.

ii. Compute the error.
iii. Propagate the sensitivities backward through the network.
11-48
Exercises
iv. Compute the gradient of the squared error with respect to the
weights and bias.
v. Update the weights and bias (assume a learning rate of = 0.1).
E11.10 Consider the following multilayer perceptron network. (The transfer func-
2
tion of the hidden layer is f n = n .)
Inputs Square
- Title
Law 1 Layer
- Linear
- TitleLayer
2-
p a1 a2
2x1 W1 2x1 W2 1x1
2x2 n1 1x2 n2
2x1 1x1
1 b1 1 b2
2x1 1x1
- Exp 1 -1p+b1)
a1 = square(W - Exp 2 2-a1+b2)
a2 = purelin(W
Figure E11.8 Two-Layer Square Law Network
The initial weights and biases are:
W 0 = 1 1 , W 0 = 2 1 , b 0 = 1 , b 0 = 1 .
1 2 1 2
1 0 1
Perform one iteration of the standard steepest descent backpropagation

(use matrix operations) with learning rate = 0.5 for the following input/
target pair:
1 t=
p = 2
1
E11.11 Consider the network shown in Figure E11.9.
11-49
11 Backpropagation
w1 n1 a1 w2 n2 a2
p
SxR

SxR
b1 b2
1 1
a1 = (w1p+b1) 3 a2 = (w2a1+b2) 2
The initial weights and biases are chosen to be

1 1 2 2
w 0 = 2 b 0 = 1 w 0 = 1 b 0 = 2 .
p 1 = 1 t 1 = 0 ,
Perform one iteration of backpropagation (steepest descent) with = 1.
E11.12 Consider the multilayer perceptron network in Figure E11.10. (The trans-
3
fer function of the hidden layer is f n = n .)
Inputs Cubic Law Layer Linear Layer
p1 1
a1 2
a2
2x1
W 2x1
W 1x1
2x2
n1 1x2
n
2
2x1 1x1
1 b 1
1 b 2
1
2 2x1 S 1x1 1
1 1 1 2 2 1 2
a = cube(W p + b ) a = W a +b
Figure E11.10 Cubic Law Neural Network
The initial weights and biases are:
11-50
Exercises
W 0 = 1 1 , b 0 = 1 , W 0 = 1 1 , b 0 = 1 .
1 1 2 2
1 0 2
Perform one iteration of the standard steepest descent backpropagation

(use matrix operations) with learning rate = 0.5 for the following input/
target pair:
1 t =
p = 1 .
1
E11.13 Someone has proposed that the standard multilayer network should be
modified to include a scalar gain at each layer. This means that the net in-
put at layer m would be computed as
m m m m1 m
n = W a +b ,
m
where is the scalar gain at layer m. This gain would be trained like the
weights and biases of the network. Modify the backpropagation algorithm
(Eq. (11.41) to Eq. (11.47)) for this new network. (There will need to be a
m
new equation added to update , but some of the other equations may
have to be modified as well.)
E11.14 Consider the two-layer network shown in Figure E11.11.
Inputs Log-Sigmoid Layer Linear Layer
1 1
n1 a1
2 S 2
p -1 n2 a2
1 S
1
1
n2 a12
S 3
1
-1
1
1

i. If p = 1, use a (slightly) modified form of backpropagation (as devel-
oped in Eq. (11.41) through Eq. (11.47)) to find
11-51
11 Backpropagation
2 2 2
a a a
2
, 1
, 1
.
n n1 n2
2
a
ii. Use the results of i. and the chain rule to find
.
p
Your answers to both parts should be numerical.
E11.15 Consider the network shown in Figure E11.12, where the inputs to the neu-
ron involve both the original inputs and their product. This is a type of
higher-order network.
Inputs Log-Sigmoid Layer
p1 w1
w1,2 n a
S
p2 w2 b
1
a = logsig(w1 p1 + w2 p2 + w1,2 p1 p2 + b)
Figure E11.12 Higher-Order Network

i. Find a learning rule for the network parameters, using the approx-
imate steepest descent algorithm (as was done for backpropaga-
tion).
ii. For the following initial parameter values, inputs and target, per-
form one iteration of your learning rule with learning rate = 1:
w 1 = 1 w 2 = 1 w 1 2 = 0.5 b 1 = 1 p 1 = 0 p 2 = 1 t = 0.75
E11.16 In Figure E11.13 we have a two-layer network that has an additional con-
nection from the input directly to the second layer. Derive the backpropa-
gation algorithm for this network.
11-52
Exercises
w 2,1
w1 n 1 a 1
w2 n2 a2
p f1 f2
b 1 b2
1 1
a 1 = f 1(w 1 p + b 1) a 2 = f 2 (w 2 a 1 + w 2,1 p + b 2)
Figure E11.13 Two-Layer Network with Bypass Connection
E11.17 In the multilayer network, the net input is computed as follows

m
S
w i j
m+1 m+1 m m+1 m+1 m+1 m m+1
n = W a +b or ni = aj + bi .
j=1
If the net input calculation is changed to the following equation (squared

distance calculation), how will the sensitivity backpropagation (Eq.
(11.35)) change?
m
S
m 2
w i j
m+1 m+1
ni = aj
j=1
E11.18 Consider again the net input calculation, as described in Exercise E11.17.
If the net input calculation is changed to the following equation (multiply
by the bias, instead of add), how will the sensitivity backpropagation (Eq.
(11.35)) change?
m
S
= w i j a j b i
m+1 m+1 m m+1
ni .

j = 1
E11.19 Consider the system shown in Figure E11.14. There are a series of stages,
with different transfer functions in each stage. (There are no weights or bi-
M
ases.) We want to take the derivative of the output of this system ( a ) with
respect to the input of the system ( p ). Derive a recursive algorithm that
you can use to compute this derivative. Use the concepts that we used to
derive the backpropagation algorithm, and use the following intermediate
variable in your algorithm:
11-53
11 Backpropagation
M
a
q = ---------i- .
i
a
M
a
1
a
2
a
3
a
p f1 f2 f
3
f
M
Figure E11.14 Cascade System
E11.20 The backpropagation algorithm is used to compute the gradient of the

squared error with respect to the weights and biases of a multilayer net-
work. How would the algorithm be changed if you wanted to compute the
gradient with respect to the inputs of the network (i.e., with respect to the
elements of the input vector p )? Carefully explain all of your steps, and
write out the final algorithm.
E11.21 With the standard backpropagation algorithm, we want to compute the de-
rivative
F-
------ .
w
To calculate this derivative, we use the chain rule in the form
F F n
------- = ------ ------- .
w n w
Suppose that we want to use Newton's method. We would need to find the
second derivative
2
F
--------2- .
w
What form will the chain rule take in this case?
E11.22 The standard steepest descent backpropagation algorithm, which is sum-

marized in Eq. (11.41) through Eq. (11.47), was designed to minimize the
performance function that was the sum of squares of the network errors, as
given in Eq. (11.12). Suppose that we want to change the performance func-
tion to the sum of the fourth powers of the errors (e4) plus the sum of the
squares of the weights and biases in the network. Show how Eq. (11.41)
11-54
Exercises
through Eq. (11.47) will change for this new performance function. (You
don't need to rederive any steps which are already given in this chapter and
do not change.)
E11.23 Repeat Problem P11.4 using the backward method described below.
In Problem P11.4. we had the dynamic system
y k + 1 = fyk .
We had to choose the initial condition y 0 so that at some final time k = K

the system output y K would be as close as possible to some target output
t . We minimized the performance index
2 2
Fy0 = t yK = e K
using steepest descent, so we needed the gradient
Fy0 .
y0
We developed a procedure for computing this gradient using the chain rule.
The procedure involved a recursive equation for the term
y k ,
rk
y 0
which evolved forward in time. The gradient can also be computed in a dif-
ferent way by evolving the term
2
qk e K
yk
backward through time.
E11.24 Consider the recurrent neural network in Figure E11.15.
p(k) 1
n(k) a(k)
S
a(k-1) w
D
a(k) = p(k) + w a(k-1)
Figure E11.15 Recurrent Network
11-55
11 Backpropagation
We want to find the weight value w so that at some final time k = K the
system output a K will be as close as possible to some target output t . We
2
will minimize the performance index F w = t a K using steepest de-
scent, so we need the gradient F w /w .
i. Find a general procedure to compute this gradient using the chain
rule. Develop an equation to evolve the following term forward
through time:

s k ak .
w
Show each step of your entire procedure carefully. This will involve
updating s k and also computing the gradient F w /w .
ii. Assume that K = 3 . Write out the complete expression for a 3 as
a function of p 1 , p 2 , p 3 and w (assuming a 0 = 0 ). Take the
derivative of this expression with respect to w , and show that it
equals s 3 .
E11.25 Write a MATLAB program to implement the backpropagation algorithm

1
for a 1 S 1 network. Write the program using matrix operations, as in
2+2
Eq. (11.41) to Eq. (11.47). Choose the initial weights and biases to be ran-
ans = dom numbers uniformly distributed between -0.5 and 0.5 (using the MAT-
4 LAB function rand), and train the network to approximate the function

g p = 1 + sin --- p for 2 p 2 .
2
1 1
Use S = 2 and S = 10 . Experiment with several different values for the
learning rate , and use several different initial conditions. Discuss the
convergence properties of the algorithm as the learning rate changes.
11-56
Objectives
12 Variations on
Backpropagation
Objectives 12-1
Drawbacks of Backpropagation 12-3
Performance Surface Example 12-3
Convergence Example 12-7
Heuristic Modifications to Backpropagation 12-9
Momentum 12-9
Variable Learning Rate 12-12
Numerical Optimization Techniques 12-14
Levenberg-Marquardt Algorithm 12-19
Epilogue 12-46
Exercises 12-50
Objectives
The backpropagation algorithm introduced in Chapter 11 was a major
breakthrough in neural network research. However, the basic algorithm is
too slow for most practical applications. In this chapter we present several
variations of backpropagation that provide significant speedup and make
the algorithm more practical.
We will begin by using a function approximation example to illustrate why
the backpropagation algorithm is slow in converging. Then we will present
several modifications to the algorithm. Recall that backpropagation is an
approximate steepest descent algorithm. In Chapter 9 we saw that steepest
descent is the simplest, and often the slowest, minimization method. The
conjugate gradient algorithm and Newtons method generally provide fast-
er convergence. In this chapter we will explain how these faster procedures
can be used to speed up the convergence of backpropagation.
12-1
Theory and Examples

When the basic backpropagation algorithm is applied to a practical prob-
lem the training may take days or weeks of computer time. This has en-
couraged considerable research on methods to accelerate the convergence
of the algorithm.
The research on faster algorithms falls roughly into two categories. The
first category involves the development of heuristic techniques, which arise
out of a study of the distinctive performance of the standard backpropaga-
tion algorithm. These heuristic techniques include such ideas as varying
the learning rate, using momentum and rescaling variables (e.g.,
[VoMa88], [Jacob88], [Toll90] and [RiIr90]). In this chapter we will discuss
the use of momentum and variable learning rates.
Another category of research has focused on standard numerical optimiza-
tion techniques (e.g., [Shan90], [Barn92], [Batt92] and [Char92]). As we
have discussed in Chapters 10 and 11, training feedforward neural net-
works to minimize squared error is simply a numerical optimization prob-
lem. Because numerical optimization has been an important research
subject for 30 or 40 years (see Chapter 9), it seems reasonable to look for
fast training algorithms in the large number of existing numerical optimi-
zation techniques. There is no need to reinvent the wheel unless absolute-
ly necessary. In this chapter we will present two existing numerical
optimization techniques that have been very successfully applied to the
training of multilayer perceptrons: the conjugate gradient algorithm and
the Levenberg-Marquardt algorithm (a variation of Newtons method).
We should emphasize that all of the algorithms that we will describe in this
chapter use the backpropagation procedure, in which derivatives are pro-
cessed from the last layer of the network to the first. For this reason they
could all be called backpropagation algorithms. The differences between
the algorithms occur in the way in which the resulting derivatives are used
to update the weights. In some ways it is unfortunate that the algorithm
we usually refer to as backpropagation is in fact a steepest descent algo-
rithm. In order to clarify our discussion, for the remainder of this chapter
we will refer to the basic backpropagation algorithm as steepest descent
SDBP backpropagation (SDBP).
In the next section we will use a simple example to explain why SDBP has
problems with convergence. Then, in the following sections, we will present
various procedures to improve the convergence of the algorithm.
12-2
Drawbacks of Backpropagation
Recall from Chapter 10 that the LMS algorithm is guaranteed to converge
to a solution that minimizes the mean squared error, so long as the learn-
ing rate is not too large. This is true because the mean squared error for a
single-layer linear network is a quadratic function. The quadratic function
has only a single stationary point. In addition, the Hessian matrix of a qua-
dratic function is constant, therefore the curvature of the function in a giv-
en direction does not change, and the function contours are elliptical.
SDBP is a generalization of the LMS algorithm. Like LMS, it is also an ap-
proximate steepest descent algorithm for minimizing the mean squared er-
ror. In fact, SDBP is equivalent to the LMS algorithm when used on a
single-layer linear network. (See Problem P11.10.) When applied to multi-
layer networks, however, the characteristics of SDBP are quite different.
This has to do with the differences between the mean squared error perfor-
mance surfaces of single-layer linear networks and multilayer nonlinear
networks. While the performance surface for a single-layer linear network
has a single minimum point and constant curvature, the performance sur-
face for a multilayer network may have many local minimum points, and
the curvature can vary widely in different regions of the parameter space.
This will become clear in the example that follows.
Performance Surface Example

To investigate the mean squared error performance surface for multilayer
networks we will employ a simple function approximation example. We will
use the 1-2-1 network shown in Figure 12.1, with log-sigmoid transfer func-
tions in both layers.
Input Log-Sigmoid Layer Log-Sigmoid Layer
w11,1 n11 a11

w21,1
b11 n2 a2
p
1
n12 a12 b2
w12,1 w21,2
1
b12
1
a1 = logsig (W1p + b1) a2 = logsig (W2a1 + b2)
Figure 12.1 1-2-1 Function Approximation Network

In order to simplify our analysis, we will give the network a problem for
which we know the optimal solution. The function we will approximate is
12-3
the response of the same 1-2-1 network, with the following values for the
weights and biases:
1 1 1 1
w 1 1 = 10 , w 2 1 = 10 , b 1 = 5 , b 2 = 5 , (12.1)
2 2 2
w 1 1 = 1 , w 1 2 = 1 , b = 1 . (12.2)
The network response for these parameters is shown in Figure 12.2, which
2
plots the network output a as the input p is varied over the range 2 2 .
1
0.75
2
a 0.5
0.25
0
-2 -1 0 1 2
Figure 12.2 Nominal Function

We want to train the network of Figure 12.1 to approximate the function
displayed in Figure 12.2. The approximation will be exact when the net-
work parameters are set to the values given in Eq. (12.1) and Eq. (12.2).
This is, of course, a very contrived problem, but it is simple and it illus-
trates some important concepts.
Lets now consider the performance index for our problem. We will assume
that the function is sampled at the values
p = 2 1.9 1.8 1.9 2 , (12.3)
and that each occurs with equal probability. The performance index will be
the sum of the squared errors at these 41 points. (We wont bother to find
the mean squared error, which just requires dividing by 41.)
In order to be able to graph the performance index, we will vary only two
parameters at a time. Figure 12.3 illustrates the squared error when only
1 2
w 1 1 and w 1 1 are being adjusted, while the other parameters are set to
their optimal values given in Eq. (12.1) and Eq. (12.2). Note that the mini-
1 2
mum error will be zero, and it will occur when w 1 1 = 10 and w 1 1 = 1 , as
indicated by the open blue circle in the figure.
12-4
There are several features to notice about this error surface. First, it is
clearly not a quadratic function. The curvature varies drastically over the
parameter space. For this reason it will be difficult to choose an appropri-
ate learning rate for the steepest descent algorithm. In some regions the
surface is very flat, which would allow a large learning rate, while in other
regions the curvature is high, which would require a small learning rate.
(Refer to discussions in Chapters 9 and 10 on the choice of learning rate for
the steepest descent algorithm.)
It should be noted that the flat regions of the performance surface should
not be unexpected, given the sigmoid transfer functions used by the net-
work. The sigmoid is very flat for large inputs.
A second feature of this error surface is the existence of more than one local
1
minimum point. The global minimum point is located at w 1 1 = 10 and
2 1
w 1 1 = 1 , along the valley that runs parallel to the w 1 1 axis. However,
there is also a local minimum, which is located in the valley that runs par-
2
allel to the w 1 1 axis. (This local minimum is actually off the graph at
1 2
w 1 1 = 0.88 , w 1 1 = 38.6 .) In the next section we will investigate the per-
formance of backpropagation on this surface.
15
10
10
5
2
w 1 1 5
0
0 -5
0 -5
5 0
2
w 1 1 5 1
10
10
w 1 1
-5 15 15
-5 0 5 10 15
1
w 1 1
1 2
Figure 12.3 Squared Error Surface Versus w 1 1 and w 1 1
1 1
Figure 12.4 illustrates the squared error when w 1 1 and b 1 are being ad-
justed, while the other parameters are set to their optimal values. Note
1
that the minimum error will be zero, and it will occur when w 1 1 = 10 and
1
b 1 = 5 , as indicated by the open blue circle in the figure.
Again we find that the surface has a very contorted shape, steep in some
regions and very flat in others. Surely the standard steepest descent algo-
rithm will have some trouble with this surface. For example, if we have an
1 1
initial guess of w 1 1 = 0 , b 1 = 10 , the gradient will be very close to zero,
12-5
and the steepest descent algorithm would effectively stop, even though it is
not close to a local minimum point.
15
2.5
5
2
1.5
1
b -5 1
1
0.5
0
-15 -10
0 20
10
1 10 0
w 1 1 20 -10 b
1
-20 1
-25 30 -30
-10 0 10 20 30
1
w 1 1
1 1
Figure 12.4 Squared Error Surface Versus w 1 1 and b 1
1 1
Figure 12.5 illustrates the squared error when b 1 and b 2 are being adjust-
ed, while the other parameters are set to their optimal values. The mini-
1 1
mum error is located at b 1 = 5 and b 2 = 5 , as indicated by the open blue
circle in the figure.
This surface illustrates an important property of multilayer networks: they
have a symmetry to them. Here we see that there are two local minimum
points and they both have the same value of squared error. The second so-
lution corresponds to the same network being turned upside down (i.e., the
top neuron in the first layer is exchanged with the bottom neuron). It is be-
cause of this characteristic of neural networks that we do not set the initial
weights and biases to zero. The symmetry causes zero to be a saddle point
of the performance surface.
This brief study of the performance surfaces for multilayer networks gives
us some hints as to how to set the initial guess for the SDBP algorithm.
First, we do not want to set the initial parameters to zero. This is because
the origin of the parameter space tends to be a saddle point for the perfor-
mance surface. Second, we do not want to set the initial parameters to large
values. This is because the performance surface tends to have very flat re-
gions as we move far away from the optimum point.
Typically we choose the initial weights and biases to be small random val-
ues. In this way we stay away from a possible saddle point at the origin
without moving out to the very flat regions of the performance surface. (An-
other procedure for choosing the initial parameters is described in
[NgWi90].) As we will see in the next section, it is also useful to try several
different initial guesses, in order to be sure that the algorithm converges to
a global minimum point.
12-6
10
1.4
0.7
2
b1 0
0
-5 -10
-5 -10
0 -5
1
b1 0 1
5 b
5 2
-10 10 10
-10 -5 0 5 10
1
b 1
1 1
Figure 12.5 Squared Error Surface Versus b 1 and b 2
Convergence Example
Now that we have examined the performance surface, lets investigate the
performance of SDBP. For this section we will use a variation of the stan-
Batching dard algorithm, called batching, in which the parameters are updated only
after the entire training set has been presented. The gradients calculated
at each training example are averaged together to produce a more accurate
estimate of the gradient. (If the training set is complete, i.e., covers all pos-
sible input/output pairs, then the gradient estimate will be exact.)
In Figure 12.6 we see two trajectories of SDBP (batch mode) when only two
1 2
parameters, w 1 1 and w 1 1 are adjusted. For the initial condition labeled
a the algorithm does eventually converge to the optimal solution, but the
convergence is slow. The reason for the slow convergence is the change in
curvature of the surface over the path of the trajectory. After an initial
moderate slope, the trajectory passes over a very flat surface, until it falls
into a very gently sloping valley. If we were to increase the learning rate,
the algorithm would converge faster while passing over the initial flat sur-
face, but would become unstable when falling into the valley, as we will see
in a moment.
Trajectory b illustrates how the algorithm can converge to a local mini-
mum point. The trajectory is trapped in a valley and diverges from the op-
timal solution. If allowed to continue the trajectory converges to
1 2
w 1 1 = 0.88 , w 1 1 = 38.6 . The existence of multiple local minimum points
is typical of the performance surface of multilayer networks. For this rea-
son it is best to try several different initial guesses in order to ensure that
a global minimum has been obtained. (Some of the local minimum points
may have the same value of squared error, as we saw in Figure 12.5, so we
would not expect the algorithm to converge to the same parameter values
for each initial guess. We just want to be sure that the same minimum error
is obtained.)
12-7
15
10 b
2 5
w 1 1
a
-5
-5 0 5 10 15
1
w 1 1
Figure 12.6 Two SDBP (Batch Mode) Trajectories

The progress of the algorithm can also be seen in Figure 12.7, which shows
the squared error versus the iteration number. The curve on the left corre-
sponds to trajectory a and the curve on the right corresponds to trajectory
b. These curves are typical of SDBP, with long periods of little progress
and then short periods of rapid advance.
2 6
a b
1.5
4
2
0.5
0 0 2 4
0 0 1 2 3
10 10 10 10 10 10 10
Iteration Number Iteration Number
Figure 12.7 Squared Error Convergence Patterns

We can see that the flat sections in Figure 12.7 correspond to times when
the algorithm is traversing a flat section of the performance surface, as
shown in Figure 12.6. During these periods we would like to increase the
learning rate, in order to speed up convergence. However, if we increase the
learning rate the algorithm will become unstable when it reaches steeper
portions of the performance surface.
This effect is illustrated in Figure 12.8. The trajectory shown here corre-
sponds to trajectory a in Figure 12.6, except that a larger learning rate
was used. The algorithm converges faster at first, but when the trajectory
reaches the narrow valley that contains the minimum point the algorithm
begins to diverge. This suggests that it would be useful to vary the learning
rate. We could increase the learning rate on flat surfaces and then decrease
the learning rate as the slope increased. The question is: How will the al-
12-8
Heuristic Modifications of Backpropagation
gorithm know when it is on a flat surface? We will discuss this in a later

section.
15
1.5
Squared Error
10
1
2
w 5
1 1
0.5
0 0 1 2 3
10 10 10 10
Iteration Number
-5
-5 0 5 10 15
1
w 1 1
Figure 12.8 Trajectory with Learning Rate Too Large

Another way to improve convergence would be to smooth out the trajectory.
Note in Figure 12.8 that when the algorithm begins to diverge it is oscillat-
ing back and forth across a narrow valley. If we could filter the trajectory,
by averaging the updates to the parameters, this might smooth out the os-
cillations and produce a stable trajectory. We will discuss this procedure in
the next section.
To experiment with this backpropagation example, use the MATLAB Neu-
ral Network Design Demonstration Steepest Descent Backpropagation
(nnd12sd).

Now that we have investigated some of the drawbacks of backpropagation
(steepest descent), lets consider some procedures for improving the algo-
rithm. In this section we will discuss two heuristic methods. In a later sec-
tion we will present two methods based on standard numerical optimiza-
tion algorithms.
Momentum
The first method we will discuss is the use of momentum. This is a modifi-
cation based on our observation in the last section that convergence might
be improved if we could smooth out the oscillations in the trajectory. We
can do this with a low-pass filter.
Before we apply momentum to a neural network application, lets investi-
gate a simple example to illustrate the smoothing effect. Consider the fol-
lowing first-order filter:
12-9
y k = y k 1 + 1 w k , (12.4)
where w k is the input to the filter, y k is the output of the filter and
is the momentum coefficient that must satisfy
01. (12.5)
The effect of this filter is shown in Figure 12.9. For these examples the in-
put to the filter was taken to be the sine wave:
w k = 1 + sin --------- ,
2k
(12.6)
16
and the momentum coefficient was set to = 0.9 (left graph) and = 0.98
(right graph). Here we can see that the oscillation of the filter output is less
than the oscillation in the filter input (as we would expect for a low-pass
filter). In addition, as is increased the oscillation in the filter output is
reduced. Notice also that the average filter output is the same as the aver-
age filter input, although as is increased the filter output is slower to re-
spond.
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 50 100 150 200 0 50 100 150 200
a) = 0.9 b) = 0.98
Figure 12.9 Smoothing Effect of Momentum

To summarize, the filter tends to reduce the amount of oscillation, while
still tracking the average value. Now lets see how this works on the neural
network problem. First, recall that the parameter updates for SDBP (Eq.
(11.46) and Eq. (11.47)) are
m m m1 T
W k = s a , (12.7)
m m
b k = s . (12.8)
12-10
Momentum When the momentum filter is added to the parameter changes, we obtain
the following equations for the momentum modification to backpropaga-
MOBP tion (MOBP):
m m m m1 T
W k = W k 1 1 s a , (12.9)
m m m
b k = b k 1 1 s . (12.10)
If we now apply these modified equations to the example in the preceding

section, we obtain the results shown in Figure 12.10. (For this example we
have used a batching form of MOBP, in which the parameters are updated
only after the entire training set has been presented. The gradients calcu-
lated at each training example are averaged together to produce a more ac-
curate estimate of the gradient.) This trajectory corresponds to the same
initial condition and learning rate shown in Figure 12.8, but with a momen-
tum coefficient of = 0.8 . We can see that the algorithm is now stable. By
the use of momentum we have been able to use a larger learning rate, while
maintaining the stability of the algorithm. Another feature of momentum
is that it tends to accelerate convergence when the trajectory is moving in
a consistent direction.
15
1.5
Squared Error
10
2 5
w 1 1
0.5
0
0 0 1 2 3
10 10 10 10
Iteration Number
-5
-5 0 5 10 15
1
w 1 1
Figure 12.10 Trajectory with Momentum

If you look carefully at the trajectory in Figure 12.10, you can see why the
procedure is given the name momentum. It tends to make the trajectory
continue in the same direction. The larger the value of , the more mo-
mentum the trajectory has.
To experiment with momentum, use the MATLAB Neural Network Design
Demonstration Momentum Backpropagation (nnd12mo).
12-11
Variable Learning Rate

We suggested earlier in this chapter that we might be able to speed up con-
vergence if we increase the learning rate on flat surfaces and then decrease
the learning rate when the slope increases. In this section we want to ex-
plore this concept.
Recall that the mean squared error performance surface for single-layer
linear networks is always a quadratic function, and the Hessian matrix is
therefore constant. The maximum stable learning rate for the steepest de-
scent algorithm is two divided by the maximum eigenvalue of the Hessian
matrix. (See Eq. (9.25).)
As we have seen, the error surface for the multilayer network is not a qua-
dratic function. The shape of the surface can be very different in different
regions of the parameter space. Perhaps we can speed up convergence by
adjusting the learning rate during the course of training. The trick will be
to determine when to change the learning rate and by how much.
There are many different approaches for varying the learning rate. We will
describe a very straightforward batching procedure [VoMa88], where the
Variable Learning Rate learning rate is varied according to the performance of the algorithm. The
VLBP rules of the variable learning rate backpropagation algorithm (VLBP) are:
1. If the squared error (over the entire training set) increases by more
than some set percentage (typically one to five percent) after a
weight update, then the weight update is discarded, the learning rate
is multiplied by some factor 0 1 , and the momentum coefficient
(if it is used) is set to zero.
2. If the squared error decreases after a weight update, then the weight
update is accepted and the learning rate is multiplied by some factor
1 . If has been previously set to zero, it is reset to its original val-
ue.
3. If the squared error increases by less than , then the weight update
is accepted but the learning rate is unchanged. If has been previously
set to zero, it is reset to its original value.
(See Problem P12.3 for a numerical example of VLBP.)
To illustrate VLBP, lets apply it to the function approximation problem of
the previous section. Figure 12.11 displays the trajectory for the algorithm
using the same initial guess, initial learning rate and momentum coeffi-
cient as was used in Figure 12.10. The new parameters were assigned the
values
= 1.05 , = 0.7 and = 4% . (12.11)
12-12
15
10
2
w 1 1 5
-5
-5 0 5 10 15
1
w 1 1
Figure 12.11 Variable Learning Rate Trajectory

Notice how the learning rate, and therefore the step size, tends to increase
when the trajectory is traveling in a straight line with constantly decreas-
ing error. This effect can also be seen in Figure 12.12, which shows the
squared error and the learning rate versus iteration number.
When the trajectory reaches a narrow valley, the learning rate is rapidly
decreased. Otherwise the trajectory would have become oscillatory, and the
error would have increased dramatically. For each potential step where the
error would have increased by more than 4% the learning rate is reduced
and the momentum is eliminated, which allows the trajectory to make the
quick turn to follow the valley toward the minimum point. The learning
rate then increases again, which accelerates the convergence. The learning
rate is reduced again when the trajectory overshoots the minimum point
when the algorithm has almost converged. This process is typical of a
VLBP trajectory.
1.5 60
Squared Error Learning Rate
1 40
0.5 20
0 0 1 2 3
0 0 1 2 3
10 10 10 10 10 10 10 10
Iteration Number Iteration Number
Figure 12.12 Convergence Characteristics of Variable Learning Rate

There are many variations on this variable learning rate algorithm. Jacobs
[Jaco88] proposed the delta-bar-delta learning rule, in which each network
parameter (weight or bias) has its own learning rate. The algorithm in-
creases the learning rate for a network parameter if the parameter change
12-13
has been in the same direction for several iterations. If the direction of the
parameter change alternates, then the learning rate is reduced. The Su-
perSAB algorithm of Tollenaere [Toll90] is similar to the delta-bar-delta
rule, but it has more complex rules for adjusting the learning rates.
Another heuristic modification to SDBP is the Quickprop algorithm of
Fahlman [Fahl88]. It assumes that the error surface is parabolic and con-
cave upward around the minimum point and that the effect of each weight
can be considered independently.
The heuristic modifications to SDBP can often provide much faster conver-
gence for some problems. However, there are two main drawbacks to these
methods. The first is that the modifications require that several parame-
ters be set (e.g., , and ), while the only parameter required for SDBP
is the learning rate. Some of the more complex heuristic modifications can
have five or six parameters to be selected. Often the performance of the al-
gorithm is sensitive to changes in these parameters. The choice of param-
eters is also problem dependent. The second drawback to these
modifications to SDBP is that they can sometimes fail to converge on prob-
lems for which SDBP will eventually find a solution. Both of these draw-
backs tend to occur more often when using the more complex algorithms.
To experiment with VLBP, use the MATLAB Neural Network Design Dem-
onstration Variable Learning Rate Backpropagation (nnd12vl).
Numerical Optimization Techniques

Now that we have investigated some of the heuristic modifications to
SDBP, lets consider those methods that are based on standard numerical
optimization techniques. We will investigate two techniques: conjugate
gradient and Levenberg-Marquardt. The conjugate gradient algorithm for
quadratic functions was presented in Chapter 9. We need to add two proce-
dures to this algorithm in order to apply it to more general functions.
The second numerical optimization method we will discuss in this chapter
is the Levenberg-Marquardt algorithm, which is a modification to New-
tons method that is well-suited to neural network training.
Conjugate Gradient
In Chapter 9 we presented three numerical optimization techniques: steep-
est descent, conjugate gradient and Newtons method. Steepest descent is
the simplest algorithm, but is often slow in converging. Newtons method
is much faster, but requires that the Hessian matrix and its inverse be cal-
culated. The conjugate gradient algorithm is something of a compromise; it
does not require the calculation of second derivatives, and yet it still has
the quadratic convergence property. (It converges to the minimum of a qua-
dratic function in a finite number of iterations.) In this section we will de-
scribe how the conjugate gradient algorithm can be used to train
12-14
multilayer networks. We will call this algorithm conjugate gradient back-

CGBP propagation (CGBP).
Lets begin by reviewing the conjugate gradient algorithm. For ease of ref-
erence, we will repeat the algorithm steps from Chapter 9 (page 9-18):
1. Select the first search direction p 0 to be the negative of the gradient,
as in Eq. (9.59):
p0 = g0 , (12.12)
where
g k F x . (12.13)
x = xk
2. Take a step according to Eq. (9.57), selecting the learning rate k to

minimize the function along the search direction:
xk + 1 = xk + k pk . (12.14)
3. Select the next search direction according to Eq. (9.60), using Eq.
(9.61), Eq. (9.62), or Eq. (9.63) to calculate k :
pk = gk + k pk 1 , (12.15)
with
T T T
g k 1 g k gk gk g k 1 g k
k = --------------------------
T
- or k = ----------------------
T
- or k = ----------------------
T
-. (12.16)
g k 1 p k 1 gk 1 gk 1 gk 1 gk 1
4. If the algorithm has not converged, continue from step 2.

This conjugate gradient algorithm cannot be applied directly to the neural
network training task, because the performance index is not quadratic.
This affects the algorithm in two ways. First, we will not be able to use Eq.
(9.31) to minimize the function along a line, as required in step 2. Second,
the exact minimum will not normally be reached in a finite number of
steps, and therefore the algorithm will need to be reset after some set num-
ber of iterations.
Lets address the linear search first. We need to have a general procedure
for locating the minimum of a function in a specified direction. This will in-
volve two steps: interval location and interval reduction. The purpose of the
interval location step is to find some initial interval that contains a local
minimum. The interval reduction step then reduces the size of the initial
interval until the minimum is located to the desired accuracy.
12-15
We will use a function comparison method [Scal85] to perform the interval

Interval Location location step. This procedure is illustrated in Figure 12.13. We begin by
evaluating the performance index at an initial point, represented by a 1 in
the figure. This point corresponds to the current values of the network
weights and biases. In other words, we are evaluating
F x0 . (12.17)
The next step is to evaluate the function at a second point, represented by

b 1 in the figure, which is a distance from the initial point, along the first
search direction p 0 . In other words, we are evaluating
F x 0 + p 0 . (12.18)
F(x)
8
4
2
x
a1 b1
a2 b2
a3 b3
a4 b4
a5 b5
Figure 12.13 Interval Location

We then continue to evaluate the performance index at new points b i , suc-
cessively doubling the distance between points. This process stops when
the function increases between two consecutive evaluations. In Figure
12.13 this is represented by b 3 to b 4 . At this point we know that the mini-
mum is bracketed by the two points a 5 and b 5 . We cannot narrow the in-
terval any further, because the minimum may occur either in the interval
[a 4, b 4] or in the interval [a 3, b 3] . These two possibilities are illustrated in
Figure 12.14 (a).
Now that we have located an interval containing the minimum, the next
Interval Reduction step in the linear search is interval reduction. This will involve evaluating
the function at points inside the interval [a 5, b 5] , which was selected in the
interval location step. From Figure 12.14 we can see that we will need to
evaluate the function at two internal points (at least) in order to reduce the
size of the interval of uncertainty. Figure 12.14 (a) shows that one internal
function evaluation does not provide us with any information on the loca-
tion of the minimum. However, if we evaluate the function at two points c
and d , as in Figure 12.14 (b), we can reduce the interval of uncertainty. If
12-16
F c F d , as shown in Figure 12.14 (b), then the minimum must occur in

the interval [c, b] . Conversely, if F c F d , then the minimum must oc-
cur in the interval [a, d] . (Note that we are assuming that there is a single
minimum located in the initial interval. More about that later.)
F(x) F(x)
a c b a c d b
(a) Interval is not reduced. (b) Minimum must occur
between c and b.
Figure 12.14 Reducing the Size of the Interval of Uncertainty

The procedure described above suggests a method for reducing the size of
the interval of uncertainty. We now need to decide how to determine the lo-
cations of the internal points c and d . There are several ways to do this
Golden Section Search (see [Scal85]). We will use a method called the Golden Section search,
which is designed to reduce the number of function evaluations required.
At each iteration one new function evaluation is required. For example, in
the case illustrated in Figure 12.14 (b), point a would be discarded and
point c would become the new a . Then point d would become the new point
c , and a new d would be placed between the original points d and b . The
trick is to place the new point so that the interval of uncertainty will be re-
duced as quickly as possible.
The algorithm for the Golden Section search is as follows [Scal85]:
= 0.618
Set c1 = a1 + 1 b1 a1 , Fc = F c1 .
d1 = b1 1 b1 a1 , Fd = F d1 .
For k = 1 2 repeat
If F c F d then
Set ak + 1 = ak ; bk + 1 = dk ; dk + 1 = ck
ck + 1 = ak + 1 + 1 bk + 1 ak + 1
Fd = Fc ; Fc = F ck + 1
else
12-17
Set ak + 1 = ck ; bk + 1 = bk ; ck + 1 = dk
dk + 1 = bk + 1 1 bk + 1 ak + 1
Fc = Fd ; Fd = F dk + 1
end
end until b k + 1 a k + 1 tol
Where tol is the accuracy tolerance set by the user.
(See Problem P12.4 for a numerical example of the interval location and in-
terval reduction procedures.)
There is one more modification to the conjugate gradient algorithm that
needs to be made before we apply it to neural network training. For qua-
dratic functions the algorithm will converge to the minimum in at most n
iterations, where n is the number of parameters being optimized. The
mean squared error performance index for multilayer networks is not qua-
dratic, therefore the algorithm would not normally converge in n itera-
tions. The development of the conjugate gradient algorithm does not
indicate what search direction to use once a cycle of n iterations has been
completed. There have been many procedures suggested, but the simplest
method is to reset the search direction to the steepest descent direction
(negative of the gradient) after n iterations [Scal85]. We will use this meth-
od.
Lets now apply the conjugate gradient algorithm to the function approxi-
mation example that we have been using to demonstrate the other neural
network training algorithms. We will use the backpropagation algorithm to
compute the gradient (using Eq. (11.23) and Eq. (11.24)) and the conjugate
gradient algorithm to determine the weight updates. This is a batch mode
algorithm, as the gradient is computed after the entire training set has
been presented to the network.
Figure 12.15 shows the intermediate steps of the CGBP algorithm for the
first three iterations. The interval location process is illustrated by the
open blue circles; each one represents one evaluation of the function. The
final interval is indicated by the larger open black circles. The black dots in
Figure 12.15 indicate the location of the new interior points during the
Golden Section search, one for each iteration of the procedure. The final
point is indicated by a blue dot.
Figure 12.16 shows the total trajectory to convergence. Notice that the
CGBP algorithm converges in many fewer iterations than the other algo-
rithms that we have tested. This is a little deceiving, since each iteration of
CGBP requires more computations than the other methods; there are
many function evaluations involved in each iteration of CGBP. Even so,
CGBP has been shown to be one of the fastest batch training algorithms for
multilayer networks [Char92].
12-18
15
12
10
2
w 1 1 5
-5
-5 0 5 10 15
1
w 1 1
Figure 12.15 Intermediate Steps of CGBP

15
1.5
Squared Error
10
2
w 1 1 5
0.5
0
0 0 1 2
10 10 10
Iteration Number
-5
-5 0 5 10 15
1
w 1 1
Figure 12.16 Conjugate Gradient Trajectory

To experiment with CGBP, use the MATLAB Neural Network Design
Demonstrations Conjugate Gradient Line Search (nnd12ls) and Conjugate
Gradient Backpropagation (nnd12cg).
Levenberg-Marquardt Algorithm
The Levenberg-Marquardt algorithm is a variation of Newtons method
that was designed for minimizing functions that are sums of squares of oth-
er nonlinear functions. This is very well suited to neural network training
where the performance index is the mean squared error.
Basic Algorithm
Lets begin by considering the form of Newtons method where the perfor-
mance index is a sum of squares. Recall from Chapter 9 that Newtons
method for optimizing a performance index F x is
12-19
1
xk + 1 = xk Ak gk , (12.19)
where A k 2F x and g k F x .
x = xk x = xk
If we assume that F x is a sum of squares function:

N
vi x
2 T
Fx = = v x v x , (12.20)
i=1
then the jth element of the gradient would be

N
F x v i x
F x j = -------------- = 2 v i x --------------
-. (12.21)
x j x j
i=1
The gradient can therefore be written in matrix form:

T
F x = 2J x v x , (12.22)
where
v 1 x v 1 x v 1 x
- ---------------- ---------------
--------------- -
x 1 x 2 x n
v 2 x v 2 x v 2 x
- ---------------- ---------------
--------------- -
Jx = x 1 x 2 x n

v N x v N x v N x
---------------- ---------------- ----------------
x 1 x 2 x n
. (12.23)
Jacobian Matrix is the Jacobian matrix.
Next we want to find the Hessian matrix. The k j element of the Hessian
matrix would be
N 2
2
Fx v i x v i x vi x
2F x = ----------------- = 2 - --------------- + v i x -----------------
-------------- x k x j
k j . (12.24)
x k x j x k x j
i=1
The Hessian matrix can then be expressed in matrix form:

T
2F x = 2J x J x + 2S x , (12.25)
where
12-20
Sx = vi x 2vi x . (12.26)
i=1
If we assume that S x is small, we can approximate the Hessian matrix as

T
2F x 2J x J x . (12.27)
If we then substitute Eq. (12.27) and Eq. (12.22) into Eq. (12.19), we obtain
Gauss-Newton the Gauss-Newton method:
T 1 T
x k + 1 = x k 2J x k J x k 2J x k v x k
T 1 T
= x k J x k J x k J x k v x k .
(12.28)
Note that the advantage of Gauss-Newton over the standard Newtons
method is that it does not require calculation of second derivatives.
T
One problem with the Gauss-Newton method is that the matrix H = J J
may not be invertible. This can be overcome by using the following modifi-
cation to the approximate Hessian matrix:
G = H + I . (12.29)
To see how this matrix can be made invertible, suppose that the eigenval-
ues and eigenvectors of H are 1 2 n and z 1 z 2 z n . Then
Gz i = H + I z i = Hz i + z i = i z i + z i = i + z i . (12.30)
Therefore the eigenvectors of G are the same as the eigenvectors of H , and

the eigenvalues of G are i + . G can be made positive definite by in-
creasing until i + 0 for all i , and therefore the matrix will be in-
vertible.
Levenberg-Marquardt This leads to the Levenberg-Marquardt algorithm [Scal85]:
T 1 T
x k + 1 = x k J x k J x k + k I J x k v x k . (12.31)
or
T 1 T
x k = J x k J x k + k I J x k v x k . (12.32)
This algorithm has the very useful feature that as k is increased it ap-
proaches the steepest descent algorithm with small learning rate:
12-21
1 T 1
x k + 1 x k ----- J x k v x k = x k -------- F x , for large k , (12.33)
k 2 k
while as k is decreased to zero the algorithm becomes Gauss-Newton.

The algorithm begins with k set to some small value (e.g., k = 0.01 ). If a
step does not yield a smaller value for F x , then the step is repeated with
k multiplied by some factor 1 (e.g., = 10 ). Eventually F x should
decrease, since we would be taking a small step in the direction of steepest
descent. If a step does produce a smaller value for F x , then k is divided
by for the next step, so that the algorithm will approach Gauss-Newton,
which should provide faster convergence. The algorithm provides a nice
compromise between the speed of Newtons method and the guaranteed
convergence of steepest descent.
Now lets see how we can apply the Levenberg-Marquardt algorithm to the
multilayer network training problem. The performance index for multilay-
er network training is the mean squared error (see Eq. (11.11)). If each tar-
get occurs with equal probability, the mean squared error is proportional
to the sum of squared errors over the Q targets in the training set:
Q
tq aq
T
Fx = tq aq
q=1
M
Q Q S N
e j q vi
T 2 2
= eq eq = = ,
q=1 q = 1j = 1 i=1 (12.34)
where e j q is the j th element of the error for the q th input/target pair.
Eq. (12.34) is equivalent to the performance index, Eq. (12.20), for which
Levenberg-Marquardt was designed. Therefore it should be a straightfor-
ward matter to adapt the algorithm for network training. It turns out that
this is true in concept, but it does require some care in working out the de-
tails.
Jacobian Calculation
The key step in the Levenberg-Marquardt algorithm is the computation of
the Jacobian matrix. To perform this computation we will use a variation
of the backpropagation algorithm. Recall that in the standard backpropa-
gation procedure we compute the derivatives of the squared errors, with re-
spect to the weights and biases of the network. To create the Jacobian
matrix we need to compute the derivatives of the errors, instead of the de-
rivatives of the squared errors.
It is a simple matter conceptually to modify the backpropagation algorithm
to compute the elements of the Jacobian matrix. Unfortunately, although
12-22
the basic concept is simple, the details of the implementation can be a little
tricky. For that reason you may want to skim through the rest of this sec-
tion on your first reading, in order to obtain an overview of the general flow
of the presentation, and return later to pick up the details. It may also be
helpful to review the development of the backpropagation algorithm in
Chapter 11 before proceeding.
Before we present the procedure for computing the Jacobian, lets take a
closer look at its form (Eq. (12.23)). Note that the error vector is
T
v = v 1 v 2 v N = e 1 1 e 2 1 e M e 1 2 e M , (12.35)
S 1 S Q
the parameter vector is
T
x = x 1 x 2 x n = w 11 1 w 11 2 w 11 1
b1 b
1
1
2
w 1 1 b
M
M
, (12.36)
S R S S
and n = S R + 1 + S S + 1 + + S S
M 1 2 1 M M1
N = QS + 1 .
Therefore, if we make these substitutions into Eq. (12.23), the Jacobian
matrix for multilayer network training can be written
e 1 1 e 1 1 e 1 1 e 1 1
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w S 1 R b 1
e 2 1 e 2 1 e 2 1 e 2 1
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w S 1 R b 1

Jx = . (12.37)
e S M 1 e S M 1 e S M 1 e SM 1
--------------
1 1
- --------------
------------- 1
-
- -------------
1
w 1 1 w 1 2 w S 1 R b 1
e 1 2 e 1 2 e 1 2 e 1 2
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w S 1 R b 1

The terms in this Jacobian matrix can be computed by a simple modifica-

tion to the backpropagation algorithm.
Standard backpropagation calculates terms like
T
F x e q e q
-------------- = --------------- . (12.38)
x l x l
12-23
For the elements of the Jacobian matrix that are needed for the Levenberg-
Marquardt algorithm we need to calculate terms like
v e k q
J h l = -------h = ----------- . (12.39)
x l x l
Recall from Eq. (11.18) in our derivation of backpropagation that

m
F F n i
- = --------m- ----------
----------
m m
-, (12.40)
w i j n i w i j
where the first term on the right-hand side was defined as the sensitivity:
m F
s i --------m- . (12.41)
n i
The backpropagation process computed the sensitivities through a recur-

rence relationship from the last layer backward to the first layer. We can
use the same concept to compute the terms needed for the Jacobian matrix
Marquardt Sensitivity (Eq. (12.37)) if we define a new Marquardt sensitivity:
m v h e k q
s i h ----------
m
- = -----------
m
, (12.42)
n i q n i q
M
where, from Eq. (12.35), h = q 1 S + k .
Now we can compute elements of the Jacobian by
m m
v e k q e k q n i q m n i q m m1
J h l = -------h = ----------- = ----------- ----------
- = s i h ----------
- = s i h a j q , (12.43)
x l w i j
m m
n i q w i j
m m
w i j
or if x l is a bias,
m m
v e k q e k q n i q m n i q m
J h l = -------h = ----------- = ----------- ----------
- = s i h ----------
- = s i h . (12.44)
x l b i
m m
n i q b i
m
b i
m
The Marquardt sensitivities can be computed through the same recurrence

relations as the standard sensitivities (Eq. (11.35)) with one modification
at the final layer, which for standard backpropagation is computed with
Eq. (11.40). For the Marquardt sensitivities at the final layer we have
12-24
M M
M v h e k q t k q a k q a k q
s i h = ----------
M
- = -----------
M
= ------------------------------
M
- = -----------
M
-
n i q n i q n i q n i q
M M
f n i q for i = k
= .
0 for i k

(12.45)
Therefore when the input p q has been applied to the network and the cor-
M
responding network output a q has been computed, the Levenberg-Mar-
quardt backpropagation is initialized with
S q = F n q ,
M M M
(12.46)
where F n is defined in Eq. (11.34). Each column of the matrix S q

M M M
must be backpropagated through the network using Eq. (11.35) to produce

one row of the Jacobian matrix. The columns can also be backpropagated
together using
m+1 T m+1
S q = F (n q ) W
m m m
S q . (12.47)
The total Marquardt sensitivity matrices for each layer are then created by
augmenting the matrices computed for each input:
m m m m
S = S 1 S 2 S Q . (12.48)
Note that for each input that is presented to the network we will backprop-
M
agate S sensitivity vectors. This is because we are computing the deriva-
tives of each individual error, rather than the derivative of the sum of
squares of the errors. For every input applied to the network there will be
M
S errors (one for each element of the network output). For each error
there will be one row of the Jacobian matrix.
After the sensitivities have been backpropagated, the Jacobian matrix is

computed using Eq. (12.43) and Eq. (12.44). See Problem P12.5 for a nu-
merical illustration of the Jacobian computation.
The iterations of the Levenberg-Marquardt backpropagation algorithm
LMBP (LMBP) can be summarized as follows:
1. Present all inputs to the network and compute the corresponding net-
work outputs (using Eq. (11.41) and Eq. (11.42)) and the errors
M
e q = t q a q . Compute the sum of squared errors over all inputs, F x ,
12-25
using Eq. (12.34).
2. Compute the Jacobian matrix, Eq. (12.37). Calculate the sensitivities

with the recurrence relations Eq. (12.47), after initializing with Eq.
(12.46). Augment the individual matrices into the Marquardt sensitiv-
ities using Eq. (12.48). Compute the elements of the Jacobian matrix
with Eq. (12.43) and Eq. (12.44).
3. Solve Eq. (12.32) to obtain x k .
4. Recompute the sum of squared errors using x k + x k . If this new sum
of squares is smaller than that computed in step 1, then divide by ,
let x k + 1 = x k + x k and go back to step 1. If the sum of squares is not
reduced, then multiply by and go back to step 3.
The algorithm is assumed to have converged when the norm of the gradi-
ent, Eq. (12.22), is less than some predetermined value, or when the sum
of squares has been reduced to some error goal.
To illustrate LMBP, lets apply it to the function approximation problem in-
troduced at the beginning of this chapter. We will begin by looking at the
basic Levenberg-Marquardt step. Figure 12.17 illustrates the possible
steps the LMBP algorithm could take on the first iteration.
15
10
2
w 1 1 5
-5
-5 0 5 10 15
1
w 1 1
Figure 12.17 Levenberg-Marquardt Step

The black arrow represents the direction taken for small k , which corre-
sponds to the Gauss-Newton direction. The blue arrow represents the di-
rection taken for large k , which corresponds to the steepest descent
direction. (This was the initial direction taken by all of the previous algo-
rithms discussed.) The blue curve represents the Levenberg-Marquardt
step for all intermediate values of k . Note that as k is increased the al-
gorithm moves toward a small step in the direction of steepest descent.
This guarantees that the algorithm will always be able to reduce the sum
of squares at each iteration.
12-26
Figure 12.18 shows the path of the LMBP trajectory to convergence, with
0 = 0.01 and = 5 . Note that the algorithm converges in fewer itera-
tions than any of the methods we have discussed so far. Of course this al-
gorithm also requires more computation per iteration than any of the other
algorithms, since it involves a matrix inversion. Even given the large num-
ber of computations, however, the LMBP algorithm appears to be the fast-
est neural network training algorithm for moderate numbers of network
parameters [HaMe94].
15
1.5
Squared Error
10
2
w 1 1 5
0.5
0
0 0 1 2
10 10 10
Iteration Number
-5
-5 0 5 10 15
1
w 1 1
Figure 12.18 LMBP Trajectory

To experiment with the LMBP algorithm, use the MATLAB Neural Net-
work Design Demonstrations Marquardt Step (nnd12ms) and Marquardt
Backpropagation (nnd12m).
The key drawback of the LMBP algorithm is the storage requirement. The
T
algorithm must store the approximate Hessian matrix J J . This is an n n
matrix, where n is the number of parameters (weights and biases) in the
network. Recall that the other methods discussed need only store the gra-
dient, which is an n-dimensional vector. When the number of parameters
is very large, it may be impractical to use the Levenberg-Marquardt algo-
rithm. (What constitutes very large depends on the available memory on
your computer, but typically a few thousand parameters is an upper limit.)
12-27
Summary of Results
Heuristic Variations of Backpropagation

Batching
The parameters are updated only after the entire training set has been pre-
sented. The gradients calculated for each training example are averaged
together to produce a more accurate estimate of the gradient. (If the train-
ing set is complete, i.e., covers all possible input/output pairs, then the gra-
dient estimate will be exact.)
Backpropagation with Momentum (MOBP)

m m m m1 T
W k = W k 1 1 s a
m m m
b k = b k 1 1 s
Variable Learning Rate Backpropagation (VLBP)

1. If the squared error (over the entire training set) increases by more
than some set percentage (typically one to five percent) after a
weight update, then the weight update is discarded, the learning rate
is multiplied by some factor 1 , and the momentum coefficient (if
it is used) is set to zero.
2. If the squared error decreases after a weight update, then the weight
update is accepted and the learning rate is multiplied by some factor
1 . If has been previously set to zero, it is reset to its original val-
ue.
3. If the squared error increases by less than , then the weight update
is accepted but the learning rate and the momentum coefficient are un-
changed.
12-28
Summary of Results

12
Conjugate Gradient
Interval Location
F(x)
8
4
2
x
a1 b1
a2 b2
a3 b3
a4 b4
a5 b5
Interval Reduction (Golden Section Search)

= 0.618
Set c1 = a1 + 1 b1 a1 , Fc = F c1 .
d1 = b1 1 b1 a1 , Fd = F d1 .
For k = 1 2 repeat
If F c F d then
Set ak + 1 = ak ; bk + 1 = dk ; dk + 1 = ck
ck + 1 = ak + 1 + 1 bk + 1 ak + 1
Fd = Fc ; Fc = F ck + 1
else
Set ak + 1 = ck ; bk + 1 = bk ; ck + 1 = dk
dk + 1 = bk + 1 1 bk + 1 ak + 1
Fc = Fd ; Fd = F dk + 1
end
end until b k + 1 a k + 1 tol
12-29
Levenberg-Marquardt Backpropagation (LMBP)

T 1 T
x k = J x k J x k + k I J x k v x k
T
v = v 1 v 2 v N = e 1 1 e 2 1 e M e 1 2 e M
S 1 S Q
T
x = x 1 x 2 x n = w 11 1 w 11 2 w 11 1
b1 b
1
1
2
w 1 1 b
M
M
S R S S
M 1 2 1 M M1
N = QS and n = S R + 1 + S S + 1 + + S S + 1
e 1 1 e 1 1 e 1 1 e 1 1
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w 1 b 1
S R
e 2 1 e 2 1 e 2 1 e 2 1
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w S 1 R b 1

Jx =
e S M 1 e SM 1 e S M 1 e SM 1
--------------
1 1
- --------------
------------- 1
-
- -------------
1
w 1 1 w 1 2 w S 1 R b 1
e 1 2 e 1 2 e 1 2 e 1 2
------------
1 1
- --------------
- ------------ 1
-
- -----------
1
w 1 1 w 1 2 w S 1 R b 1

m m
v e k q e k q n i q m n i q m m1
J h l = -------h = ----------- = ----------- ----------
- = s i h ----------
- = s i h a j q for weight x l
x l w i j
m m
n i q w i j
m
w i j
m
m m
v e k q e k q n i q m n i q m
J h l = -------h = ----------- = ----------- ----------
- = s i h ----------
- = s i h for bias x l
x l b i
m m
n i q b i
m
b i
m
m v h e k q M
s i h ----------
m
- = -----------
m
(Marquardt Sensitivity) where h = q 1 S + k
n i q n i q
S q = F n q
M MM
S q = F (n q ) W
m m m m+1 T m+1
S q
m m m m
S = S 1 S 2 S Q
12-30
Summary of Results
Levenberg-Marquardt Iterations
1. Present all inputs to the network and compute the corresponding net-
work outputs (using Eq. (11.41) and Eq. (11.42)) and the errors
M
e q = t q a q . Compute the sum of squared errors over all inputs, F x ,
using Eq. (12.34).
2. Compute the Jacobian matrix, Eq. (12.37). Calculate the sensitivities

with the recurrence relations Eq. (12.47), after initializing with Eq.
(12.46). Augment the individual matrices into the Marquardt sensitiv-
ities using Eq. (12.48). Compute the elements of the Jacobian matrix
with Eq. (12.43) and Eq. (12.44).
3. Solve Eq. (12.32) to obtain x k .
4. Recompute the sum of squared errors using x k + x k . If this new sum
of squares is smaller than that computed in step 1, then divide by ,
let x k + 1 = x k + x k and go back to step 1. If the sum of squares is not
reduced, then multiply by and go back to step 3.
12-31
Solved Problems
P12.1 We want to train the network shown in Figure P12.1 on the train-
ing set

p 1 = 3 t 1 = 0.5 , p 2 = 2 t 2 = 1 ,

w 0 = 0.4 , b 0 = 0.15 .
Demonstrate the effect of batching by computing the direction of

the initial step for SDBP with and without batching.
Input Log-Sigmoid Layer
w n a
p
b
1
a = logsig (w p + b)
Figure P12.1 Network for Problem P12.1
Lets begin by computing the direction of the initial step if batching is not
used. In this case the first step is computed from the first input/target pair.
The forward and backpropagation steps are
1
a = logsig wp + b = ------------------------------------------------------------------- = 0.2592
1 + exp 0.4 3 + 0.15
e = t a = 0.5 0.2592 = 0.2408

s = 2f n e = 2a 1 a e = 2 0.2592 1 0.2592 0.2408 = 0.0925 .
The direction of the initial step is the negative of the gradient. For the
weight this will be
sp = 0.0925 3 = 0.2774 .
For the bias we have
s = 0.0925 = 0.0925 .
12-32
Solved Problems
Therefore the direction of the initial step in the w b plane would be
0.2774 .
0.0925
Now lets consider the initial direction for the batch mode algorithm. In this
case the gradient is found by adding together the individual gradients
found from the two sets of input/target pairs. For this we need to apply the
second input to the network and perform the forward and backpropagation
steps:
1
a = logsig wp + b = ---------------------------------------------------------------- = 0.7211
1 + exp 0.4 2 + 0.15
e = t a = 1 0.7211 = 0.2789

s = 2f n e = 2a 1 a e = 2 0.7211 1 0.7211 0.2789 = 0.1122 .
The direction of the step is the negative of the gradient. For the weight this
will be
sp = 0.1122 2 = 0.2243 .
For the bias we have
s = 0.1122 = 0.1122 .
The partial gradient for the second input/target pair is therefore
0.2243 .
0.1122
If we now add the results from the two input/target pairs we find the direc-
tion of the first step of the batch mode SDBP to be
1 0.2774 1
--- + 0.2243 = --- 0.0531 = 0.0265 .
2 0.0925 0.1122 2 0.2047 0.1023
The results are illustrated in Figure P12.2. The blue circle indicates the ini-
tial guess. The two blue arrows represent the directions of the partial gra-
dients for each of the two input/target pairs, and the black arrow
represents the direction of the total gradient. The function that is plotted
is the sum of squared errors for the entire training set. Note that the indi-
vidual partial gradients can point in quite different directions than the true
gradient. However, on the average, over several iterations, the path will
generally follow the steepest descent trajectory.
12-33
The relative effectiveness of the batch mode over the incremental approach
depends very much on the particular problem. The incremental approach
requires less storage, and, if the inputs are presented randomly to the net-
work, the trajectory is stochastic, which makes the algorithm somewhat
less likely to be trapped in a local minimum. It may also take longer to con-
verge than the batch mode algorithm.
4
1.5
0.5
b 0
0
-4 -2
-2 -4
0 -2
0
b 2
2 w
4 4 -4
-4 -2 0 2 4
Figure P12.2 Effect of Batching in Problem P12.1
P12.2 In Chapter 9 we proved that the steepest descent algorithm, when

applied to a quadratic function, would be stable if the learning
rate was less than 2 divided by the maximum eigenvalue of the
Hessian matrix. Show that if a momentum term is added to the
steepest descent algorithm there will always be a momentum coef-
ficient that will make the algorithm stable, regardless of the learn-
ing rate. Follow the format of the proof on page 9-6.
The standard steepest descent algorithm is
x k = F x k = g k ,
If we add momentum this becomes
x k = x k 1 1 g k .
Recall from Chapter 8 that the quadratic function has the form
1 T T
F x = --- x Ax + d x + c ,
2
and the gradient of the quadratic function is
F x = Ax + d .
12-34
Solved Problems
If we now insert this expression into our expression for the steepest descent
12
algorithm with momentum we obtain
x k = x k 1 1 Ax k + d .
Using the definition x k = x k + 1 x k this can be rewritten
x k + 1 x k = x k x k 1 1 Ax k + d
or
x k + 1 = 1 + I 1 A x k x k 1 1 d .
Now define a new vector
xk 1
x k = .
xk
The momentum variation of steepest descent can then be written
x k + 1 = 0 I x k + 0 = Wx k + v .
I 1 + I 1 A 1 d
This is a linear dynamic system that will be stable if the eigenvalues of W

are less than one in magnitude. We will find the eigenvalues of W in stag-
es. First, rewrite W as
W = 0 I where T = 1 + I 1 A .
I T
The eigenvalues and eigenvectors of W should satisfy
w w
Wz
w w w
= z , or 0 I z1 = w z1 .
I T z w w
z2
2
This means that

w w w w w w w
z 2 = z 1 and z 1 + Tz 2 = z 2 .
w
At this point we will choose z 2 to be an eigenvector of the matrix T , with
t
corresponding eigenvalue . (If this choice is not appropriate it will lead
to a contradiction.) Therefore the previous equations become
12-35
w w w w t w w w
z 2 = z 1 and z 1 + z 2 = z 2 .
If we substitute the first equation into the second equation we find
w t w w w w 2 t w w
-----w- z 2 + z 2 = z 2 or + z 2 = 0 .

t w
Therefore for each eigenvalue of T there will be two eigenvalues of
W that are roots of the quadratic equation
w 2 t w
+ = 0 .
From the quadratic formula we have
t t 2
w 4
= -------------------------------------- .
2
For the algorithm to be stable the magnitude of each eigenvalue must be

less than 1. We will show that there always exists some range of for
which this is true.
w
Note that if the eigenvalues are complex then their magnitude will be
:
t 2 t 2
w 4
= ----------- + ----------------------- = .
4 4
t t
(This is true only for real . We will show later that is real.) Since is
between 0 and 1, the magnitude of the eigenvalue must be less than 1. It
remains to show that there exists some range of for which all of the eigen-
values are complex.
w
In order for to be complex we must have
t 2 t
4 0 or 2 .
t
Lets now consider the eigenvalues of T . These eigenvalues can be ex-
pressed in terms of the eigenvalues of A . Let 1 2 n and
z 1 z 2 z n be the eigenvalues and eigenvectors of the Hessian matrix.
Then
Tz i = 1 + I 1 A z i = 1 + z i 1 Az i
t
= 1 + z i 1 i z i = 1 + 1 i z i = i z i
.
12-36
Solved Problems
Therefore the eigenvectors of T are the same as the eigenvectors of A , and

the eigenvalues of T are
t
i = 1 + 1 i .
t
(Note that i is real, since , and i for symmetric A are real.) There-
w
fore, in order for to be complex we must have
t
2 or 1 + 1 i 2 .
For = 1 both sides of the inequality will equal 2. The function on the
right of the inequality, as a function of , has a slope of 1 at = 1 . The
function on the left of the inequality has a slope of 1 + i . Since the eigen-
values of the Hessian will be positive real numbers if the function has a
strong minimum, and the learning rate is a positive number, this slope
must be greater than 1. This shows that the inequality will always hold for
close enough to 1.
To summarize the results, we have shown that if a momentum term is add-
ed to the steepest descent algorithm on a quadratic function, then there
will always be a momentum coefficient that will make the algorithm stable,
regardless of the learning rate. In addition we have shown that if is close
enough to 1, then the magnitudes of the eigenvalues of W will be . It can
be shown (see [Brog91]) that the magnitudes of the eigenvalues determine
how fast the algorithm will converge. The smaller the magnitude, the fast-
er the convergence. As the magnitude approaches 1, the convergence time
increases.
We can demonstrate these results using the example on page 9-7. There we
showed that the steepest descent algorithm, when applied to the function
2 2
F x = x 1 + 25x 2 , was unstable for a learning rate 0.4 . In Figure P12.3
we see the steepest descent trajectory (with momentum) with = 0.041
and = 0.2 . Compare this trajectory with Figure 9.3, which uses the same
learning rate but no momentum.
12-37
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure P12.3 Trajectory for = 0.041 and = 0.2
P12.3 Execute three iterations of the variable learning rate algorithm on

the following function (from the Chapter 9 example on page 9-7):
2 2
F x = x 1 + 25x 2 ,
x 0 = 0.5 ,
0.5
and use the following values for the algorithm parameters:
= 0.05 , = 0.2 , = 1.5 , = 0.5 , = 5% .
The first step is to evaluate the function at the initial guess:
1 T 1
F x 0 = --- x 0 2 0 x 0 = --- 0.5 0.5 2 0 0.5 = 6.5 .
2 0 50 2 0 50 0.5
The next step is to find the gradient:

Fx
x1 2x 1
F x = = .
50x 2
Fx
x2
If we evaluate the gradient at the initial guess we find:
12-38
Solved Problems
12
g 0 = F x = 1 .
x = x0 25
With the initial learning rate of = 0.05 , the tentative first step of the al-
gorithm is
x 0 = x 1 1 g 0 = 0.2 0 0.8 0.05 1 = 0.04

0 25 1
x 1 = x 0 + x 0 = 0.5 + 0.04 = 0.46 .

t
0.5 1 0.5
To verify that this is a valid step we must test the value of the function at
this new point:
1 t T 1
F x 1 = --- x 1 2 0 x 1 = --- 0.46 0.5 2 0 0.46 = 6.4616 .
t t
2 0 50 2 0 50 0.5
This is less than F x 0 . Therefore this tentative step is accepted and the
learning rate is increased:
x 1 = x 1 = 0.46 , F x 1 = 6.4616 and = = 1.5 0.05 = 0.075 .

t
0.5
The tentative second step of the algorithm is
x 1 = x 0 1 g 1 = 0.2 0.04 0.8 0.075 0.92 = 0.0632

1 25 1.3
x 2 = x 1 + x 1 = 0.46 + 0.0632 = 0.3968 .

t
0.5 1.3 0.8
We evaluate the function at this point:
1 t T 1
F x 2 = --- x 2 2 0 x 2 = --- 0.3968 0.8 2 0 0.3968 = 16.157 .
t t
2 0 50 2 0 50 0.8
Since this is more than 5% larger than F x 1 , we reject this step, reduce
the learning rate and set the momentum coefficient to zero.
x 2 = x 1 , F x 2 = F x 1 = 6.4616 , = = 0.5 0.075 = 0.0375 , = 0
12-39
Now a new tentative step is computed (momentum is zero).
x 2 = g 2 = 0.0375 0.92 = 0.0345

25 0.9375
x 3 = x 2 + x 2 = 0.46 + 0.0345 = 0.4255

t
0.5 0.9375 0.4375
1 t T 1
F x 3 = --- x 3 2 0 x 3 = --- 0.4255 0.4375 2 0 0.4255 = 4.966
t t
2 0 50 2 0 50 0.4375
This is less than F x 2 . Therefore this step is accepted, the momentum is

reset to its original value, and the learning rate is increased.
t
x 3 = x 3 , = 0.2 , = = 1.5 0.0375 = 0.05625
This completes the third iteration.
P12.4 Recall the example from Chapter 9 that we used to demonstrate

the conjugate gradient algorithm (page 9-18):
1 T
F x = --- x 2 1 x ,
2 1 2
with initial guess
x0 = 0.8 .
0.25
Perform one iteration of the conjugate gradient algorithm. For the

linear minimization use interval location by function evaluation
and interval reduction by the Golden Section search.
2x 1 + x 2
F x = .
x 1 + 2x 2
As with steepest descent, the first search direction for the conjugate gradi-
ent algorithm is the negative of the gradient:
12-40
Solved Problems
= 1.35 .
T
p 0 = g 0 = F x
x = x0 0.3
For the first iteration we need to minimize F x along the line
x1 = x0 + 0 p0 = 0.8 + 1.35 .
0
0.25 0.3
The first step is interval location. Assume that the initial step size is
= 0.075 . Then the interval location would proceed as follows:

F a 1 = F 0.8 = 0.5025 ,
0.25

b 1 = = 0.075 , F b 1 = F 0.8 + 0.075 1.35 = 0.3721
0.25 0.3

b 2 = 2 = 0.15 , F b 2 = F 0.8 + 0.15 1.35 = 0.2678
0.25 0.3

b 3 = 4 = 0.3 , F b 3 = F 0.8 + 0.3 1.35 = 0.1373
0.25 0.3

b 4 = 8 = 0.6 , F b 4 = F 0.8 + 0.6 1.35 = 0.1893 .
0.25 0.3
Since the function increases between two consecutive evaluations we know

that the minimum must occur in the interval [0.15, 0.6] . This process is il-
lustrated by the open blue circles in Figure P12.4, and the final interval is
indicated by the large open black circles.
The next step in the linear minimization is interval reduction using the
Golden Section search. This proceeds as follows:
c 1 = a 1 + 1 b 1 a 1 = 0.15 + 0.382 0.6 0.15 = 0.3219 ,
d 1 = b 1 1 b 1 a 1 = 0.6 0.382 0.6 0.15 = 0.4281 ,
F a = 0.2678 , F b = 0.1893 , F c = 0.1270 , F d = 0.1085 .
12-41
Since F c F d , we have
a 2 = c 1 = 0.3219 , b 2 = b 1 = 0.6 , c 2 = d 1 = 0.4281
d 2 = b 2 1 b 2 a 2 = 0.6 0.382 0.6 0.3219 = 0.4938 ,
F a = F c = 0.1270 , F c = F d = 0.1085 , F d = F d 2 = 0.1232 .
This time F c F d , therefore
a 3 = a 2 = 0.3219 , b 3 = d 2 = 0.4938 , d 3 = c 2 = 0.4281 ,
c 3 = a 3 + 1 b 3 a 3 = 0.3219 + 0.382 0.4938 0.3219 = 0.3876 ,
F b = F d = 0.1232 , F d = F c = 0.1085 , F c = F c 3 = 0.1094 .
This routine continues until b k + 1 a k + 1 tol . The black dots in Figure

P12.4 indicate the location of the new interior points, one for each iteration
of the procedure. The final point is indicated by a blue dot. Compare this
result with the first iteration shown in Figure 9.10.
1
0.5
-0.5
-1
-1 -0.5 0 0.5 1
Figure P12.4 Linear Minimization Example
P12.5 To illustrate the computation of the Jacobian matrix for the Lev-
enberg-Marquardt method, consider using the network of Figure
P12.5 for function approximation. The network transfer functions
are chosen to be
1 2 2
f n = n , f n = n .
Therefore their derivatives are
f n = 2n , f n = 1 .
1 2
12-42
Solved Problems
Assume that the training set consists of

12
{ p 1 = 1 , t 1 = 1 } , { p 2 = 2 , t 2 = 2 } ,
and that the parameters are initialized to

1 1 2 1
W = 1 ,b = 0 ,W = 2 ,b = 1 .
Find the Jacobian matrix for the first step of the Levenberg-Mar-
quardt method.
Input Layer 1 Layer 2
w11,1 n11 a11 w21,1 n21 a21

p f1 f2
b11 b21
1 1
a1 = f 1 (w1 p + b1) a2 = f 2 (w2 a1 + b2)
Figure P12.5 Two-Layer Network for LMBP Demonstration

The first step is to propagate the inputs through the network and compute
the errors.
0
a1 = p1 = 1
1 1 0 1 1 1 1 2
n1 = W a1 + b = 1 1 + 0 = 1 , a1 = f n1 = 1 = 1
2 2 1 2 2 2 2
n1 = W a1 + b = 2 1 + 1 = 3 , a1 = f n1 = 3 = 3
2
e1 = t1 a1 = 1 3 = 2
0
a2 = p2 = 2
1 1 0 1 1 1 1 2
n2 = W a2 + b = 1 2 + 0 = 2 , a2 = f n2 = 2 = 4
2 2 1 2 2 2 2
n2 = W a2 + b = 2 4 + 1 = 9 , a2 = f n2 = 9 = 9
12-43
2
e2 = t2 a2 = 2 9 = 7
The next step is to initialize and backpropagate the Marquardt sensitivi-

ties using Eq. (12.46) and Eq. (12.47).
S 1 = F n 1 = 1
2 2 2
S 1 = F (n 1) W S 1 = 2n 1
1 1 1 2 T 2
= 2 1 2 1 = 4
1 1 2 1
S 2 = F n 2 = 1
2 2 2
2 T 2
S 2 = F (n 2) W S 2 = 2n 2
1 1 1
= 2 2 2 1 = 8
1 2 2 1
1 2
S = S 11 S 12 = 4 8 , S = S 21 S 22 = 1 1
We can now compute the Jacobian matrix using Eq. (12.43), Eq. (12.44) and
Eq. (12.37).
v 1 v 1 v 1 v 1 e 1 1 e 1 1 e 1 1 e 1 1
------- ------- ------- ------- ------------
1
- -----------
1
- ------------
2
- -----------
2
-
x 1 x 2 x 3 x 4 w 1 1 b 1 w 1 1 b 1
Jx = =
v 2 v 2 v 2 v 2 e 1 2 e 1 2 e 1 2 e 1 2
------- ------- ------- ------- ------------
- -----------
- ------------
- -----------
-
x 1 x 2 x 3 x 4 1 1
w 1 1 b 1 w 1 1 b 1
2 2
1 1
v e 1 1 e 1 1 n 1 1 1 n 1 1 1 0
J 1 1 = -------1 = ------------ - ------------
- = ----------- - = s 1 1 ------------
- = s 1 1 a 1 1
x 1 1
w 1 1
1
n 1 1 w 1 1
1
w 1 1
1
= 4 1 = 4
1 1
v e 1 1 e 1 1 n 1 1 1 n 1 1 1
J 1 2 = -------1 = ----------- - -----------
- = ----------- - = s 1 1 -----------
- = s 1 1 = 4
x 2 b 1
1 1
n 1 1 b 1
1
b 1
1
2 2
v e 1 1 n 1 1 2 n 1 1 2 1
J 1 3 = -------1 = -----------
- ------------
- = s 1 1 ------------
- = s 1 1 a 1 1 = 1 1 = 1
x 3 2
n 1 1 w 1 1
2
w 1 1
2
12-44
Solved Problems
1 2
12
v e 1 1 n 1 1 2 n 1 1 2
J 1 4 - -----------
= -------1 = ----------- - = s 1 1 -----------
- = s 1 1 = 1
x 4 2
n 1 1 b 1
2
b 1
2
1 1
v e 1 2 n 1 2 1 n 1 2 1 0
J 2 1 = -------2 = -----------
- ------------
- = s 1 2 ------------
- = s 1 2 a 1 2 = 8 2 = 16
x 1 1
n 1 2 w 1 1
1
w 1 1
1
1 1
v e 1 2 e 1 2 n 1 2 1 n 1 2 1
J 2 2 = -------2 = ----------- - -----------
- = ----------- - = s 1 2 -----------
- = s 1 2 = 8
x 2 b 1
1 1
n 1 2 b 1
1
b 1
1
2 2
v e 1 2 n 1 2 2 n 1 2 2 1
J 2 3 - ------------
= -------2 = ----------- - = s 1 2 ------------
- = s 1 2 a 1 2 = 1 4 = 4
x 3 2
n 1 2 w 1 1
2
w 1 1
2
2 2
v e 1 2 e 1 2 n 1 2 2 n 1 2 2
J 2 4 = -------2 = ----------- - -----------
- = ----------- - = s 1 2 -----------
- = s 1 2 = 1
x 4 b 1
2 2
n 1 2 b 1
2
b 1
2
Therefore the Jacobian matrix is
Jx = 4 4 1 1 .
16 8 4 1
12-45
Epilogue
One of the major problems with the basic backpropagation algorithm
(steepest descent backpropagation SDBP) has been the long training
times. It is not feasible to use SDBP on practical problems, because it can
take weeks to train a network, even on a large computer. Since backpropa-
gation was first popularized, there has been considerable work on methods
to accelerate the convergence of the algorithm. In this chapter we have dis-
cussed the reasons for the slow convergence of SDBP and have presented
several techniques for improving the performance of the algorithm.
The techniques for speeding up convergence have fallen into two main cat-
egories: heuristic methods and standard numerical optimization methods.
We have discussed two heuristic methods: momentum (MOBP) and vari-
able learning rate (VLBP). MOBP is simple to implement, can be used in
batch mode or incremental mode and is significantly faster than SDBP. It
does require the selection of the momentum coefficient, but is limited to
the range [0, 1] and the algorithm is not extremely sensitive to this choice.
The VLBP algorithm is faster than MOBP but must be used in batch mode.
For this reason it requires more storage. VLBP also requires the selection
of a total of five parameters. The algorithm is reasonably robust, but the
choice of the parameters can affect the convergence speed and is problem
dependent.
We also presented two standard numerical optimization techniques: conju-
gate gradient (CGBP) and Levenberg-Marquardt (LMBP). CGBP is gener-
ally faster than VLBP. It is a batch mode algorithm, which requires a linear
search at each iteration, but its storage requirements are not significantly
different than VLBP. There are many variations of the conjugate gradient
algorithm proposed for neural network applications. We have presented
only one.
The LMBP algorithm is the fastest algorithm that we have tested for train-
ing multilayer networks of moderate size, even though it requires a matrix
inversion at each iteration. It requires that two parameters be selected, but
the algorithm does not appear to be sensitive to this selection. The main
T
drawback of LMBP is the storage requirement. The J J matrix, which
must be inverted, is n n , where n is the total number of weights and bi-
ases in the network. If the network has more than a few thousand param-
eters, the LMBP algorithm becomes impractical on current machines.
There are many other variations on backpropagation that have not been
discussed in this chapter. Some references to other techniques are given in
Chapter 19.
12-46
Further Reading
12
Further Reading
[Barn92] E. Barnard, Optimization for training neural nets, IEEE
Trans. on Neural Networks, vol. 3, no. 2, pp. 232240, 1992.
A number of optimization algorithms that have promise for
neural network training are discussed in this paper.
[Batt92] R. Battiti, First- and second-order methods for learning:
Between steepest descent and Newton's method, Neural
Computation, vol. 4, no. 2, pp. 141166, 1992.
This paper is an excellent survey of the current optimiza-
tion algorithms that are suitable for neural network train-
ing.
[Char92] C. Charalambous, Conjugate gradient algorithm for effi-
cient training of artificial neural networks, IEE Proceed-
ings, vol. 139, no. 3, pp. 301310, 1992.
This paper explains how the conjugate gradient algorithm
can be used to train multilayer networks. Comparisons are
made to other training algorithms.
[Fahl88] S. E. Fahlman, Faster-learning variations on back-propa-
gation: An empirical study, In D. Touretsky, G. Hinton &
T. Sejnowski, eds., Proceedings of the 1988 Connectionist
Models Summer School, San Mateo, CA: Morgan Kauf-
mann, pp. 3851, 1988.
The QuickProp algorithm, which is described in this paper,
is one of the more popular heuristic modifications to back-
propagation. It assumes that the error curve can be approx-
imated by a parabola, and that the effect of each weight can
be considered independently. QuickProp provides signifi-
cant speedup over standard backpropagation on many
problems.
[HaMe94] M. T. Hagan and M. Menhaj, Training feedforward net-
works with the Marquardt algorithm, IEEE Transactions
on Neural Networks, vol. 5, no. 6, 1994.
This paper describes the use of the Levenberg-Marquardt
algorithm for training multilayer networks and compares
the performance of the algorithm with variable learning
rate backpropagation and conjugate gradient. The Leven-
berg-Marquardt algorithm is faster, but requires more
storage.
12-47
[Jaco88] R. A. Jacobs, Increased rates of convergence through

learning rate adaptation, Neural Networks, vol. 1, no. 4,
pp. 295308, 1988.
This is another early paper discussing the use of variable
learning rate backpropagation. The procedure described
here is called the delta-bar-delta learning rule, in which
each network parameter has its own learning rate that var-
ies at each iteration.
[NgWi90] D. Nguyen and B. Widrow, Improving the learning speed
of 2-layer neural networks by choosing initial values of the
adaptive weights, Proceedings of the IJCNN, vol. 3, pp.
2126, July 1990.
This paper describes a procedure for setting the initial
weights and biases for the backpropagation algorithm. It
uses the shape of the sigmoid transfer function and the
range of the input variables to determine how large the
weights should be, and then uses the biases to center the
sigmoids in the operating region. The convergence of back-
propagation is improved significantly by this procedure.
[RiIr90] A. K. Rigler, J. M. Irvine and T. P. Vogl, Rescaling of vari-
ables in back propagation learning, Neural Networks, vol.
4, no. 2, pp. 225230, 1991.
This paper notes that the derivative of a sigmoid function
is very small on the tails. This means that the elements of
the gradient associated with the first few layers will gener-
ally be smaller that those associated with the last layer.
The terms in the gradient are then scaled to equalize them.
[Scal85] L. E. Scales, Introduction to Non-Linear Optimization. New
Scales has written a very readable text describing the ma-
jor optimization algorithms. The book emphasizes methods
of optimization rather than existence theorems and proofs
of convergence. Algorithms are presented with intuitive ex-
planations, along with illustrative figures and examples.
Pseudocode is presented for most algorithms.
12-48
Further Reading
[Shan90] D. F. Shanno, Recent advances in numerical techniques

12
for large-scale optimization, Neural Networks for Control,
Miller, Sutton and Werbos, eds., Cambridge MA: MIT
Press, 1990.
This paper discusses some conjugate gradient and quasi-
Newton optimization algorithms that could be used for
neural network training.
[Toll90] T. Tollenaere, SuperSAB: Fast adaptive back propagation
with good scaling properties, Neural Networks, vol. 3, no.
5, pp. 561573, 1990.
This paper presents a variable learning rate backpropaga-
tion algorithm in which different learning rates are used
for each weight.
[VoMa88] T. P. Vogl, J. K. Mangis, A. K. Zigler, W. T. Zink and D. L.
Alkon, Accelerating the convergence of the backpropaga-
tion method, Biological Cybernetics., vol. 59, pp. 256264,
Sept. 1988.
This was one of the first papers to introduce several heuris-
tic techniques for accelerating the convergence of back-
propagation. It included batching, momentum and variable
learning rate.
12-49
Exercises
E12.1 We want to train the network shown in Figure E12.1 on the training set

p 1 = 2 t 1 = 0.8 , p 2 = 2 t 2 = 1 ,

where each pair is equally likely to occur.

2+2
ans = Write a MATLAB M-file to create a contour plot for the mean squared error
4 performance index.
Input Log-Sigmoid Layer
w n a
p
b
1
a = logsig (w p + b)
E12.2 Demonstrate the effect of batching by computing the direction of the initial
step for SDBP with and without batching for the problem described in Ex-
ercise E12.1, starting from the initial guess
w 0 = 0 , b 0 = 0.5 .
E12.3 Recall the quadratic function used in Problem P9.1:
1 T
F x = --- x 10 6 x + 4 4 x .
2 6 10
We want to use the steepest descent algorithm with momentum to mini-

mize this function.
i. Suppose that the learning rate is = 0.2 . Find a value for the mo-
mentum coefficient for which the algorithm will be stable. Use the
ideas presented in Problem P12.2.
ii. Suppose that the learning rate is = 20 . Find a value for the mo-
mentum coefficient for which the algorithm will be stable.
12-50
Exercises
iii. Write a MATLAB program to plot the trajectories of the algorithm

2+2 for the and values of both part (i) and part (ii) on the contour
ans =
4
plot of F x , starting from the initial guess
x0 = 1 .
2.5
E12.4 Consider the following quadratic function.
1 T
F x = --- x 3 1 x + 4 4 x .
2 1 3

mize this function.
i. Perform two iterations (finding x 1 and x 2 ) of steepest descent with
T
momentum, starting from the initial condition x 0 = 0 0 . Use a
learning rate of = 1 and a momentum coefficient of = 0.75 .
ii. Is the algorithm stable with this learning rate and this momentum?
Use the ideas presented in Problem P12.2.
iii. Would the algorithm be stable with this learning rate, if the momen-
tum were zero?
1 T
F x = --- x 3 1 x + 1 2 x + 2 .
2 1 3

mize this function.
i. Suppose the learning rate is = 1 . Is the algorithm stable, if the
momentum coefficient is = 0 ? Use the ideas presented in Problem
P12.2.
ii. Suppose the learning rate is = 1 . Is the algorithm stable, if the

momentum coefficient is = 0.6 ?
12-51
1 T
F x = --- x 2 1 x + 1 2 x + 2 .
2 1 2

mize this function. Suppose the learning rate is = 1 . Find a value for the
momentum coefficient so that the algorithm will be stable. Use the ideas
presented in Problem Eq. P12.2.
E12.7 For the function of Exercise E12.3, perform three iterations of the variable
learning rate algorithm, with initial guess
x0 = 1 .
2.5
Plot the algorithm trajectory on a contour plot of F x . Use the algorithm

parameters
= 0.4 , = 0.1 , = 1.5 , = 0.5 , = 5% .

2 2
F x = x 1 + 2x 2 .
Perform three iterations of the variable learning rate algorithm, with ini-
tial guess
x0 = 0 .
1
Use the algorithm parameters
= 1 , = 0.2 , = 1.5 , = 0.5 , = 5% .
(Count an iteration each time the function is evaluated after the initial
guess.)
E12.9 For the function of Exercise E12.3, perform one iteration of the conjugate
gradient algorithm, with initial guess
x0 = 1 .
2.5
12-52
Exercises
For the linear minimization use interval location by function evaluation

and interval reduction by the Golden Section search. Plot the path of the
search on a contour plot of F x .
1 T
F x = --- x 4 0 x + 2 1 x .
2 0 2
We want to minimize this function along the line
x = 0 + 1 .
0 1
i. Sketch this line in the x 1 , x 2 plane.
ii. The learning rate must fall somewhere between 0 and 3. Perform
one iteration of the golden section search. You should find a 2 , b 2 , c 2
and d 2 , and indicate these points along the line that you drew in
part i.
1 T
F x = --- x 1 1 x + 1 1 x .
2 1 1
x = 0 + 1 .
0 0
i. Use the method described on page 12-16 to determine an initial in-

terval containing the minimum. Use = 0.5 .
ii. Take one iteration of the golden section search to reduce the inter-
val you obtained in part i.
1 T
F x = --- x 1 0 x .
2 0 2
12-53
x = 1 + 1 .
1 1
Perform two iterations of the Golden Section search ( k = 1 2 ) to find the

interval a 3 b 3 . Assume that the initial interval is defined by a 1 = 0 and
b 1 = 1 . Make a rough sketch of the contour plot of F x , draw the search
line in the same figure and indicate your search points (points where you
evaluated F x ) on the line.
1 T
F x = --- x 2 0 x .
2 0 1
x = 0 + 1 .
1 1
Perform two iterations of the Golden Section search ( k = 1 2 ) to find the

interval a 3 b 3 . Assume that the initial interval is defined by a 1 = 0 and
b 1 = 1 . Make a rough sketch of the contour plot of F x , draw the search
line in the same figure and indicate your search points (points where you
evaluated F x ) on the line.
E12.14 We want to use the network of Figure E12.2 to approximate the function

g p = 1 + sin --- p for 2 p 2 .
4
The initial network parameters are chosen to be
w 0 = 0.27 , b 0 = 0.48 , w 0 = 0.09 0.17 , b 0 = 0.48 .

1 1 2 2
0.41 0.13
To create the training set we sample the function g p at the points p = 1

and p = 0 . Find the Jacobian matrix for the first step of the LMBP algo-
rithm. (Some of the information you will need has been computed in the ex-
ample starting on page 11-14.)
12-54
Exercises
w11,1 n11 a11

w21,1
b11 n2 a2
p
1
n12 a12 b2
w12,1 w22,1
1
b12
1
E12.15 Show that for a linear network the LMBP algorithm will converge to an op-
timum solution in one iteration if = 0 .
E12.16 In Exercise E11.25 you wrote a MATLAB program to implement the SDBP
1
algorithm for a 1 S 1 network, and trained the network to approximate
the function

g p = 1 + sin --- p for 2 p 2 .
4
2+2
ans = Repeat this exercise, modifying your program to use the training proce-
4 dures discussed in this chapter: batch mode SDBP, MOBP, VLBP, CGBP
and LMBP. Compare the convergence results of the various methods.
12-55
Objectives
13 Generalization
13
Objectives 13-1
Methods for Improving Generalization 13-5
Estimating Generalization Error - The Test Set 13-6
Early Stopping 13-6
Regularization 13-8
Bayesian Analysis 13-10
Bayesian Regularization 13-12
Relationship Between Early Stopping and Regularization 13-19
Epilogue 13-44
Exercises 13-47
Objectives
One of the key issues in designing a multilayer network is determining the
number of neurons to use. In effect, that is the objective of this chapter.
In Chapter 11 we showed that if the number of neurons is too large, the net-
work will overfit the training data. This means that the error on the train-
ing data will be very small, but the network will fail to perform as well
when presented with new data. A network that generalizes well will per-
form as well on new data as it does on the training data.
The complexity of a neural network is determined by the number of free pa-
rameters that it has (weights and biases), which in turn is determined by
the number of neurons. If a network is too complex for a given data set,
then it is likely to overfit and to have poor generalization.
In this chapter we will see that we can adjust the complexity of a network
to fit the complexity of the data. In addition, this can be done without
changing the number of neurons. We can adjust the effective number of
free parameters without changing the actual number of free parameters.
13-1
13 Generalization
Theory and Examples

Mark Twain once said We should be careful to get out of an experience
only the wisdom that is in it-and stop there; lest we be like the cat that sits
down on a hot stove-lid. She will never sit down on a hot stove-lid again-
and that is well; but also she will never sit down on a cold one any more.
(From Following the Equator, 1897.)
That is the objective of this chapter. We want to train neural networks to
Generalization get out of the data only the wisdom that is in it. This concept is called gen-
eralization. A network trained to generalize will perform as well in new sit-
uations as it does on the data on which it was trained.
The key strategy we will use for obtaining good generalization is to find the
simplest model that explains the data. This is a variation of a principle
Ockhams Razor called Ockhams razor, which is named after the English logician William
of Ockham, who worked in the 14th Century. The idea is that the more
complexity you have in your model, the greater the possibility for errors.
In terms of neural networks, the simplest model is the one that contains
the smallest number of free parameters (weights and biases), or, equiva-
lently, the smallest number of neurons. To find a network that generalizes
well, we need to find the simplest network that fits the data.
There are at least five different approaches that people have used to pro-
duce simple networks: growing, pruning, global searches, regularization,
and early stopping. Growing methods start with no neurons in the network
and then add neurons until the performance is adequate. Pruning methods
start with large networks, which likely overfit, and then remove neurons
(or weights) one at a time until the performance degrades significantly.
Global searches, such as genetic algorithms, search the space of all possible
network architectures to locate the simplest model that explains the data.
The final two approaches, regularization and early stopping, keep the net-
work small by constraining the magnitude of the network weights, rather
than by constraining the number of network weights. In this chapter we
will concentrate on these two approaches. We will begin by defining the
problem of generalization and by showing examples of both good and poor
generalization. We will then describe the regularization and early stopping
methods for training neural networks. Finally, we will demonstrate how
these two methods are, in effect, performing the same operation.
Problem Statement
Lets begin our discussion of generalization by defining the problem. We
start with a training set of example network inputs and corresponding tar-
get outputs:
13-2
Problem Statement
{p 1, t 1} { p 2, t 2} {p Q, tQ} . (13.1)
For our development of the concept of generalization, we will assume that

the target outputs are generated by
tq = g pq + q , (13.2)
where g . is some unknown function, and q is a random, independent

and zero mean noise source. Our training objective will be to produce a neu-
ral network that approximates g . , while ignoring the noise.
The standard performance index for neural network training is the sum
squared error on the training set:
Q
tq aq
T
F x = ED = tq aq , (13.3)
q=1
where a q is the network output for input p q . We are using the variable E D
to represent the sum squared error on the training data, because later we
will modify the performance index to include an additional term.
Overfitting The problem of overfitting is illustrated in Figure 13.1. The blue curve rep-
resents the function g . . The large open circles represent the noisy target
points. The black curve represents the trained network response, and the
smaller circles filled with crosses represent the network response at the
training points. In this figure we can see that the network response exactly
matches the training points. However, it does a very poor job of matching
the underlying function. It overfits.
There are actually two kinds of errors that occur in Figure 13.1. The first
type of error, which is caused by overfitting, occurs for input values be-
tween -3 and 0. This is the region where all of the training data points oc-
ccur. The network response in this region overfits the training data and
will fail to perform well for input values that are not in the training set. The
Interpolation network does a poor job of interpolation; it fails to accurately approximate
the function near the training points.
The second type of error occurs for inputs in the region between 0 and 3.
The network fails to perform well in this region, not because it is overfit-
Extrapolation ting, but because there is no training data there. The network is extrapo-
lating beyond the range of the input data.
In this chapter we will discuss methods for preventing errors of interpola-
tion (overfitting). There is no way to prevent errors of extrapolation, unless
the data that is used to train the network covers all regions of the input
space where the network will be used. The network has no way of knowing
what the true function looks like in regions where there is no data.
13-3
13 Generalization
25
20
15
10
10
15
20
25
30
3 2 1 0 1 2 3
Figure 13.1 Example of Overfitting and Poor Extrapolation

In Figure 13.2 we have an example of a network that has been trained to
generalize well. The network has the same number of weights as the net-
work of Figure 13.1, and it was trained using the same data set, but it has
been trained in such a way that it does not fully use all of the weights that
are available. It only uses as many weights as necessary to fit the data. The
network response does not fit the function perfectly, but it does the best job
it can, based on limited and noisy data.
25
20
15
10
10
15
20
25
30
3 2 1 0 1 2 3
Figure 13.2 Example of Good Interpolation and Poor Extrapolation

In both Figure 13.1 and Figure 13.2 we can see that the network fails to ex-
trapolate accurately. This is understandable, since the network has been
provided with no information about the characteristics of the function out-
13-4
Methods for Improving Generalization
side of the range 3 p 0 . The network response outside this range will
be unpredictable. This is why it is important to have training data for all
regions of the input space where the network will be used. It is usually not
difficult to determine the required input range when the network has a sin-
gle input, as in this example. However, when the network has many inputs,
it becomes more difficult to determine when the network is interpolating
and when it is extrapolating.
This problem is illustrated in a simple way in Figure 13.3. On the left side
of this figure we see the function that is to be approximated. The range for
the input variables is 3 p 1 3 and 3 p 2 3 . The neural network was
trained over these ranges of the two variables, but only for p 1 p 2 . There-
fore, both p 1 and p 2 cover their individual ranges, but only half of the total
input space is covered. When p 1 p 2 , the network is extrapolating, and we
can see on the right side of Figure 13.3 that the network performs poorly in
this region. (See Problem P13.4 for another example of extrapolation.) If
there are many input variables, it will be quite difficult to determine when
the network is interpolating and when it is extrapolating. We will discuss
some practical ways of dealing with this problem in Chapter 22.
a) b)
8 8
6 6
4 4
2 2
t 0 a 0
2 2
4 4
6 6
3 3
2 2
3 3
1 2 1 2
0 1 0 1
p2 1 0 p2 1 0
2
3 3
2
1
p1 2
3 3
2
1
p1
Figure 13.3 Function (a) and Neural Network Approximation (b)

The remainder of this chapter will discuss methods for improving the gen-
eralization capability of neural networks. As we discussed earlier, there are
a number of approaches to this problem - all of which try to find the sim-
plest network that will fit the data. These approaches fit into two general
categories: restricting the number of weights (or, equivalently, the number
of neurons) in the network, or restricting the magnitude of the weights. We
will concentrate on two methods that we have found to be particularly use-
ful: early stopping and regularization. Both of these approaches attempt to
restrict the magnitude of the weights, although they do so in very different
13-5
13 Generalization
ways. At the end of this chapter, we will demonstrate the approximate

equivalence of the two methods.
We should note that in this chapter we are assuming that there is a limited
amount of data with which to train the network. If the amount of data is
unlimited, which in practical terms means that the number of data points
is significantly larger than the number of network parameters, then there
will not be a problem of overfitting.
Estimating Generalization Error - The Test Set

Before we discuss methods for improving the generalization capability of
neural networks, we should first discuss how we can estimate this error for
a specific neural network. Given a limited amount of available data, it is
important to hold aside a certain subset during the training process. After
the network has been trained, we will compute the errors that the trained
Test Set network makes on this test set. The test set errors will then give us an in-
dication of how the network will perform in the future; they are a measure
of the generalization capability of the network.
In order for the test set to be a valid indicator of generalization capability,
there are two important things to keep in mind. First, the test set must
never be used in any way to train the neural network, or even to select one
network from a group of candidate networks. The test set should only be
used after all training and selection is complete. Second, the test set must
be representative of all situations for which the network will be used. This
can sometimes be difficult to guarantee, especially when the input space is
high-dimensional or has a complex shape. We will discuss this problem in
more detail in Chapter 22, Practical Training Issues.
In the remaining sections of this chapter, we will assume that a test set has
been removed from the data set before training begins, and that this set
will be used at the completion of training to measure generalization capa-
bility.
Early Stopping
The first method we will discuss for improving generalization is also the
simplest method. It is called early stopping [WaVe94]. The idea behind this
method is that as training progresses the network uses more and more of
its weights, until all weights are fully used when training reaches a mini-
mum of the error surface. By increasing the number of iterations of train-
ing, we are increasing the complexity of the resulting network. If training
is stopped before the minimum is reached, then the network will effectively
be using fewer parameters and will be less likely to overfit. In a later sec-
tion of this chapter we will demonstrate how the number of parameters
changes as the number of iterations increases.
In order to use early stopping effectively, we need to know when to stop the
Cross-Validation training. We will describe a method, called cross-validation, that uses a
13-6
Validation Set validation set to decide when to stop [Sarl95]. The available data (after re-
moving the test set, as described above) is divided into two parts: a training
set and a validation set. The training set is used to compute gradients or
Jacobians and to determine the weight update at each iteration. The vali-
dation set is an indicator of what is happening to the network function in
between the training points, and its error is monitored during the training
process. When the error on the validation set goes up for several iterations,
the training is stopped, and the weights that produced the minimum error
on the validation set are used as the final trained network weights.
This process is illustrated in Figure 13.4. The graph at the bottom of this
figure shows the progress of the training and validation performance indi-
ces, F (the sum squared errors), during training. Although the training er-
ror continues to go down throughout the training process, a minimum of
the validation error occurs at the point labeled a, which corresponds to
training iteration 14. The graph at the upper left shows the network re-
sponse at this early stopping point. The resulting network provides a good
fit to the true function. The graph at the upper right demonstrates the net-
work response if we continue to train to point b, where the validation er-
ror has increased and the network is overfitting.
1.5 1.5
1 a 1 b
0.5 0.5
2 2
a 0 a 0
0.5 0.5
1 1
1.5 1.5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
p p
3
10
2
Validation
10
a b
F 10
1
Training
0
10
1
10
0 1 2
10 10 10
Iteration
Figure 13.4 Illustration of Early Stopping
13-7
13 Generalization
The basic concept for early stopping is simple, but there are several practi-
cal issues to be addressed. First, the validation set must be chosen so that
it is representative of all situations for which the network will be used. This
is also true for the test and training sets, as we mentioned earlier. Each set
must be roughly equivalent in its coverage of the input space, although the
size of each set may be different.
When we divide the data, approximately 70% is typically used for training,
with 15% for validation and 15% for testing. These are only approximate
numbers. A complete discussion of how to select the amount of data for the
validation set is given in [AmMu97].
Another practical point to be made about early stopping is that we should
use a relatively slow training method. During training, the network will
use more and more of the available network parameters (as we will explain
in the last section of this chapter). If the training method is too fast, it will
likely jump past the point at which the validation error is minimized.
To experiment with the effect of early stopping, use the MATLAB Neural
Network Design Demonstration Early Stopping (nnd13es).
Regularization
The second method we will discuss for improving generalization is called
regularization. For this method, we modify the sum squared error perfor-
mance index of Eq. (13.3) to include a term that penalizes network com-
plexity. This concept was introduced by Tikhonov [Tikh63]. He added a
penalty, or regularization, term that involved the derivatives of the approx-
imating function (neural network in our case), which forced the resulting
function to be smooth. Under certain conditions, this regularization term
can be written as the sum of squares of the network weights, as in
Q n
tq aq tq aq + xi ,
T 2
F x = E D + E W = (13.4)
q=1 i=1
where the ratio controls the effective complexity of the network solu-
tion. The larger this ratio is, the smoother the network response. (Note that
we could have used a single parameter here, but developments in later sec-
tions will require two parameters.)
Why do we want to penalize the sum squared weights, and how is this sim-
ilar to reducing the number of neurons? Consider again the example mul-
tilayer network shown in Figure 11.4. Recall how increasing a weight
increased the slope of the network function. You can see this effect again in
2
Figure 13.5, where we have changed the weight w 1 1 from 0 to 2. When the
weights are large, the function created by the network can have large
slopes, and is therefore more likely to overfit the training data. If we re-
strict the weights to be small, then the network function will create a
13-8
smooth interpolation through the training data - just as if the network had
a small number of neurons.
2
w 1 1 = 2
2
2
a 1
2
w 1 1 = 0
1
2 1 0 1 2
Figure 13.5 Effect of Weight on Network Response

To experiment with the effect of weight changes on the network function, use
the MATLAB Neural Network Design Demonstration Network Function
(nnd11nf).
The key to the success of the regularization method in producing a network
that generalizes well is the correct choice of the regularization ratio .
Figure 13.6 illustrates the effect of changing this ratio. Here we have
trained a 1-20-1 network on 21 noisy samples of a sine wave.
In the figure, the blue line represents the true function, and the large open
circles represent the noisy data. The black curve represents the trained
network response, and the smaller circles filled with crosses represent the
network response at the training points. From the figure, we can see that
the ratio = 0.01 produces the best fit to the true function. For ratios
larger than this, the network response is too smooth, and for ratios smaller
than this, the network overfits.
There are several techniques for setting the regularization parameter. One
approach is to use a validation set, such as we described in the section on
early stopping; the regularization parameter is set to minimize the squared
error on the validation set [GoLa98]. In the next two sections we will de-
scribe a different technique for automatically setting the regularization pa-
rameter. It is called Bayesian regularization.
13-9
13 Generalization
1.5 1.5
1 = 0 1
= 0.01
0.5 0.5
0 0
0.5 0.5
1 1
1.5 1.5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
1.5 1.5
1 = 0.25 1
= 1
0.5 0.5
0 0
0.5 0.5
1 1
1.5 1.5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Figure 13.6 Effect of Regularization Ratio

To experiment with the effect of regularization, use the MATLAB Neural
Network Design Demonstration Regularization (nnd13reg).
Bayesian Analysis
Thomas Bayes was a Presbyterian minister who lived in England during
the 1700s. He was also an amateur mathematician. His most important
work was published after his death. In it, he presented what is now known
as Bayes Theorem. The theorem states that if you have two random
events, A and B , then the conditional probability of the occurrence of A ,
given the occurrence of B can be computed as
P B A P A
P A B = ------------------------------- . (13.5)
P B
Eq. (13.5) is called Bayes rule. Each of the terms in this expression has a
name by which it is commonly referred. P A is called the prior probability.
It tells us what we know about A before we know the outcome of B . P A B
is called the posterior probability. This tells us what we know about A after
we learn about B . P B A is the conditional probability of B given A . Nor-
mally this term is given by our knowledge of the system that describes the
relationship between B and A . P B is the marginal probability of the
event B , and it acts as a normalization factor in Bayes rule.
13-10
To illustrate how Bayes rule can be used, consider the following medical
2
+2 situation. Assume that 1% of the population have a certain disease. There
is a test that can be performed to detect the presence of this disease. The
test is 80% accurate in detecting the disease in people who have it. Howev-
er, 10% of the time, someone without the disease will register a positive
test. If you take the test and register positive, your question would be:
What is the probability that I actually have the disease? Most of us (includ-
ing most physicians, as has been shown in many studies), would guess that
the probability is very high, considering that the test is 80% accurate in de-
tecting the disease in a sick person. However, this turns out not to be the
case, and Bayes rule can help us overcome this lack of intuition, when it
comes to probability.
Let A represent the event that you have the disease. Let B represent the
event that you have a positive test result. We can then use Bayes rule to
find P A B , which is the probability that you have the disease, given that
you have a positive test. We know that the prior probability P A would be
0.01, because 1% of the population have the disease. P B A is 0.8, because
the test is 80% accurate in detecting the disease in people who have it. (No-
tice that this conditional probability is based on our knowledge of the test
procedure and its accuracy.) In order to use Bayes rule, we need one more
term, which is P B . This is the probability of getting a positive test,
whether or not you have the disease. This can be obtained by adding the
probability of having a positive test when you have the disease to the prob-
ability of having a positive test when you dont have the disease:
P B = P A B + P A B = P B A P A + P B A P A , (13.6)
where we have used the definition of conditional probability:
PA B
P B A = ----------------------- , or P A B = P B A P A . (13.7)
P A
If we plug in our known probabilities into Eq. (13.6), we find
P B = 0.8 0.01 + 0.1 0.99 = 0.107 , (13.8)
where P B A is 0.1, because 10% of healthy people register a positive test.

We can now use Bayes rule to find the posterior probability P A B :
P B A P A 0.8 0.01
P A B = ------------------------------- = ------------------------ = 0.0748 . (13.9)
P B 0.107
This tells us that even if you get a positive test, you only have a 7.5% chance
of having the disease. For most of us, this result is not intuitive.
The key to Bayes rule is the prior probability P A . In this case, the prior
odds of having the disease were only 1 in 100. If this number had been
much higher, then our posterior probability P A B would have also in-
13-11
13 Generalization
creased significantly. It is important when using Bayes rule to have the

prior probability P A accurately reflect our prior knowledge.
For another example of using Bayes rule and the effect of the prior density,
see Solved Problem P13.2 and its associated demonstration.
In the next section, we will apply Bayesian analysis to the training of mul-
tilayer networks. The advantage of Bayesian methods is that we can insert
prior knowledge through the selection of the prior probability. For neural
network training, we will make the prior assumption that the function we
are approximating is smooth. This means that the weights cannot be too
large, as was demonstrated in Figure 13.5. The trick will be to incorporate
this prior knowledge into an appropriate choice for the prior probability.
Bayesian Regularization
Although there have been many approaches to the automatic selection of
the regularization parameter, we will concentrate on one developed by
David MacKay [MacK92]. This approach puts the training of neural net-
works into a Bayesian statistical framework. This framework is useful for
many aspects of training, in addition to the selection of the regularization
parameter, so it is an important concept to become familiar with. There are
two levels to this Bayesian analysis. We will begin with Level I.
Level I Bayesian Framework

The Bayesian framework begins with the assumption that the network
weights are random variables. We then choose the weights that maximize
the conditional probability of the weights given the data. Bayes rule is
used to find this probability function:
P D x M P x M
P x D M = ----------------------------------------------------------- , (13.10)
P D M
where x is the vector containing all of the weights and biases in the net-
work, D represents the training data set, and are parameters associ-
ated with the density functions P D x M and P x M , and M is the
selected model - the architecture of the network we have chosen (i.e., how
many layers and how may neurons in each layer).
It is worth taking some time to investigate each of the terms in Eq. (13.10).
First, P D x M is the probability density for the data, given a certain
set of weights x , the parameter (which we will explain shortly), and the
choice of model M . If we assume that the noise terms in Eq. (13.2) are in-
dependent and have a Gaussian distribution, then
1
P D x M = --------------- exp E D , (13.11)
ZD
13-12
2 2
where = 1 2 , is the variance of each element of q , E D is the
squared error (as defined in Eq. (13.3)), and
2 N2 N2
Z D = 2 = , (13.12)
M
where N is Q S , as in Eq. (12.34).
Likelihood Function Eq. (13.11) is called the likelihood function. It is a function of the network
weights x , and it describes how likely a given data set is to occur, given a
Maximum Likelihood specific set of weights. The maximum likelihood method selects the weights
so as to maximize the likelihood function, which in this Gaussian case is
the same as minimizing the squared error E D . Therefore, our standard
sum squared error performance index can be derived statistically with the
assumption of Gaussian noise in the training set, and our standard choice
for the weights is the maximum likelihood estimate.
Now consider the second term on the right side of Eq. (13.10): P x M .
Prior Density This is called the prior density. It embodies our knowledge about the net-
work weights before we collect any data. Bayesian statistics allows us to in-
corporate prior knowledge through the prior density. For example, if we
assume that the weights are small values centered around zero, we might
select a zero-mean Gaussian prior density:
1
P x M = ---------------- exp E W (13.13)
ZW
2 2
where = 1 2 w , w is the variance of each of the weights, E W is the
sum squared weights (as defined in Eq. (13.4)), and
2 n2 n2
Z W = 2 w = , (13.14)
where n is the number of weights and biases in the network, as in Eq.

(12.35).
The final term on the right side of Eq. (13.10) is P D M . This is called
Evidence the evidence, and it is a normalizing term that is not a function of x . If our
Posterior Density objective is to find the weights x that maximize the posterior density
P x D M , then we do not need to be concerned with P D M .
(However, it will be important later for estimating and .)
With the Gaussian assumptions that we made earlier, we can rewrite the
posterior density, using Eq. (13.10), in the following form:
13-13
13 Generalization
1 1
---------------- --------------- exp E D + E W
ZW ZD
P x D M = -----------------------------------------------------------------------------------
Normalization Factor (13.15)
1
= --------------------- exp F x
Z F
where Z F is a function of and (but not a function of x ), and F x

is our regularized performance index, which we defined in Eq. (13.4). To
find the most probable value for the weights, we should maximize the pos-
terior density P x D M . This is equivalent to minimizing the regu-
larized performance index F x = E D + E W .
Therefore, our regularized performance index can be derived using Baye-
sian statistics, with the assumption of Gaussian noise in the training set
and a Gaussian prior density for the network weights. We will identify the
Most Probable weights that maximize the posterior density as x MP , or most probable. This
is to be contrasted with the weights that maximize the likelihood function:
x ML .
Note how this statistical framework provides a physical meaning for the
parameters and . The parameter is inversely proportional to the
variance in the measurement noise q . Therefore, if the noise variance is
large, will be small, and the regularization ratio will be large. This
will force the resulting weights to be small and the network function to be
smooth (as seen in Figure 13.6). The larger the measurement noise, the
more we will smooth the network function, in order to average out the af-
fects of the noise.
The parameter is inversely proportional to the variance in the prior dis-
tribution for the network weights. If this variance is large, it means that
we have very little certainty about the values of the network weights, and,
therefore, they might be very large. The parameter will then be small,
and the regularization ratio will also be small. This will allow the net-
work weights to be large, and the network function will be allowed to have
more variation (as seen in Figure 13.6). The larger the variance in the prior
density for the network weights, the more variation the network function
will be allowed to have.
Level II Bayesian Framework

So far we have an interesting statistical derivation of the regularized per-
formance index and some new insight into the meanings of the parameters
and , but what we really want to find is a way to estimate these param-
eters from the data. In order to do this, we need to take the Bayesian anal-
ysis to another level. If we want to estimate and using Bayesian
analysis, we need the probability density P D M . Using Bayes rule
this can written
13-14
P D M P M
P D M = ------------------------------------------------------------ . (13.16)
PD M
This has the same format as Eq. (13.10), with the likelihood function and
the prior density in the numerator of the right hand side. If we assume a
uniform (constant) prior density P M for the regularization parame-
ters and , then maximizing the posterior is achieved by maximizing the
likelihood function P D M . However, note that this likelihood func-
tion is the normalization factor (evidence) from Eq. (13.10). Since we have
assumed that all probabilities have a Gaussian form, we know the form for
the posterior density of Eq. (13.10). It is shown in Eq. (13.15). Now we can
solve Eq. (13.10) for the normalization factor (evidence).
P D x M P x M
P D M = ------------------------------------------------------------
P x D M
1 1
--------------- exp E D ---------------- exp E W
ZD ZW (13.17)
= -----------------------------------------------------------------------------------------------------
1
--------------------- exp F x
Z F
Z F exp E D E W Z F
- ---------------------------------------------
= ------------------------------ = ------------------------------
-
Z D Z W exp F x Z D Z W
Note that we know the constants Z D and Z W from Eq. (13.12) and
Eq. (13.14). The only part we do not know is Z F . However, we can es-
timate it by using a Taylor series expansion.
Since the objective function has the shape of a quadratic in a small area
surrounding a minimum point, we can expand F x in a second order Tay-
lor series (see Eq. (8.9)) around its minimum point, x MP , where the gradi-
ent is zero:
1 T MP
F x F x MP + --- x x MP H x x MP , (13.18)
2
MP
where H = 2E D + 2E W is the Hessian matrix of F x , and H is the
Hessian evaluated at x MP . We can now substitute this approximation into
the expression for the posterior density, Eq. (13.15):
1 1 T MP
P x D M ------ exp F x MP --- x x MP H x x MP , (13.19)
ZF 2
which can be rewritten as
13-15
13 Generalization
1 1 T MP
P x D M ------ exp F x MP exp --- x x MP H x x MP . (13.20)
Z
F 2
The standard form of the Gaussian density is
P x = ------------------------------------------ exp --- x x MP H x x MP .

1 1 T MP
(13.21)
n MP 1 2
2 H
Therefore, equating Eq. (13.21) with Eq. (13.20), we can solve for Z F :
n2 MP 1 12
Z F 2 det H exp F x MP . (13.22)
Placing this result into Eq. (13.17), we can solve for the optimal values for
and at the minimum point. We do this by taking the derivative with
respect to each of the log of Eq. (13.17) and set them equal to zero. This
yields (see Solved Problem P13.3):
MP MP N
- and
= ------------------------
MP
-,
= ------------------------
MP
(13.23)
2E W x 2E D x
MP MP 1
Effective # of Parameters where = n 2 tr H is called the effective number of parameters,
and n is the total number of parameters in the network. The term is a
measure of how many parameters (weights and biases) in the neural net-
work are effectively used in reducing the error function. It can range from
zero to n . (See the example on page 13-23 for more analysis of .)
Bayesian Regularization Algorithm

The Bayesian optimization of the regularization parameters requires the
computation of the Hessian matrix of F x at the minimum point x MP . We
propose using the Gauss-Newton approximation to the Hessian matrix
[FoHa97], which is readily available if the Levenberg-Marquardt optimiza-
tion algorithm is used to locate the minimum point (see Eq. (12.31)). The
additional computation required for optimization of the regularization is
minimal.
Here are the steps required for Bayesian optimization of the regularization
parameters, with the Gauss-Newton approximation to the Hessian matrix:
0. Initialize , and the weights. The weights are initialized randomly,
and then E D and E W are computed. Set = n , and compute and
using Eq. (13.23).
1. Take one step of the Levenberg-Marquardt algorithm toward minimiz-
ing the objective function F x = E D + E W .
1
2. Compute the effective number of parameters = n 2tr H , mak-
13-16
ing use of the Gauss-Newton approximation to the Hessian available in

the Levenberg-Marquardt training algorithm:
T
H = 2F x 2J J + 2I n , where J is the Jacobian matrix of the
training set errors (see Eq. (12.37)).

3. Compute new estimates for the regularization parameters = ------------------
N 2E Wx
and = ------------------ .
2E D x
4. Now iterate steps 1 through 3 until convergence.
Bear in mind that with each reestimate of the regularization parameters
and the objective function F x changes; therefore, the minimum
point is moving. If traversing the performance surface generally moves to-
ward the next minimum point, then the new estimates for the regulariza-
tion parameters will be more precise. Eventually, the precision will be good
enough that the objective function will not significantly change in subse-
quent iterations. Thus, we will obtain convergence.
When this Gauss-Newton approximation to Bayesian regularization
GNBR (GNBR) algorithm is used, the best results are obtained if the training data
is first mapped into the range [-1,1] (or some similar region). We will dis-
cuss this preprocessing of the training data in Chapter 22.
In Figure 13.7 you can see the results of training a 1-20-1 network with
GNBR on the same data set represented in Figure 13.4 and Figure 13.6.
The network has fit the underlying function, without overfitting to the
noise. The fit looks similar to that obtained in Figure 13.6, with the regu-
larization ratio set to = 0.01 . In fact, at the completion of training
with GNBR, the final regularization ratio for this example was
= 0.0137 .
The training process for this example is illustrated in Figure 13.8. In the
upper left of this figure, you see the squared error on the training set. No-
tice that it does not necessarily go down at each iteration. In the upper
right of the figure, you see the squared testing error. This was obtained by
comparing the network function to the true function at a number of points
between -1 and 1. It is a measure of the generalization capability of the net-
work. (This would not be possible in a practical case, where the true func-
tion was unknown.) Note that the testing error is at its minimum at the
completion of training.
13-17
13 Generalization
1.5
0.5
0.5
1.5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Figure 13.7 Bayesian Regularization Fit

Figure 13.8 also shows the regularization ratio and the effective num-
ber of parameters during training. These parameters have no particular
meaning during the training process, but at the completion of training they
are significant. As we mentioned earlier, the final regularization ratio was
= 0.0137 , which is consistent with our earlier investigation of regular-
izationillustrated in Figure 13.6. The final effective number of parame-
ters was = 5.2 . This is out of a total of 61 total weights and biases in the
network.
2 3
10 10
Training Testing
2
0
10
10
ED ED
1
10
2
10
0
10
0 1 2 3 0 1 2 3
10 10 10 10 10 10 10 10
Iteration Iteration
1 2
10 10
2
10
10
1
3
10
4 0
10 10
0 1 2 3 0 1 2 3
10 10 10 10 10 10 10 10
Iteration Iteration
Figure 13.8 Bayesian Regularization Training Process

The fact that in this example the effective number of parameters is much
less than the total number of parameters (6 versus 61) means that we
13-18
might well have been able to use a smaller network to fit this data. There
are two disadvantages of a large network: 1) it may overfit the data, and 2)
it requires more computation to calculate the network output. We have
overcome the first disadvantage by training with GNBR; although the net-
work has 61 parameters, it is equivalent to a network with only 6 parame-
ters. The second disadvantage is only important if the calculation time for
the network response is critical to the application. This is not usually the
case, since the time to calculate a network response to a particular input is
measured in milliseconds. In those cases where the calculation time is sig-
nificant, you can train a smaller network on the data.
On the other hand, when the effective number of parameters is close to the
total number of parameters, this can mean that the network is not large
enough to fit the data. In this case, you should increase the size of the net-
work and retrain on the data set.
To experiment with Bayesian Regularization, use the MATLAB Neural
Network Design Demonstration Bayesian Regularization (nnd17breg).
Relationship Between Early Stopping and Regularization

We have discussed two techniques for improving network generalization:
early stopping and regularization. These two methods were developed in
very different ways, but they both improve generalization by restricting the
network weights and, therefore, producing a network with fewer effective
parameters. Early stopping restricts the network weights by stopping the
training before the weights have converged to the minimum of the squared
error. Regularization restricts the weights by adding a term to the squared
error that penalizes large weights. In this section we want to demonstrate,
using a linear example, an approximate equivalence of these two methods.
During the process, we will also shed some light on the meaning of the ef-
fective number of parameters, . This development is based on the more
general procedures described in [SjLj94].
Early Stopping Analysis

Consider the single layer linear network shown in Figure 10.1. We have
shown in Eq. (10.12) and Eq. (10.14) that the mean square error perfor-
mance function for this linear network is quadratic, of the form
T 1 T
F x = c + d x + --- x Ax , (13.24)
2
where A is the Hessian matrix. In order to study the performance of early

stopping, we will analyze the evolution of the steepest descent algorithm on
this linear network. From Eq. (10.16), we know that the gradient of the per-
formance index is
F x = Ax + d . (13.25)
13-19
13 Generalization
Therefore, the steepest descent algorithm (see Eq. (9.10)) will be
x k + 1 = x k g k = x k Ax k + d . (13.26)
We want to know how close we come to the minimum of the squared error
at each iteration. For quadratic performance indices, we know that the
minimum will occur at the following point (see Eq. (8.62)):
ML 1
x = A d , (13.27)
where the superscript ML indicates that this result maximizes the likeli-
hood function, in addition to minimizing the squared error, as we saw in
Eq. (13.11).
We can now rewrite Eq. (13.26) as
1 ML
x k + 1 = x k A x k + A d = x k A x k x . (13.28)
With some additional algebra we can find

ML ML
x k + 1 = I A x k + Ax = Mx k + I M x , (13.29)
where M = I A . The next step is to relate x k + 1 to the initial guess x 0 .

Starting at the first iteration, using Eq. (13.29), we have
ML
x 1 = Mx 0 + I M x , (13.30)
where the initial guess x 0 usually consists of random values near zero.
Continuing to the second iteration:
ML
x 2 = Mx 1 + I M x
2 ML ML
= M x 0 + M I M x + I M x
. (13.31)
2 ML 2 ML ML ML
= M x 0 + Mx M x +x Mx
2 ML 2 ML 2 2 ML
= M x0 + x M x = M x 0 + I M x
Following similar steps, at the kth iteration we have

k k ML
x k = M x 0 + I M x , (13.32)
This key result shows how far we progress from the initial guess to the
maximum likelihood weights in k iterations. We will use this result later to
compare with regularization.
13-20
Regularization Analysis
Recall from Eq. (13.4) that the regularized performance index adds a pen-
alty term to the sum squared error, as in
F x = E D + E W . (13.33)
For the following analysis, it will more convenient to consider the following
equivalent (because the minimum occurs at the same place) performance
index
Fx
F x = ----------- = E D + ---E = E D + E W , (13.34)
W
which has only one regularization parameter.

The sum squared weight penalty term E W can be written
T
EW = x x0 x x0 , (13.35)
where the nominal value x 0 is normally taken to be the zero vector.

In order to locate the minimum of the regularized performance index,
which is also the most probable value x MP , we will set the gradient equal
to zero:
F x = E D + E W = 0 . (13.36)
The gradient of the penalty term, Eq. (13.35), is
E W = 2 x x 0 . (13.37)
From Eq. (13.25) and Eq. (13.28), the gradient of the sum squared error is
1 ML
E D = Ax + d = A x + A d = A x x . (13.38)
We can now set the total gradient to zero:
F x = A x x
ML
+ 2 x x 0 = 0 . (13.39)
The solution of Eq. (13.39) is the most probable value for the weights, x MP .
We can make that substitution and perform some algebra to obtain
MP ML MP MP ML ML
Ax x = 2 x x 0 = 2 x x +x x0
(13.40)
MP ML ML
= 2 x x 2 x x0
MP ML
Now combine the terms multiplying x x :
13-21
13 Generalization
MP ML ML
A + 2I x x = 2 x 0 x . (13.41)
MP ML
Solving for x x , we find
MP ML 1 ML ML
x x = 2 A + 2I x 0 x = M x0 x , (13.42)
1
where M = 2 A + 2I .
We want to know the relationship between the regularized solution x MP
and the minimum of the squared error x ML , so we can solve Eq. (13.42) for
x MP :
MP ML
x = M x 0 + I M x . (13.43)
This is the key result that describes the relationship between the regular-
ized solution and the minimum of the squared error. By comparing Eq.
(13.43) with Eq. (13.32), we can investigate the relationship between early
stopping and regularization. We will do that in the next section.
Connection Between Early Stopping and Regularization

To compare early stopping and regularization, we need to compare Eq.
(13.43) and Eq. (13.32). They are summarized in Figure 13.9. We would like
to find out when these two solutions are equal. In other words, when do ear-
ly stopping and regularization produce the same weights?
Early Stopping Regularization

MP ML
x = M x 0 + I M x
k k ML
x k = M x 0 + I M x
1
M = I A M = 2 A + 2I
Figure 13.9 Early Stopping and Regularization Solutions

k k
The key matrix for early stopping is M = I A . The key matrix for
1
regularization is M = 2 A + 2I . If these two matrices are equal, then
the weights for early stopping will be the same as the weights for regular-
ization. In Eq. (9.22) we showed that the eigenvectors of M are the same
as the eigenvectors of A and that the eigenvalues of M are 1 i ,
k
where the eigenvalues of A are i . The eigenvalues of M are then
k k
eig M = 1 i . (13.44)
Now lets consider the matrix M . First, using the same procedures that
led to Eq. (9.22), we can show that the eigenvectors of A + 2I are the
same as the eigenvectors of A , and the eigenvalues of A + 2I are
13-22
2 + i . Also, the eigenvectors of the inverse of a matrix are the same as

the eigenvectors of the original matrix, and the eigenvalues of the inverse
are the reciprocals of the original eigenvalues. Therefore, the eigenvectors
of M are the same as the eigenvectors of A , and the eigenvalues of M are
2
13
eig M = ---------------------- . (13.45)
i + 2
k
Therefore, in order for M to equal M , they just need to have equal eigen-
values:
2 k
---------------------- = 1 i . (13.46)
i + 2
Take the logarithm of both sides:

log 1 + -----i- = k log 1 i . (13.47)
2
These expressions are equal at i = 0 , so they will always be equal if their

derivatives are equal. Taking derivatives of both sides, we have
1 1 k
--------------------- --- = ----------------- , (13.48)

1 + -----i- 1 i
2
or
1 1 i
k = ------ ---------------------------------
-. (13.49)
2 1 + i 2
If i is small (slow, stable learning) and i 2 is small, then we have

the approximate result
1
k ------ . (13.50)
2
Therefore, early stopping is approximately equivalent to regularization. In-

creasing the number of iterations k is approximately the same as decreas-
ing the regularization parameter . This makes intuitive sense, because
increasing the number of iterations, or decreasing the regularization pa-
rameter, can lead to overfitting.
Example, Interpretation of Effective Number of Parameters

2 We will illustrate this result with a simple example. Suppose that we have
+2
a single layer, linear network with no bias. The input/target pairs are given
by
13-23
13 Generalization
1 t = 1 , p = 1 t = 1 ,
p1 = 1 2 2
1 1
where the probability of the first pair is 0.75, and the probability of the sec-
ond pair is 0.25. Following Eq. (10.13) and Eq. (10.15), we can find the qua-
dratic mean square error performance index as
2 2 2
c = E t = 1 0.75 + 1 0.25 = 1 ,
h = E tz = 0.75 1 1 + 0.25 1 1 = 1 ,
1 1 0.5
d = 2h = 2 1 = 2 ,
0.5 1

A = 2R = 2 E zz = 2 0.75 1 1 1 + 0.25 1 1 1 = 2 1 ,
T
1 1 1 2
T 1 T
E D = c + x d + --- x Ax .
2
The minimum of the mean squared error occurs at

1
= A d = R h = 1 0.5 1 = 1 .
ML 1 1
x
0.5 1 0.5 0
Now lets investigate the eigensystem of the Hessian matrix of E D :
E D x = A = 2R = 2 1 .
2
1 2
To find the eigenvalues:
2 1 2
= 4 + 3 = 1 3 ,
A I =
1 2
1 = 1 2 = 3 .
To find the eigenvectors:
13-24
A I v = 0 .
For 1 = 1 ,
11 v = 0 v1 = 1 ,
1
11 1
and for 2 = 3 ,
1 1 v = 0 v2 = 1 .
2
1 1 1
The contour plot for E D is shown in Figure 13.10
0.5
0
x2
0.5
1
0.5 0 0.5 1 1.5
x1
Figure 13.10 Contour Plot for E D
Now consider the regularized performance index of Eq. (13.34). Its Hessian
matrix will be
F x = 2E D + 2E W = 2E D + 2I = 2 1 + 2 0 = 2 + 2 1
2
.
1 2 0 2 1 2 + 2
In Figure 13.11 we have contour plots for F as is equal to 0, 1 and .
13-25
13 Generalization
= 1
0.5 = = 0
0
x2
0.5
1
0.5 0 0.5 1 1.5
x1
Figure 13.11 Contour Plot for F
In Figure 13.12 the blue curve represents the movement of x MP as is var-

ied.
x MP
0.5 = = 0
0
x2
0.5
1
0.5 0 0.5 1 1.5
x1
Figure 13.12 x MP as is Varied
Now lets compare this regularization result with early stopping. Figure
13.13 shows the steepest descent trajectory for minimizing E D , starting
from very small values for the weights. If we stop early, the result will fall
along the blue curve. Notice that this curve is very close to the regulariza-
tion curve in Figure 13.12. If the number of iterations is very small, this is
equivalent to a very large value for . As the number of iterations increas-
es, it is equivalent to reducing .
13-26
0.5
0
x2
0.5
1
0.5 0 0.5 1 1.5
x1
Figure 13.13 Steepest Descent Trajectory

To experiment with the relationship between Early Stopping and Regular-
ization, use the MATLAB Neural Network Design Demonstration Early
Stopping/Regularization (nnd17esr).
It is useful to consider the relationship between the eigenvalues and eigen-
2
vectors of the Hessian matrix E D x and the results of regularization
and early stopping. In this example, 2 is larger than 1 , so E D has higher
curvature in the v 2 direction. This means that we will get a quicker reduc-
tion in the squared error if we move in that direction first. This is shown in
Figure 13.13, as the initial steepest descent movement is almost in the di-
rection of v 2 . Note also that in regularization, as shown in Figure 13.12, as
decreases from a large value, the weights move first in the v 2 direction.
For a given change in the weights, this direction provides the largest reduc-
tion in the squared error.
Since the eigenvalue 1 is smaller than 2 , we only move in the v 1 direc-
tion after achieving significant reduction in E D in the v 2 direction. This
would be even more pronounced if the difference between 1 and 2 were
greater. In the limiting case, where 1 = 0 , we would not have to move in
the v 1 direction at all. We would only need to move in the v 2 direction to
get the complete reduction in the squared error. (This would be the case of
the stationary valley, as in Figure 8.9.) Note that in this case we would only
be effectively using one parameter, even though the network has two
weights. (Of course, this one effective parameter is some combination of the
two weights.) Therefore, the effective number of parameters is related to
2
the number of eigenvalues of E D x that are significantly different than
zero. We will analyze this in detail in the next section.
13-27
13 Generalization
Effective Number of Parameters

Recall the previous definition for the effective number of parameters:
MP MP 1
= n 2 tr H (13.51)
2
We can express this in terms of the eigenvalues of E D x . First, we can
write the Hessian matrix as
2
H x = F x = 2E D + 2E W = 2E D + 2I . (13.52)
Using arguments similar to those leading to Eq. (13.44), we can show that
the eigenvalues of H x are i + 2 . We can then use two properties of
1 1
eigenvalues to compute tr H . First, the eigenvalues of H are the re-
ciprocals of the eigenvalues of H , and, second, the trace of a matrix is equal
to the sum of its eigenvalues. Using these two properties, we can write
n
1
---------------------
1
tr H = . (13.53)
i + 2
i=1
We can now write the effective number of parameters as

n n
MP 1 2 i
---------------------
MP
= n 2 tr H = n --------------------- = , (13.54)
i + 2 i + 2
i=1 i=1
or
n n
i
= ---------------------
i + 2
= i , (13.55)
i=1 i=1
where
i
i = --------------------- . (13.56)
i + 2
Note that 0 i 1 , so the effective number of parameters must fall be-

2
tween zero and n . If all of the eigenvalues of E D x are large, then the
effective number of parameters will equal the total number of parameters.
If some of the eigenvalues are very small, then the effective number of pa-
rameters will equal the number of large eigenvalues, as was also demon-
strated by our example in the previous section. Large eigenvalues mean
large curvature, which means that the performance index changes rapidly
along those eigenvectors. Every large eigenvector represents a productive
direction for optimizing performance.
13-28
Summary of Results
Summary of Results
Problem Statement 13
A network trained to generalize will perform as well in new situations as it
does on the data on which it was trained.
Q
tq aq
T
ED = tq aq
q=1

Estimating Generalization Error - The Test Set
Given a limited amount of available data, it is important to hold aside a cer-
tain subset during the training process. After the network has been
trained, we will compute the errors that the trained network makes on this
test set. The test set errors will then give us an indication of how the net-
work will perform in the future; they are a measure of the generalization
capability of the network.
Early Stopping
The available data (after removing the test set) is divided into two parts: a
training set and a validation set. The training set is used to compute gra-
dients or Jacobians and to determine the weight update at each iteration.
When the error on the validation set goes up for several iterations, the
training is stopped, and the weights that produced the minimum error on
the validation set are used as the final trained network weights.
Regularization
Q n
tq aq tq aq + xi
T 2
F x = E D + E W =
q=1 i=1
Bayesian Regularization
Level I Bayesian Framework
P D x M P x M
P x D M = -----------------------------------------------------------
P D M
13-29
13 Generalization
1 2
P D x M = --------------- exp E D , = 1 2
ZD
2 N2 N2
Z D = 2 =
1 2
P x M = ---------------- exp E W , = 1 2 w
ZW
2 n2 n2
Z W = 2 w =
1
P x D M = --------------------- exp F x
Z F
Level II Bayesian Framework

P D M P M
P D M = ------------------------------------------------------------
PD M
MP MP N
MP
- and
= ------------------------ = ------------------------
MP
-
2E W x 2E D x
MP MP 1
= n 2 tr H
Bayesian Regularization Algorithm

0. Initialize , and the weights. The weights are initialized randomly,
and then E D and E W are computed. Set = n , and compute and
using Eq. (13.23).
1. Take one step of the Levenberg-Marquardt algorithm toward minimiz-
ing the objective function F x = E D + E W .
1
2. Compute the effective number of parameters = N 2tr H , mak-
ing use of the Gauss-Newton approximation to the Hessian available in
the Levenberg-Marquardt training algorithm:
T
H = 2F x 2J J + 2I n , where J is the Jacobian matrix of the
training set errors (see Eq. (12.37)).

3. Compute new estimates for the regularization parameters = ------------------
N 2E W x
and = ------------------ .
2E D x
4. Now iterate steps 1 through 3 until convergence.
13-30
Summary of Results
Relationship Between Early Stopping and Regularization
Early Stopping Regularization

k k
x k = M x 0 + I M x
ML
x
MP
= M x 0 + I M x
ML 13
1
M = I A M = 2 A + 2I
k k
eig M = 1 i
2
eig M = ----------------------
i + 2
1
k ------
2

n
i
= ---------------------
i + 2
i=1
0n
13-31
13 Generalization
Solved Problems
P13.1 In this problem and in the following one we want to investigate the
relationship between maximum likelihood methods and Bayesian
methods. Suppose that we have a random variable that is uniform-
ly distributed between 0 and x. We take a series of Q independent
samples of the random variable. Find the maximum likelihood es-
timate of x.
Before we begin this problem, lets review the Level I Bayesian formulation
of Eq. (13.10). We will not need the Level II formulation for this simple
problem, so we do not need the regularization parameters. Also, we only
have a single parameter to estimate, so x is a scalar. Eq. (13.10) can then
be simplified to
P D x P x
P x D = ------------------------------ .
PD
We are interested in the maximum likelihood estimate for this problem, so

we need to find the value of x that maximizes the likelihood term P D x .
The data is the Q independent samples from the uniformly distributed ran-
dom variable. A graph of the uniform density function is given in Figure
P13.1.
ft x
1
---
x
t
x
Figure P13.1 Uniform Density Function

The definition can be written
1
--- 0 t x
f t x = x .
0 elsewhere

13-32
Solved Problems
If we have Q independent samples of the random variable, then we can

multiply each of the individual probabilities to get the joint probability of
all samples:
1 1
Q
----- 0 t i x for all i ----- x max t i
P D x = f ti x = x Q Q
= x
0 x max t
i=1
0 elsewhere i
The plot of the resulting likelihood function is shown in Figure P13.1.
P D x
x
max t i
Figure P13.2 Likelihood Function for Solved Problem P13.1

From this plot, we can see that the value of x that maximizes the likelihood
function is
ML
x = max t i .
Therefore, the maximum likelihood estimate of x is the maximum value

obtained from the Q independent samples of the random variable. This
seems like a reasonable estimate of x , which is the upper limit of the ran-
dom variable.
P13.2 In this problem we will compare the maximum likelihood and

Bayesian estimators. Assume that we have a series of measure-
ments of a random signal in noise:
ti = x + i .
Assume that the noise has a Gaussian density, with zero mean:
13-33
13 Generalization
i
2
1
f i = -------------- exp --------
-
2
2
2
i. Find the maximum likelihood estimate of x .

ii. Find the most probable estimate of x . Assume that x is a
zero-mean random variable, with Gaussian prior density:
1 x2 1
f x = ---------------- exp ----------2- = ---------------- exp E W
2 x 2 x ZW
i. To find the maximum likelihood estimate, we need to find the likelihood

function P D x . This represents the density of the data, given x . The first
step is to use the noise density to find the density of the measurement.
Since, with x given, the density for the measurement would be the same as
the density for the noise, but with a mean of x , we have
ti x
2
1
f t i x = -------------- exp ------------------ .
2
2
2
Assuming that the measurement noises are independent, we can multiply

the probability densities:
P D x = f t 1 t 2 t Q x = f t 1 x f t 2 x f t Q x = P D x
Q
2
ti x
1 i=1 1
= -------------------------- exp --------------------------- = ----------- exp E D
2
Q2 Q
2
2
Z

where
Q Q
1 Q2
ei , Z
2 2
= --------2- , E D = ti x = = .
2 i=1 i=1
To maximize the likelihood, we should minimize E D . Setting the derivative

to zero, we find
Q Q Q
dE D d
dx i
t x = 2 t i x = 2 t i Qx = 0 .
2
=
dx
i=1 i=1 i=1
13-34
Solved Problems
Solving for x , we find the maximum likelihood estimate:

Q
1
= ---- t i
ML
x
Q
i=1
ii. To find the most probable estimate, we need to use Bayes rule (Eq.
(13.10)) to find the posterior density:
P D x P x
P x D = ------------------------------ .
PD
The likelihood function P D x was found above to be
1
P D x = ----------- exp E D
Z
The prior density is
1 x2 1
P x = f x = ---------------- exp ----------2- = ---------------- exp E W ,
2 x 2 x ZW
where
1 12 2
= ----------2- , Z W = , E W = x .
2 x
The posterior density can then be computed as
P x D = f x t 1 t 2 t Q
f t 1 t 2 t Q x f x
= -------------------------------------------------
f t 1 t 2 t Q
1 1
--------------- ---------------- exp E D + E W
ZD ZW
= -----------------------------------------------------------------------------------
Normalization Factor
To find the most probable value for x, we maximize the posterior density.
This is equivalent to minimizing
Q
E D + E W = t i x + x .
2 2
i=1
To find the minimum, we take the derivative with respect to x and set it
equal to zero:
13-35
13 Generalization
Q Q
d d 2
t i x + x = 2 t i x + 2x
2
E D + E W =
dx d x
i=1 i=1
Q

= 2 t i Qx + 2x

i=1
Q

= 2 t i + Q x = 0

i=1
Solving for , we obtain

Q

t i
MP

i=1
x = --------------------
-
+ Q
2 MP
Notice that as goes to zero (variance x goes to infinity), x approach-
ML
es x . Increasing the variance of the prior density represents increased
uncertainty in our prior knowledge about x. With large prior uncertainty,
we rely on the data for our estimate of x, which leads to the maximum like-
lihood estimate.
Figure P13.3 illustrates P D x , P x and P x D for the case where
2 2
x = 2 , = 1 , Q = 1 and t 1 = 1 . Here the variance associated with the
measurement is smaller than the variance associated with our prior densi-
MP ML
ty for x , so x is closer to x = t 1 = 1 than it is to the maximum of the
prior density, which occurs at 0.
To experiment with this signal in noise example, use the MATLAB Neural
Network Design Demonstration Signal Plus Noise (nnd13spn).
13-36
Solved Problems
0.6
Px D
0.4
PD x
0.2
Px
0
MP
x
3 2 1 0 1 2 3
Figure P13.3 Prior and Posterior Density Functions
P13.3 Derive Eq. (13.23).

MP MP
To solve for and , we will take the derivatives of the log of
P D M , given in Eq. (13.17), with respect to and , and set the de-
rivatives to zero. Taking the log of Eq. (13.17), and substituting Eq. (13.12),
Eq. (13.14) and Eq. (13.22), we obtain
log P D M = log Z F log Z D log Z W

n
= --- log 2 --- logdet H F x MP ---- log --- --- log ---
n 1 MP N
2 2 2 2
1 MP N n n N
= F x MP --- log det H + ---- log + --- log + --- log 2 ---- log
2 2 2 2 2
We will consider first the second term in this expression. Since H is the
Hessian of F in Eq. (13.4), we can write it as
h
H = 2F = 2 E D + 2 E W = B + 2I , where B = 2E D . If we let
b h b
be an eigenvalue of H and be an eigenvalue of B , then = + 2
for all corresponding eigenvalues. Now we take the derivative of the second
term in the above equation with respect to . Since the determinant of a
matrix can be expressed as the product of its eigenvalues, we can reduce it
1
as shown below, where tr H is the trace of the inverse of the Hessian H .
13-37
13 Generalization
n
1 1 h
--- log det H = ----------------
2 2det H
k=1
n
1
= ---------------- i + 2
b
2det H
i=1
n

j + 2 i + 2
1 b b
= ----------------
2det H
i = 1 ji
n
j + 2
b
ji
= i---------------------------------------------
=1
n
-
i + 2
b
i=1
n
1 1
= ------------------
b
+ 2
- = tr H
i=1 i
Next, we will take the derivative of the same term with respect to . First,
define the parameter , as shown below, and expand it for use in our next
step. The parameter is referred to as the effective number of parameters.
1
n 2tr H
N n n n
i
b b
1 1 ------------------
2 i
= n 2 ------------------
- = - = ------------------
- = -----
+ 2
b b b h
i=1
i + 2 i=1 i i=1 i + 2
i=1
i
1 MP
Now take the derivative of --- log det H with respect to .
2
13-38
Solved Problems
n
1 1 h
--- log det H = ---------------- k
2 2det H
k=1
1
n

= ---------------- i + 2
b
13
2det H
i=1
n

j + 2 i + 2
1 b b
= ----------------
2det H
i = 1 ji
n
i
b
j + 2 -----
b
1 = 1 ji
= --- i--------------------------------------------------------
n
-
2
i + 2
b
i=1
n b
1 i
= ------ ------------------
b
- = ------
2 + 2 2
i=1 i
b
where the fourth step is derived from the fact that i is an eigenvalue of
b
B , and therefore the derivative of i with respect to is just the eigen-
b
value of B which is i .
Now we are finally ready to take the derivatives of all terms in
log P D M and set them equal to zero. The derivative with respect to
will be
MP 1 MP n
log P D M = F w --- logdet H + --- log
2 2
MP MP 1 n
= E W w tr H + ------------
-
2
MP
MP MP 1 n
= EW w tr H + ------------
MP
-=0
2
Rearranging terms, and using our definition of , we have
MP n MP 1
EW w = ------------
MP
- tr H
2
MP MP MP MP 1
2 EW w = n 2 tr H =
MP
= --------------------------
MP
-
2E W w
13-39
13 Generalization
We now repeat the process for .
MP 1 MP N
log P D M = F w --- logdet H + ---- log
2 2
MP N
= E D w ------------ + ------------
2
MP
2
MP
MP N
= ED w ------------
MP
+ ------------
MP
=0
2 2
Rearranging terms,
MP N
ED w = ------------
MP
------------
MP
2 2
MP N
= --------------------------
MP
2E D w
P13.4 Demonstrate that extrapolation can occur in a region that is sur-

rounded by training data.
Consider the function displayed in Figure 13.3. In that example, extrapo-
lation occurred in the upper left region of the input space, because all of the
training data was in the lower right. Lets provide training data around the
outside of the input space, but without data in the region
1.5 p 1 1.5 1.5 p 2 1.5 .
The training data is distributed as shown in Figure P13.4.
3
3 2 1 0 1 2 3
Figure P13.4 Training Data Locations
13-40
Solved Problems
The result of the training is shown in Figure P13.5. The neural network ap-
proximation significantly overestimates the true function in the region
without training data, even though surrounded by regions with training
data. In addition, this result is random. With a different set of initial ran-
dom weights, the network might underestimate the true function in this re-
gion. Extrapolation occurs because there is a significantly large region
without training data. When the input space is of high dimension, it can be
very difficult to tell when a network is extrapolating. It cannot be done by
simply checking the individual ranges of each input variable.
a) b)
8 8
6 6
4 4
2 2
t 0 a 0
2 2
4 4
6 6
3 3
2 2
3 3
1 2 1 2
0 1 0 1
0 0
p2 1
2 2
1
p1 p2 1
2 2
1
p1
3 3 3 3
Figure P13.5 Function (a) and Neural Network Approximation (b)
P13.5 Consider the example starting on page 13-23. Find the effective
number of parameters if = 1 .
To find the effective number of parameters, we can use Eq. (13.55):

n
i
= ---------------------
i + 2
.
i=1
We earlier found the eigenvalues to be 1 = 1 , 2 = 3 . The regularization

parameter is

= --- = 1 .

We can rewrite in terms of as follows

n n n
i i i
= ---------------------
i + 2
= ------------------ = ----------------
i + 2
-.
i = 1 i + 2 ---
i=1
i=1
13-41
13 Generalization
Substituting our numbers, we find

n
i 1 3 1 3 14
= ----------------
i + 2
- = ------------ + ------------ = --- + --- = ------ .
1+2 3+2 3 5 15
i=1
Therefore, we are using approximately one of the two available parame-

ters. The network has two parameters: w 1 1 and w 1 2 . The parameter we
are using is not one of these two, but rather a combination. As we can see
from Figure 13.11, we move in the direction of the second eigenvector:
v2 = 1 ,
1
which means that we are changing w 1 1 and w 1 2 by the same amount. Al-
though there are two parameters, we are effectively using only one. Since
v 2 is the eigenvector with the largest eigenvalue, we move in that direction
to obtain the greatest reduction in the squared error.
P13.6 Demonstrate overfitting with polynomials. Consider fitting a poly-

nomial
2 k
gk p = x0 + x1 p + x2 p + + xk p
to a set of data {p 1, t 1} {p 2, t 2} {p Q, tQ} so as to minimize the fol-

lowing squared error performance function.
Q
tq gk pq
2
Fx =
q=1
First, we want to express the problem in matrix form. Define the following
vectors.
k
1 p1 p1
t1 x0
k
t =
t2 G = 1 p2 p2 x =
x1

k
tQ 1 pQ pQ xk
We can then write the performance index as follows.

T T T T T T
F x = t Gx t Gx = t t 2x G t + x G Gx
13-42
Solved Problems
To locate the minimum, we take the gradient and set it equal to zero.
T T
F x = 2G t + 2G Gx = 0
Solving for the weights, we obtain the least squares solution (maximum
likelihood for the Gaussian noise case).
T ML T ML T 1 T
G G x = G t x = G G G t
To demonstrate the operation of the polynomial fitting, we will use the sim-
ple linear function t = p . To create the data set, we will sample the func-
tion at five different points and will add noise as follows
t i = p i + i , p = 1 0.5 0 0.5 1 ,
where i has a uniform density with range 0.25 0.25 . The code below
shows how to generate the data and fit a 4th order polynomial. The results
of fitting 2nd and 4th order polynomials are shown in Figure P13.6. The 4th
order polynomial has five parameters, which allow it to exactly fit the five
noisy data points, but it doesnt produce an accurate approximation of the
true function.
2+2
ans = p = -1:.5:1;
4 t = p + 0.5*(rand(size(p))-0.5);
Q = length(p);
ord = 4;
G = ones(Q,1);
for i=1:ord,
G = [G (p').^i];
end
x = (G'*G)\G'*t'; % Could also use x = G\t;
1.5
Function
Data
2nd Order
1 4th Order
0.5
0.5
1
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Figure P13.6 Polynomial Approximations to a Straight Line
13-43
13 Generalization
Epilogue
The focus of this chapter has been the development of algorithms for
training multilayer neural networks so that they generalize well. A
network that generalizes well will perform as well in new situations as it
performs on the data for which it was trained.
The basic approach to producing networks that generalize well is to find
the simplest network that can represent the data. A simple network is one
that has a small number of weights and biases.
The two methods that we presented in this chapter, early stopping and reg-
ularization, produce simple networks by constraining the weights, rather
than by reducing the number of weights. We showed in this chapter that
constraining the weights is equivalent to reducing the number of weights.
Chapter 23 presents a case study that uses Bayesian regularization to pre-
vent overfitting in a practical function approximation problem. Chapter 25
presents a case study that uses early stopping to prevent overfitting in a
practical pattern recognition problem.
13-44
Further Reading
Further Reading
[AmMu97] S. Amari, N. Murata, K.-R. Muller, M. Finke, and H. H.
Yang, Asymptotic Statistical Theory of Overtraining and
Cross-Validation, IEEE Transactions on Neural Net- 13
works, vol. 8, no. 5, 1997.
When using early stopping, it is important to decide on the
number of data points to place in the validation set. This
paper provides a theoretical basis for the choice of valida-
tion set size.
[FoHa97] D. Foresee and M. Hagan, Gauss-Newton Approximation
to Bayesian Learning, Proceedings of the 1997 Interna-
tional Joint Conference on Neural Networks, vol. 3, pp.
1930 - 1935, 1997.
This paper describes a method for implementing Bayesian
regularization by using the Gauss-Newton approximation
to the Hessian matrix.
[GoLa98] C. Goutte and J. Larsen, Adaptive Regularization of Neu-
ral Networks Using Conjugate Gradient, Proceedings of
the IEEE International Conference on Acoustics, Speech
and Signal Processing, vol. 2, pp. 1201-1204, 1998.
When using regularization, the important step is setting
the regularization parameter. This paper describes a proce-
dure for setting the regularization parameter to minimize
the validation set error.
[MacK92] D. J. C. MacKay, Bayesian Interpolation, Neural Compu-
tation, vol. 4, pp. 415-447, 1992.
Bayesian approaches have been used for many years in sta-
tistics. This paper presents one of the first developments of
a Bayesian framework for training neural networks.
MacKay followed this paper with many others describing
refinements of the approach.
[Sarle95] W. S. Sarle, Stopped training and other remedies for over-
fitting, In Proceedings of the 27th Symposium on Interface,
1995.
This is one of the early papers on the use of early stopping
with a validation set to prevent overfitting. The paper de-
scribes simulation results comparing early stopping with
other methods for improving generalization.
13-45
13 Generalization
[SjLj94] J. Sjoberg and L. Ljung, Overtraining, regularization and

searching for minimum with application to neural net-
works, Linkoping University, Sweden, Tech. Rep. LiTH-
ISY-R-1567, 1994.
This report explains how early stopping and regularization
are approximately equivalent processes. It demonstrates
that the number of iterations of training is inversely pro-
portional to the regularization parameter.
[Tikh63] A. N. Tikhonov, The solution of ill-posed problems and the
regularization method, Dokl. Acad. Nauk USSR, vol. 151,
no. 3, pp. 501-504, 1963.
Regularization is a method by which a squared error per-
formance index is augmented by a penalty on the complex-
ity of the approximating function. This is the original paper
that introduced the concept of regularization. The penalty
involved the derivatives of the approximating function.
[WaVe94] C. Wang, S. S. Venkatesh, and J. S. Judd, Optimal Stop-
ping and Effective Machine Complexity in Learning, Ad-
vances in Neural Information Processing Systems, J. D.
Cowan, G. Tesauro, and J. Alspector, Eds., vol. 6, pp. 303-
310, 1994.
This paper describes how the effective number of network
parameters changes during the training process and how
the generalization capability of the network can be im-
proved by stopping the training early.
13-46
Exercises
Exercises
E13.1 Consider fitting a polynomial (kth order)
2 k
gk p = x0 + x1 p + x2 p + + xk p
to a set of data {p 1, t 1} {p 2, t 2} {p Q, tQ} . It has been proposed that min-

imizing a performance index that penalizes the derivatives of the polyno-
mial will provide improved generalization. Investigate the relationship
between this technique and regularization using squared weights.
i. Derive the least squares solution for the weights x i , which minimiz-
es the following squared error performance index. (See Solved Prob-
lem P13.6.)
Q
tq gk pq
2
Fx =
q=1
ii. Derive the regularized least squares solution, with a squared

weight penalty.
Q k
tq gk pq + xi
2 2
Fx =
q=1 i=0
iii. Derive a solution for the weights that minimizes a sum of the
squared error plus a sum of squared derivatives.
Q Q
2
d
tq gk p q +
2
Fx = g p
dp k q
q=1 i=1
iv. Derive a solution for the weights that minimizes a sum of the
squared error plus a sum of squared second derivatives.
Q Q 2 2
d
t q g k p q +
2
F x = g p
2 k q
q=1 i=1
dp
E13.2 Write a MATLAB program to implement the solutions you found in E13.1
2+2
i. through iv. Using the following data points, adjust the values to obtain
ans = the best results. Use k = 8 for all cases. Plot the data points, the noise-free
4 function ( t = p ) and the polynomial approximation in each case. Compare
the four approximations. Which do you think produces the best results?
Which cases produce similar results?
13-47
13 Generalization
t i = p i + i , p = 1 0.5 0 0.5 1 ,
where i has a uniform density with range 0.1 0.1 (use the rand com-
mand in MATLAB).
E13.3 Consider fitting a polynomial (1st order), g 1 p = x 0 + x 1 p , to the following

data set:

p1 = 1 t1 = 4 , p2 = 2 t2 = 6 .

i. Find the least squares solutions for the weights x 0 and x 1 that min-
imize the following sum squared error performance index:
2
tq g 1 pq
2
Fx = .
q=1
ii. Find the regularized least squares solution for the weights x 0 and
x 1 when the following squared weight penalty is used:
2 1
tq g 1 p q xi .
2 2
F x = +
q=1 i=0
E13.4 Investigate the extrapolation characteristics of neural networks and poly-

nomials. Consider the problem described in E11.25, where a sine wave is
fit over the range 2 p 2 . Select 11 training points evenly spaced over
this interval.
2+2
ans = i. After fitting the 1-2-1 neural network over this range, plot the actu-
4 al sine function and the neural network approximation over the
range 4 p 4 .
ii. Fit a fifth-order polynomial (which has the same number of free pa-
rameters as the 1-2-1 network) to the sine wave over the range
2 p 2 (using your results from E13.1 i.). Plot the actual function
and the polynomial approximation over the range 4 p 4 .
iii. Discuss the extrapolation characteristics of the neural network and
the polynomial.
E13.5 Suppose that we have a random variable t that is distributed according to

the following density function. We take a series of Q independent samples
ML
of the random variable. Find the maximum likelihood estimate of x ( x ).
13-48
Exercises
f t x = ----2 exp --
t t
t0
x x
E13.6 For the random variable t given in E13.5, suppose that x is a random vari-
able with the following prior density function. Find the most probable esti-
MP
mate of x ( x ).
f x = exp x x0
E13.7 Repeat E13.6 for the following prior density function. Under what condi-
MP ML
tions will x = x ?
x x
2
1
f x = ---------------- exp --------------------
-
2 x
2
2 x
MP
E13.8 In the signal plus noise example given in Solved Problem P13.2, find x
for the following prior density functions.
i. f x = exp x x0
1
ii. f x = --- exp x
2
E13.9 Suppose that we have a random variable t that is distributed according to

the following density function. We take a series of Q = 2 independent sam-
ples of the random variable.
exp t x tx
ft x =
0 tx
i. Find the likelihood function f t 1 t 2 x , and sketch versus x .

ii. Suppose that the two measurements are t 1 = 1 and t 2 = 1 . Find
ML
the maximum likelihood estimate of x ( x ).
For the random variable t above, suppose that x is a random variable with
the following prior density function.
1 2
f x = -------------- exp 2x
2
iii. Sketch the posterior density f x t 1 t 2 . (You do not need to compute

the denominator, just find the general shape. Assume the same
measurements from part ii.)
13-49
13 Generalization
MP
iv. Find the most probable estimate of x ( x ).
E13.10 We have a coin that is not fair. (The probability of heads is not the same as
the probability of tails.) We want to estimate the probability of heads ( x ).
i. If we flip the coin 10 times, the probability of getting exactly t
heads, given that the probability of heads is x , is given below. Find
ML
the maximum likelihood estimate of x ( x ). (Hint: Take the natu-
ral log of p t x before finding the maximum.) Is it reasonable? Ex-
plain.
pt x = x 1 x , where = ----------------------
10 t 10 t 10 10!
t t t! 10 t
ii. Assume that the probability of heads, x , is a random variable with

the following prior density function. Find the most probable esti-
MP
mate of x ( x ). (Hint: Take the natural log of p t x p x before
ML MP
finding the maximum.) Explain why x is different than x .
2
p x = 12x 1 x , 0 x 1 .
E13.11 Suppose that the prior density in the level I Bayesian analysis (see page 13-
12) has nonzero mean, x . Find the new performance index.
E13.12 Suppose that we have the following inputs and targets:
2 t = 1 p = 2 t = 3
p1 = 1 2 2
1 1
We want to train a single-layer linear network without a bias on this train-

ing set. Assume that each input vector occurs with equal probability. We
will train the network using the regularized performance index of Eq.
(13.34).
MP
i. Find x and the effective number of parameters , if = 1 .
MP MP
ii. Find x and as . Explain the difference between x and
ML
x . (Your answer should be specific to this problemnot a general
MP ML
discussion of the difference between x and x .)
E13.13 Suppose we have the following input pattern and target:
2 t = 1.
p1 = 1
1
13-50
Exercises
This pattern is used to train a single layer, linear network with no bias.
i. The network is to be trained with a regularized performance index,
with the regularization parameter set to = = 1 2 . Find an
expression for the regularized performance index.
MP
ii. Find x and the effective number of parameters .
iii. Sketch the contour plot of the regularized performance index.
iv. Find the maximum stable learning rate if steepest descent is used
to train the network.
v. Find the initial direction for the steepest descent algorithm trajec-
tory, if both initial weights are set to zero.
vi. Sketch an approximate complete trajectory (on your contour plot
from part iii.) for the steepest algorithm, with very small learning
rate, from the initial conditions where both weights are set to zero.
Explain your procedure for sketching the trajectory.
E13.14 Suppose that we have a single layer, linear network with no bias. The in-
put/target pairs of the training set are given by
1 t = 2 , p = 1 t = 2 , p = 2 t = 4 ,
p1 = 1 2 2 3 3
1 2 1
where each pair occurs with equal probability. We want to minimize the
regularized performance index of Eq. (13.34).
i. Find the effective number of parameters , for = 1 .
ii. Starting with zero initial weights, approximately how many itera-
tions of the steepest descent algorithm would need to be made on the
mean square performance index E D to produce results that would
be equivalent to minimizing the regularized performance index
with = 1 ? Assume a learning rate of = 0.01 .
E13.15 Repeat E11.25, but modify your program to use early stopping and to use
2+2
30 neurons. Select 10 training points and 5 validation points. Add noise to
ans = the validation and testing points that is uniformly distributed between
4 0.1 and 0.1 (using the MATLAB function rand). Measure the mean square
error of the trained network on a testing set consisting of 20 equally-spaced
points of the noise-free function. Try 10 different random sets of training
and validation data. Compare the results with early-stopping with the re-
sults without early stopping.
13-51
13 Generalization
E13.16 Repeat E13.15, but use regularization instead of early stopping. This will
2+2
require modifying your program to compute the gradient of the regularized
ans = performance index. Add the standard gradient of the squared error, which
4 is computed by the standard backpropagation algorithm, to the gradient of
times the squared weights. Try three different values of . Compare
these results with the early stopping results.
E13.17 Consider again the problem described in E10.4

i. Find the regularized performance index for = 0 1 . Sketch the
contour plot in each case. Indicate the location of the optimal
weights in each case.
ii. Find the effective number of parameters for = 0 1 .
iii. Starting with zero initial weights, approximately how many itera-
tions of the steepest descent algorithm would need to be made on the
mean square performance index to produce results that would be
equivalent to minimizing the regularized performance index with
= 1 ? Assume a learning rate of = 0.01 .
2+2
ans = iv. Write a MATLAB M-file to implement the steepest descent algo-
4 rithm to minimize the mean square error performance index that
you found in part i. (This is a quadratic function.) Start the algo-
rithm with zero initial conditions, and use a learning rate of
= 0.01 . Sketch the trajectory on a contour plot of the mean square
error (the contour plot was found in E10.4). Verify that at the itera-
tion you computed in part iii., the weights are close to the same val-
ues you found to minimize the regularized performance index with
= 1 in part i.
13-52
Objectives
14 Dynamic Networks
Objectives 14-1
Layered Digital Dynamic Networks 14-3
Example Dynamic Networks 14-5
Principles of Dynamic Learning 14-8
Dynamic Backpropagation 14-12
Preliminary Definitions 14-12
Real Time Recurrent Learning 14-12
Backpropagation-Through-Time 14-22
Summary and Comments on Dynamic Training 14-30
Epilogue 14-45
Exercises 14-47
Objectives
Neural networks can be classified into static and dynamic categories. The
multilayer network that we have discussed in the last three chapters is a
static network. This means that the output can be calculated directly from
the input through feedforward connections. In dynamic networks, the out-
put depends not only on the current input to the network, but also on the
current or previous inputs, outputs or states of the network. For example,
the adaptive filter networks we discussed in Chapter 10 are dynamic net-
works, since the output is computed from a tapped delay line of previous
inputs. The Hopfield network we discussed in Chapter 3 is also a dynamic
network. It has recurrent (feedback) connections, which means that the
current output is a function of outputs at previous times.
We will begin this chapter with a brief introduction to the operation of dy-
namic networks, and then we will describe how these types of networks can
be trained. The training will be based on optimization algorithms that use
gradients (as in steepest descent and conjugate gradient algorithms) or Ja-
cobians (as in Gauss-Newton and Levenberg-Marquardt algorithms) These
algorithms were described in Chapters 10, 11 and 12 for static networks.
The difference between the training of static and dynamic networks is in
the manner in which the gradient or Jacobian is computed. In this chapter,
we will present methods for computing gradients for dynamic networks.
14-1
14 Dynamic Networks
Theory and Examples

Dynamic Networks Dynamic networks are networks that contain delays (or integrators, for
continuous-time networks) and that operate on a sequence of inputs. (In
other words, the ordering of the inputs is important to the operation of the
network.) These dynamic networks can have purely feedforward connec-
tions, like the adaptive filters of Chapter 10, or they can also have some
Recurrent feedback (recurrent) connections, like the Hopfield network of Chapter 3.
Dynamic networks have memory. Their response at any given time will de-
pend not only on the current input, but on the history of the input sequence.
Because dynamic networks have memory, they can be trained to learn se-
quential or time-varying patterns. Instead of approximating functions, like
the static multilayer perceptron network of Chapter 11, a dynmic network
can approximate a dynamic system. This has applications in such diverse
areas as control of dynamic systems, prediction in financial markets, chan-
nel equalization in communication systems, phase detection in power sys-
tems, sorting, fault detection, speech recognition, learning of grammars in
natural languages, and even the prediction of protein structure in genetics.
Dynamic networks can be trained using the standard optimization meth-
ods that we have discussed in Chapters 9 through 12. However, the gradi-
ents and Jacobians that are required for these methods cannot be
computed using the standard backpropagation algorithm. In this chapter
we will present the dynamic backpropagation algorithms that are required
for computing the gradients for dynamic networks.
There are two general approaches (with many variations) to gradient and
Jacobian calculations in dynamic networks: backpropagation-through-
time (BPTT) [Werb90] and real-time recurrent learning (RTRL) [WiZi89].
In the BPTT algorithm, the network response is computed for all time
points, and then the gradient is computed by starting at the last time point
and working backward in time. This algorithm is efficient for the gradient
calculation, but it is difficult to implement on-line, because the algorithm
works backward in time from the last time step.
In the RTRL algorithm, the gradient can be computed at the same time as
the network response, since it is computed by starting at the first time
point, and then working forward through time. RTRL requires more calcu-
lations than BPTT for calculating the gradient, but RTRL allows a conve-
nient framework for on-line implementation. For Jacobian calculations,
the RTRL algorithm is generally more efficient than the BPTT algorithm.
In order to more easily present general BPTT and RTRL algorithms, it will
be helpful to introduce modified notation for networks that can have recur-
rent connections. In the next section we will introduce this notation, and
then the remainder of the chapter will present general BPTT and RTRL al-
gorithms for dynamic networks.
14-2
Layered Digital Dynamic Networks

In this section we want to introduce the neural network framework that we
will use to represent general dynamic networks. We call this framework
LDDN Layered Digital Dynamic Networks (LDDN). It is an extension of the nota-
tion that we have used to represent static multilayer networks. With this
new notation, we can conveniently represent networks with multiple recur-
rent (feedback) connections and tapped delay lines.
To help us introduce the LDDN notation, consider the example dynamic
network given in Figure 14.1.
Inputs Layer 1 Layer 2 Layer 3
T
D LW1,1
1
L a (t)
1
p (t)1 n (t) 2
a (t) a3(t)
T
1,1 2,1 3,2
IW D LW LW n3(t)
n2(t)
1
R x1 S2x1 3
S1xR
1 L S2xS1
2
S3xS2
3 S x1
f 1
S x1 2
f f
S x1 S3x1
1 b 1 1
S x1 1 b 2
1 b 3
1 1 2 3
R S1x1 S S2x1 S S3x1 S
T T
D LW1,3 D LW2,3
L L
Figure 14.1 Example Dynamic Network

m
The general equations for the computation of the net input n t for layer
m of an LDDN are
m l

m l
n t = LW d a t d
l L m d DL m l
f
(14.1)
m l

l m
+ IW d p t d + b
l Im d DI m l
l m l
where p t is the l th input vector to the network at time t, IW is the
m l
Input Weight input weight between input l and layer m, LW is the layer weight be-
m
Layer Weight tween layer l and layer m, b is the bias vector for layer m, DL m l is the set
of all delays in the tapped delay line between Layer l and Layer m, DI m l is
the set of all delays in the tapped delay line between Input l and Layer m,
14-3
14 Dynamic Networks
f
I m is the set of indices of input vectors that connect to layer m, and L m is
the set of indices of layers that directly connect forward to layer m. The out-
put of layer m is then computed as
m m m
a t = f n t . (14.2)
Compare this with the static multilayer network of Eq. (11.6). LDDN net-
works can have several layers connecting to layer m. Some of the connec-
tions can be recurrent through tapped delay lines. An LDDN can also have
multiple input vectors, and the input vectors can be connected to any layer
in the network; for static multilayer networks, we assumed that the single
input vector connected only to Layer 1.
With static multilayer networks, the layers were connected to each other in
numerical order. In other words, Layer 1 was connected to Layer 2, which
was connected to Layer 3, etc. Within the LDDN framework, any layer can
connect to any other layer, even to itself. However, in order to use Eq.
(14.1), we need to compute the layer outputs in a specific order. The order
in which the layer outputs must be computed to obtain the correct network
Simulation Order output is called the simulation order. (This order need not be unique; there
may be several valid simulation orders.) In order to backpropagate the de-
rivatives for the gradient calculations, we must proceed in the opposite or-
Backpropagation Order der, which is called the backpropagation order. In Figure 14.1, the
standard numerical order, 1-2-3, is the simulation order, and the backprop-
agation order is 3-2-1.
As with the multilayer network, the fundamental unit of the LDDN is the
layer. Each layer in the LDDN is made up of five components:
1. a set of weight matrices that come into that layer (which may connect
from other layers or from external inputs),
2. any tapped delay lines (represented by DL m l or DI m l ) that appear at
the input of a set of weight matrices (Any set of weight matrices can be
preceded by a TDL. For example, Layer 1 of Figure 14.1 contains the
1 3
weights LW d and the corresponding TDL.),
3. a bias vector,
4. a summing junction, and
5. a transfer function.
The output of the LDDN is a function not only of the weights, biases, and
current network inputs, but also of some layer outputs at previous points
in time. For this reason, it is not a simple matter to calculate the gradient
of the network output with respect to the weights and biases. The weights
and biases have two different effects on the network output. The first is the
direct effect, which can be calculated using the standard backpropagation
algorithm from Chapter 11. The second is an indirect effect, since some of
14-4
the inputs to the network are previous outputs, which are also functions of
the weights and biases. The main development of the next two sections is
a general gradient calculation for arbitrary LDDNs.
Example Dynamic Networks

2 Before we introduce dynamic training, lets get a feeling for the types of re-
+2
sponses we can expect to see from dynamic networks. Consider first the
feedforward dynamic network shown in Figure 14.2.
Inputs Linear Neuron
p(t) iw1,1(0)
D iw1,1(1) n(t) a(t)

S
D
iw1,1(2)
a(t) = iw1,1(0)p(t)+iw1,1(1)p(t-1)+iw1,1(2)p(t-2)
Figure 14.2 Example Feedforward Dynamic Network

This is an ADALINE filter, which we discussed in Chapter 10 (see Figure
10.5). Here we are representing it in the LDDN framework. The network
has a TDL on the input, with DI 1 1 = 0 1 2 . To demonstrate the opera-
tion of this network, we will apply a square wave as input, and we will set
all of the weight values equal to 1/3:
1 1 1
iw 1 1 0 = --- , iw 1 1 1 = --- , iw 1 1 2 = --- . (14.3)
3 3 3
The network response is calculated from:

2
at = n t = IW d p t d (14.4)
d=0
= n 1 t = iw 1 1 0 p t + iw 1 1 1 p t 1 + iw 1 1 2 p t 2
where we have left off the superscripts on the weight and the input, since
there is only one input and only one layer.
14-5
14 Dynamic Networks
The response of the network is shown in Figure 14.3. The open circles rep-
resent the square-wave input signal p t . The dots represent the network
response a t . For this dynamic network, the response at any time point
depends on the previous three input values. If the input is constant, the
output will become constant after three time steps. This type of linear net-
FIR work is called a Finite Impulse Response (FIR) filter.
1.5
0.5
0.5
1.5
0 5 10 15 20 25
Figure 14.3 Response of ADALINE Filter Network

This dynamic network has memory. Its response at any given time will de-
pend not only on the current input, but on the history of the input sequence.
If the network does not have any feedback connections, then only a finite
amount of history will affect the response. In the next example, we will con-
sider a network that has an infinite memory.
To experiment with the finite impulse response example, use the Neural Net-
work Design Demonstration Finite Impulse Response Network (nnd14fir).
2 Now consider another simple linear dynamic network, but one that has a
+2
recurrent connection. The network in Figure 14.4 is a recurrent dynamic
network. The equation of operation of the network is
1 1 1 1 1 1 1 1
a t = n t = LW 1 a t 1 + IW 0 p t
(14.5)
= lw 1 1 1 a t 1 + iw 1 1 p t
where, in the last line, we have left off the superscripts, since there is only
one neuron and one layer in the network. To demonstrate the operation of
this network, we will set the weight values to
1 1
lw 1 1 1 = --- and iw 1 1 = --- . (14.6)
2 2
14-6
Inputs Linear Neuron
iw1,1 n(t) a(t)

p(t)
S
lw1,1(1)
D
a(t) = iw1,1 p(t) + lw1,1(1) a(t-1)
Figure 14.4 Recurrent Linear Neuron

The response of this network to the square wave input is shown in Figure
14.5. The network responds exponentially to a change in the input se-
quence. Unlike the FIR filter network of Figure 14.2, the exact response of
the network at any given time is a function of the infinite history of inputs
to the network.
1.5
0.5
0.5
1.5
0 5 10 15 20 25
Figure 14.5 Recurrent Neuron Response

To experiment with this infinite impulse response example, use the Neural
Network Design Demonstration Infinite Impulse Response Network
(nnd14iir).
Compare the dynamic networks of the previous two examples with the stat-
ic, two-layer perceptron of Figure 11.4. Static networks can be trained to
approximate static functions, like sin p , where the output can be comput-
ed directly from the current input. Dynamic networks, on the other hand,
14-7
14 Dynamic Networks
can be trained to approximate dynamic systems, such as robot arms, air-

craft, biological processes and economic systems, where the current system
output depends on a history of previous inputs and outputs. Because dy-
namic systems are more complex than static functions, we expect that the
training process for dynamic networks will be more challenging than static
network training.
In the following section, we will discuss the computation of gradients for
the training of dynamic networks. For static networks, these gradients
were computed using the standard backpropagation algorithm. For dy-
namic networks, the backpropagation algorithm must be modified.
Principles of Dynamic Learning

2 Before we get into the details of training dynamic networks, lets first in-
+2
vestigate a simple example. Consider again the recurrent network of Fig-
ure 14.4. Suppose that we want to train the network using steepest
descent. The first step is to compute the gradient of the performance func-
tion. For this example we will use sum squared error:
Q Q
e tt a t
2 2
Fx = t = . (14.7)
t=1 t=1
The two elements of the gradient will be

Q Q
2
F x e t a t
------------------------ =
lw 1 1 1 ------------------------ = 2 e t ------------------------ ,
lw 1 1 1 lw 1 1 1
(14.8)
t=1 t=1
Q Q
2
F x e t a t
--------------- =
iw 1 1 --------------- = 2 e t ---------------
iw 1 1 iw 1 1
(14.9)
t=1 t=1
The key terms in these equations are the derivatives of the network output
with respect to the weights:
a t a t
------------------------ and --------------- . (14.10)
lw 1 1 1 iw 1 1
If we had a static network, then these terms would be very easy to compute.
They would correspond to a t 1 and p t , respectively. However, for re-
current networks, the weights have two effects on the network output. The
first is the direct effect, which is also seen in the corresponding static net-
work. The second is an indirect effect, caused by the fact that one of the net-
work inputs is a previous network output. Lets compute the derivatives of
the network output, in order to demonstrate these two effects.
14-8
The equation of operation of the network is
a t = lw 1 1 1 a t 1 + iw 1 1 p t . (14.11)
We can compute the terms in Eq. (14.10) by taking the derivatives of Eq.
(14.11):
a t a t 1
------------------------ = a t 1 + lw 1 1 1 ------------------------ , (14.12)
lw 1 1 1 lw 1 1 1
a t a t 1
--------------- = p t + lw 1 1 1 ---------------------- . (14.13)
iw 1 1 iw 1 1
The first term in each of these equations represents the direct effect that
each weight has on the network output. The second term represents the in-
direct effect. Note that unlike the gradient computation for static networks,
the derivative at each time point depends on the derivative at previous
time points (or at future time points, as we will see later).
The following figures illustrate the dynamic derivatives. In Figure 14.6 a)
we see the total derivatives a t iw 1 1 and also the static portions of the
derivatives. Note that if we consider only the static portion, we will under-
estimate the effect of a change in the weight. In Figure 14.6 b) we see the
original response of the network (which was also shown in Figure 14.5) and
a new response, in which iw 1 1 is increased from 0.5 to 0.6. By comparing
the two parts of Figure 14.6, we can see how the derivative indicates the
effect on the network response of a change in the weight iw 1 1 .
a) 2 b) 1.5
Total Derivative
1.5
Static Derivative
1
0.5 Original Response

0.5 Incremental Response
0 0
0.5
0.5
1
1.5
2 1.5
0 5 10 15 20 25 0 5 10 15 20 25
Figure 14.6 Derivatives for iw 1 1 and Response of Network in Figure 14.4
In Figure 14.7 we see similar results for the weight lw 1 1 1 . The key ideas
to get from this example are: 1) the derivatives have static and dynamic
components, and 2) the dynamic component depends on other time points.
14-9
14 Dynamic Networks
To experiment with dynamic derivatives, use the Neural Network Design

Demonstration Dynamic Derivatives (nnd14dynd).
a) 2 b) 1.5
1.5
1
Original Response
0.5
Incremental Response
0.5
0 0
0.5
0.5
1
1.5 Total Derivative
Static Derivative
2 1.5
0 5 10 15 20 25 0 5 10 15 20 25
Figure 14.7 Derivatives for lw 1 1 1 and Response of Network in Figure 14.4
Having made this initial investigation of a single-neuron network, lets con-

sider the slightly more complex dynamic network that is shown in Figure
14.8. It consists of a static multilayer network with a single feedback loop
added from the output of the network to the input of the network through
a single delay. In this figure, the vector x represents all of the network pa-
rameters (weights and biases), and the vector a t represents the output of
the multilayer network at time step t. This network will help us demon-
strate the key steps of dynamic training.
p(t)
a(t)
Multilayer Network
a(t) = NN(p(t),a(t-1),x)
D
Figure 14.8 Simple Dynamic Network

As with a standard multilayer network, we want to adjust the weights and
biases of the network to minimize the performance index, F x , which is
normally chosen to be the mean squared error. In Chapter 11, we derived
the backpropagation algorithm for computing the gradient of F x , which
we could then use with any of the optimization methods from Chapter 12
to minimize F x . With dynamic networks, we need to modify the standard
backpropagation algorithm. There are two different approaches to this
problem. They both use the chain rule, but are implemented in different
ways:
14-10
Q
e
F a t T F
------ =
x ------------
x
T
------------ ,
a t
(14.14)
t=1
or
Q T
e
F at F
------ =
x --------------
x
T
- ------------ .
a t
(14.15)
t=1
where the superscript e indicates an explicit derivative, not accounting for

indirect effects through time. The explicit derivatives can be obtained with
the standard backpropagation algorithm of Chapter 11. To find the com-
plete derivatives that are required in Eq. (14.14) and Eq. (14.15), we need
the additional equations:
e e
a t at at a t 1
------------
T
= --------------
T T
- ---------------------
- + ----------------------- T
(14.16)
x x a t 1 x
and
e e
F F at + 1 F
- ---------------------- .
------------ = ------------ + ----------------------- (14.17)
a t a t T
a t a t + 1
Eq. (14.14) and Eq. (14.16) make up the real-time recurrent learning
RTRL (RTRL) algorithm. Note that the key term is
a t
------------
T
, (14.18)
x
which must be propagated forward through time. Eq. (14.15) and Eq.
BPTT (14.17) make up the backpropagation-through-time (BPTT) algorithm.
Here the key term is
F
------------ (14.19)
a t
which must be propagated backward through time.

In general, the RTRL algorithm requires somewhat more computation
than the BPTT algorithm to compute the gradient. However, the BPTT al-
gorithm cannot be conveniently implemented in real time, since the out-
puts must be computed for all time steps, and then the derivatives must be
backpropagated back to the initial time point. The RTRL algorithm is well
suited for real time implementation, since the derivatives can be calculated
at each time step. (For Jacobian calculations, which are needed for Leven-
berg-Marquardt algorithms, the RTRL algorithm is often more efficient
than the BPTT algorithm. See [DeHa07].)
14-11
14 Dynamic Networks
Dynamic Backpropagation
In this section, we will develop general RTRL and BPTT algorithms for dy-
namic networks represented in the LDDN framework. This development
will involve generalizing Eq. (14.14) through Eq. (14.17).
Preliminary Definitions
In order to simplify the description of the training algorithm, some layers
of the LDDN will be assigned as network outputs, and some will be as-
Input Layer signed as network inputs. A layer is an input layer if it has an input weight,
or if it contains any delays with any of its weight matrices. A layer is an
Output Layer output layer if its output will be compared to a target during training, or if
it is connected to an input layer through a matrix that has any delays as-
sociated with it.
For example, the LDDN shown in Figure 14.1 has two output layers (1 and
3) and two input layers (1 and 2). For this network the simulation order is
1-2-3, and the backpropagation order is 3-2-1. As an aid in later deriva-
tions, we will define U as the set of all output layer numbers and X as the
set of all input layer numbers. For the LDDN in Figure 14.1, U={1,3} and
X={1,2}.
The general equations for simulating an arbitrary LDDN network are giv-
en in Eq. (14.1) and Eq. (14.2). At each time point, these equations are it-
erated forward through the layers, as m is incremented through the
simulation order. Time is then incremented from t = 1 to t = Q .
Real Time Recurrent Learning

In this subsection we will generalize the RTRL algorithm, given in Eq.
(14.14) and Eq. (14.16), for LDDN networks. This development will follow
in many respects the development of the backpropagation algorithm for
static multilayer networks in Chapter 11. You may want to quickly review
that material before proceeding.
Eq. (14.14)
The first step in developing the RTRL algorithm is to generalize Eq.
(14.14). For the general LDDN network, we can calculate the terms of the
gradient by using the chain rule, as in
Q T
u e
F a t F
------ =
x --------------
x
T
- --------------
u
a t
- . (14.20)
t = 1u U
If we compare this equation with Eq. (14.14), we notice that in addition to

each time step, we also have a term in the sum for each output layer. How-
ever, if the performance index F x is not explicitly a function of a specific
u
output a t , then that explicit derivative will be zero.
14-12
Eq. (14.16)
The next step of the development of the RTRL algorithm is the generaliza-
tion of Eq. (14.16). Again, we use the chain rule:
u e u e u e x u'
a t a t a t n t a t d
--------------
T
-+
- = ----------------
T -----------------T --------------------------T- ------------------------
T
- . (14.21)
x x x
u' U x X d DL x u' n t
u'
a t d x
In Eq. (14.16) we only had one delay in the system. Now we need to account
for each output and also for the number of times each output is delayed be-
fore it is input to another layer. That is the reason for the first two summa-
tions in Eq. (14.21). These equations must be updated forward in time, as
t is varied from 1 to Q. The terms
u
a t
--------------
T
- (14.22)
x
are generally set to zero for t 0 .

To implement Eq. (14.21), we need to compute the terms
e u e x
a t n t
-----------------T --------------------------T- . (14.23)
x u'
n t a t d
To find the second term on the right, we can use

l
S
x l
lw k i d' ai t d'
x l
nk t =
l L x d' DL x l i = 1
f
(14.24)
l
R
x l
iwk i d' p i t d'
l x
+ + bk
l Ix d' DI x l i = 1
(Compare with Eq. (11.20).) We can now write

e x
nk t x u'
------------------------
u'
= lw k j d . (14.25)
a j t d
If we define the following sensitivity term

e u
u m ak t
s k i t ----------------
m
-, (14.26)
n i t
which can be used to make up the following matrix
14-13
14 Dynamic Networks
u m u m u m
s 1 1 t s 1 2 t s 1 Sm t
e u u m u m u m
u m a t s 2 1 t s 2 2 t s 1 Sm t
S t = ------------------T- = , (14.27)
m
n t

u m u m u m
s Su 1 t s Su 2 t s S u Sm t
then we can write Eq. (14.23) as

x
S
e u e x
a t n t u x x u'
-----------------T --------------------------T-
x u'
= si k t + d lwk j d , (14.28)
n t a t d i j k=1
or in matrix form
e u e x
a t n t u x x u'
-----------------T --------------------------T- = S t LW d . (14.29)
x u'
n t a t d
Therefore Eq. (14.21) can be written

u e u u'
a t a t u x x u' a t d
--------------
T
-+
- = ----------------
T S t LW d ------------------------
T
- (14.30)
x x u' U x X d DL x u' x
Many of the terms in the summation on the right hand side of Eq. (14.30)
will be zero and will not have to be computed. To take advantage of these
efficiencies, we introduce some indicator sets. They are sets that tell us for
which layers the weights and the sensitivities are nonzero.
The first type of indicator set contains all of the output layers that connect
to a specified layer x (which will always be an input layer) with at least
some nonzero delay:
U x u
E LW x = u U LW d 0 d 0 , (14.31)
where means such that, and means there exists.

The second type of indicator set contains the input layers that have a non-
zero sensitivity with a specified layer u :
X u x
E S u = x X S 0 . (14.32)
u x
(When S is nonzero, there is a static connection from layer x to ouput
layer u .) The third type of indicator set contains the layers that have a non-
zero sensitivity with a specified layer u :
u x
E S u = x S 0 . (14.33)
14-14
X X
The difference between E S u and E S u is that E S u contains only input
layers. E S u will not be needed in the simplification of Eq. (14.30), but it
will be used for the calculation of sensitivities in Eq. (14.38).
Using Eq. (14.31) and Eq. (14.32), we can rearrange the order of the sum-
mations in Eq. (14.30) and sum only over nonzero terms:
u e u
a t a t
--------------
T
- = ----------------
T
-
x x
u' . (14.34)
u x x u' a t d
+ S t LW d ------------------------
x
T
-
X d DL x u'
x ES u u'
U
E LW x
Eq. (14.34) makes up the generalization of Eq. (14.16) for the LDDN net-
u m
work. It remains to compute the sensitivity matrices S t and the explic-
e u
it derivatives a t w , which are described in the next two subsections.
Sensitivities
In order to compute the elements of the sensitivity matrix, we use a form
of standard static backpropagation. The sensitivities at the outputs of the
network can be computed as
u u
u u
e u
ak t f n i t for i = k
s k i t = ----------------
- = , uU, (14.35)
u
n i t 0 for i k
or, in matrix form,

u u
t = F n t ,
u u
S (14.36)
where F n t is defined as
u u
f n 1 t
u u
0 0
u u
F n t =
u u 0 f n2 t 0 (14.37)

f n Su t
u u
0 0
u m
(see also Eq. (11.34)). The matrices S t can be computed by backpropa-
gating through the network, from each network output, using
u m u l l m
0 F n t , u U ,
m

m
S t = S t LW (14.38)
b
l ES u Lm
14-15
14 Dynamic Networks
where m is decremented from u through the backpropagation order, and

b
L m is the set of indices of layers that are directly connected backwards to
layer m (or to which layer m connects forward) and that contain no delays
in the connection. The backpropagation step given in Eq. (14.38) is essen-
tially the same as that given in Eq. (11.45), but it is generalized to allow for
arbitrary connections between layers.
Explicit Derivatives
We also need to compute the explicit derivatives
e u
a t
-.
----------------
T
(14.39)
x
Using the chain rule of calculus, we can derive the following expansion of
Eq. (14.39) for input weights:
e u e u e m
ak t ak t ni t u m l
----------------------
m l m
- ----------------------
- = ---------------- m l
- = s k i t p j t d . (14.40)
iw i j d n i t iw i j d
In vector form we can write

e u
a t u m l
- = si t pj t d .
----------------------
m l
(14.41)
iw i j d
In matrix form we have

e u
a t l T u m
------------------------------------------T = p t d S t , (14.42)
m l
vec IW d
and in a similar way we can derive the derivatives for layer weights and
biases:
e u
a t l T u m
-------------------------------------------T- = a t d S t , (14.43)
m l
vec LW d
e u
a t u m
----------------T- = S t , (14.44)
m
b
where the vec operator transforms a matrix into a vector by stacking the
columns of the matrix one underneath the other, and A B is the Kroneck-
er product of A and B [MaNe99].
The total RTRL algorithm for the LDDN network is summarized in the fol-
lowing pseudo code.
14-16
Initialize: Real-Time Recurrent Learning Gradient

u
a t
- = 0 t 0 , for all u U ,
--------------
T
x
For t = 1 to Q,
X
U = , E S u = and E S u = for all u U .
For m decremented through the BP order
b
For all u U , if E S u L m
u m u l l m
0 F n t
m
14

m
S t = S t LW
b
l ES u Lm
add m to the set E S u
X
if m X , add m to the set E S u
EndFor u
If m U
m m
S t = F n t
m m
add m to the sets U and E S m

X
if m X , add m to the set E S m
EndIf m
EndFor m
For u U incremented through the simulation order
For all weights and biases (x is a vector containing all weights and biases)
e u
a t l T u m
------------------------------------------T = p t d S t
m l
vec IW d
e u
a t l T u m
-------------------------------------------T- = a t d S t
m l
vec LW d
e u
a t u m
----------------T- = S t
m
b
EndFor weights and biases
u e u u'
--------------
x
T
- = ----------------
x
T
-+ S t LW d ------------------------
x
T
-
X d DL x u'
x ES u U
u' E LW x
EndFor u
EndFor t
Compute Gradients
Q T
u e
F a t F
------ =
x --------------
x
T
- --------------
u
a t
-
t = 1u U
14-17
14 Dynamic Networks
Example RTRL Implementations (FIR and IIR)

2 To demonstrate the RTRL algorithm, consider again the feedforward dy-
+2
namic network of Figure 14.2. The equation of operation of this network is
a t = n t = iw 1 1 0 p t + iw 1 1 1 p t 1 + iw 1 1 2 p t 2 .
The architecture of the network is defined by

f U
U = 1 , X = 1 , I 1 = 1 , DI 1 1 = 0 1 2 , L 1 = , E LW 1 = .
We will choose the following standard performance function with three

time points:
Q 3
t t a t e
2 2 2 2 2
F = = t = e 1 + e 2 + e 3 ,
t=1 t=1
with the following inputs and targets:
p 1 t 1 p 2 t 2 p 3 t 3 .
The RTRL algorithm begins with some initialization:

X
U = , E S 1 = , E S 1 = .
In addition, the initial conditions for the delays, p 0 p 1 , must be pro-

vided.
The network response is then computed for the first time step:
a 1 = n 1 = iw 1 1 0 p 1 + iw 1 1 1 p 0 + iw 1 1 2 p 1
Because the RTRL algorithm proceeds forward through time, we can im-
mediately compute the derivatives for the first time step. We will see in the
next section that the BPTT algorithm, which proceeds backward through
time, will require that we proceed through all of the time points before we
can compute the derivatives.
From the preceding pseudo-code, the first step in the derivative calculation
will be
1 1
1 = F n 1 = 1 ,
1 1
S
since the transfer function is linear. We also update the following sets:
X
ES 1 = 1 , ES 1 = 1 .
14-18
The next step is the computation of the explicit derivatives from Eq.
(14.42):
e 1 e
a 1 a1 1 T 1 1
------------------------------------------T = ----------------------- = p 1 S t = p 1 ,
vec IW 0
1 1 iw 1 1 0
e 1 e
a 1 a1 1 T 1 1
------------------------------------------T = ----------------------- = p 0 S t = p 0 ,
vec IW 1
1 1 iw 1 1 1
e 1 e
a 1 a1 1 T 1 1
------------------------------------------T = ----------------------- = p 1 S t = p 1 .
vec IW 2
1 1 iw 1 1 2
The next step would be to compute the total derivative, using Eq. (14.34).
U
However, since E LW 1 = , the total derivatives are equal to the explicit
derivatives.
All of the above steps would be repeated for each time point, and then the
final step is to compute the derivatives of the performance index with re-
spect to the weights, using Eq. (14.20):
Q T 3 T
u e 1 e
F a t F a t F
------ =
x --------------
x
T
- --------------
u
a t
- = --------------
x
T
- --------------
1
a t
- .
t = 1u U t=1
If we break this down for each weight, we have
F
----------------------- = p 1 2e 1 + p 2 2e 2 + p 3 2e 3 ,
iw 1 1 0
F
----------------------- = p 0 2e 1 + p 1 2e 2 + p 2 2e 3 ,
iw 1 1 1
F
----------------------- = p 1 2e 1 + p 0 2e 2 + p 1 2e 3 .
iw 1 1 2
We can then use this gradient in any of our standard optimization algo-
rithms from Chapters 9 and 12. Note that if we use steepest descent, this
result is a batch form of the LMS algorithm (see Eq. (10.33)).
2 Lets now do an example using a recurrent neural network. Consider again
+2
the simple recurrent network in Figure 14.4. From Eq. (14.5), the equation
of operation of this network is
a t = lw 1 1 1 a t 1 + iw 1 1 p t
The architecture of the network is defined by
14-19
14 Dynamic Networks
U = 1 , X = 1 , I 1 = 1 , DI 1 1 = 0 ,
f U
DL 1 1 = 1 , L 1 = 1 , E LW 1 = 1 .
We will choose the same performance function as the previous example:

Q 3
e
2 2 2 2 2
F = t t a t = t = e 1 + e 2 + e 3 ,
D t=1 t=1
with the following inputs and targets:
p 1 t 1 p 2 t 2 p 3 t 3 .
We initialize with
X
U = , E S 1 = , E S 1 = .
In addition, the initial condition for the delay, a 0 , and the initial deriva-
tives
a 0 a 0
--------------- and -----------------------
iw 1 1 lw 1 1 1
must be provided. (The initial derivatives are usually set to zero.)

The network response is then computed for the first time step:
a 1 = lw 1 1 1 a 0 + iw 1 1 p 1
The derivative calculation begins with

1 1
1 = F n 1 = 1 ,
1 1
S
X
ES 1 = 1 , ES 1 = 1 .
The next step is the computation of the explicit derivatives:

e 1 e
a 1 a1 1 T 1 1
------------------------------------------T = ---------------- = p 1 S 1 = p 1 ,
vec IW 0
1 1 iw 1 1
e 1 e
a 1 a1 1 T 1 1
-------------------------------------------T- = ----------------------- = a 0 S 1 = a 0 .
vec LW 1
1 1 lw 1 1 1
14-20
The next step is to compute the total derivative, using Eq. (14.34):
1 e 1 1
a t a t 1 1 1 1 a t 1
--------------
T
- + S t LW 1 -----------------------
- = ----------------
T T
-. (14.45)
x x x
Replicating this formula for each of our weights for this network, for t = 1 ,
we have
a 1 a 0
--------------- = p 1 + lw 1 1 1 --------------- = p 1 ,
iw 1 1 iw 1 1
a 1 a 0
----------------------- = a 0 + lw 1 1 1 ----------------------- = a 0 .
lw 1 1 1 lw 1 1 1
Note that unlike the corresponding equation in the previous example, these
equations are recursive. The derivative at the current time depends on the
derivative at the previous time. (Note that the two initial derivatives on the
right side of this equation would normally be set to zero, but at the next
time step they would be nonzero.) As we mentioned earlier, the weights in
a recurrent network have two different effects on the network output. The
first is the direct effect, which is represented by the explicit derivative in
Eq. (14.45). The second is an indirect effect, since one of the inputs to the
network is a previous output, which is also a function of the weights. This
effect causes the second term in Eq. (14.45).
All of the above steps would be repeated for each time point:
e e
a2 a2
---------------- = p 2 , ----------------------- = a 1 ,
iw 1 1 lw 1 1 1
a 2 a 1
--------------- = p 2 + lw 1 1 1 --------------- = p 2 + lw 1 1 1 p 1 ,
iw 1 1 iw 1 1
a 2 a 1
----------------------- = a 1 + lw 1 1 1 ----------------------- = a 1 + lw 1 1 1 a 0 ,
lw 1 1 1 lw 1 1 1
e e
a3 a3
---------------- = p 3 , ----------------------- = a 2 ,
iw 1 1 lw 1 1 1
a 3 a 2 2
--------------- = p 3 + lw 1 1 1 --------------- = p 3 + lw 1 1 1 p 2 + lw 1 1 1 p 1 ,
iw 1 1 iw 1 1
a 3 a 2 2
----------------------- = a 2 + lw 1 1 1 ----------------------- = a 2 + lw 1 1 1 a 1 + lw 1 1 1 a 0 .
lw 1 1 1 lw 1 1 1
14-21
14 Dynamic Networks
The final step is to compute the derivatives of the performance index with
respect to the weights, using Eq. (14.20):
Q T 3 T
u e 1 e
F a t F a t F
------ =
x --------------
x
T
- --------------
u
a t
- = --------------
x
T
- --------------
1
a t
- .
t = 1u U t=1
If we break this down for each weight, we have
F a 1 a 2 a 3
--------------- = --------------- 2e 1 + --------------- 2e 2 + --------------- 2e 3
iw 1 1 iw 1 1 iw 1 1 iw 1 1
= 2e 1 p 1 2e 2 p 2 + lw 1 1 1 p 1
2
2e 3 p 3 + lw 1 1 1 p 2 + lw 1 1 1 p 1
F a 1 a 2 a 3
----------------------- = ----------------------- 2e 1 + ----------------------- 2e 2 + ----------------------- 2e 3
lw 1 1 1 lw 1 1 1 lw 1 1 1 lw 1 1 1
= 2e 1 a 0 2e 2 a 1 + lw 1 1 1 a 0
2
2e 3 a 2 + lw 1 1 1 a 1 + lw 1 1 1 a 0
The expansions that we show in the final two lines of the above equations
(and also in some of the previous equations) would not be necessary in prac-
tice, since the results would be numerical. We have included them here so
that we can compare this result with the BPTT algorithm, which we
present next.
Backpropagation-Through-Time
In this section we will generalize the Backpropagation-Through-Time
(BPTT) algorithm, given in Eq. (14.15) and Eq. (14.17), for LDDN net-
works.
Eq. (14.15)
The first step is to generalize Eq. (14.15). For the general LDDN network,
we can calculate the terms of the gradient by using the chain rule, as in
u
Q S e u e m
F F ak t ni t
----------------------
m l
lw i j d
- = --------------
u
- ----------------
m
- ----------------------
m l
- (14.46)
t=1 u U k = 1 a k t n i t lw i j d
(for the layer weights), where u is an output layer, U is the set of all output
u
layers, and S is the number of neurons in layer u .
From Eq. (14.24) we can write
14-22
e m
ni t l
- = aj t d .
----------------------
m l
(14.47)
lw i j d
We will also define

u
S e u
F ak t

m
di t = - ----------------
--------------
u m
-. (14.48)
u Uk = 1 a k t n i t
The terms of the gradient for the layer weights can then be written
Q
F
d i t aj t d ,
m l
----------------------
m l
- = (14.49)
lw i j d
t=1
If we use the sensitivity term defined in Eq. (14.26),

e u
u m ak t
s k i t ----------------
m
-, (14.50)
n i t
m
then the elements d i t can be written
u
S
F u m
--------------
m
di t = u
- s k i t . (14.51)
a t
u Uk = 1 k
In matrix form this becomes
u m T F

m
d t = S t --------------
u
- (14.52)
uU a t
where
T
F F F F
- = --------------
-------------- u u
- ----------------
- -------------- u (14.53)
u
a t a 1 t a 2 t a S u t
Now the gradient can be written in matrix form.

Q
F T
d
m l
---------------------------
m l
- = t a t d , (14.54)
LW d t=1
and by similar steps we can find the derivatives for the biases and input
weights:
14-23
14 Dynamic Networks
Q
F T
d
m l
-------------------------
m l
- = t p t d , (14.55)
IW d t=1
Q
F
d
m
--------m- = t . (14.56)
b t=1
Eq. (14.54) through Eq. (14.56) make up the generalization of Eq. (14.15)
for the LDDN network.
Eq. (14.17)
The next step in the development of the BPTT algorithm is the generaliza-
tion of Eq. (14.17). Again, we use the chain rule:
e
F F
--------------
u
- = --------------
u
-
a t a t
e u' e x T (14.57)
a t + d n t + d F
+ --------------------------T- --------------------------
x u T
- ------------------------
u'
a t + d
-
u' U x X d DL x u n t + d a t
(Many of the terms in these summations will be zero. We will provide a

more efficient representation later in this section.) In Eq. (14.17) we only
had one delay in the system. Now we need to account for each network out-
put, how that network output is connected back through a network input,
and also for the number of times each network output is delayed before it
is applied to a network input. That is the reason for the three summations
in Eq. (14.57). This equation must be updated backward in time, as t is var-
ied from Q to 1. The terms
F
---------------
u'
(14.58)
a t
are generally set to zero for t > Q.

If we consider the matrix in the brackets on the right side of Eq. (14.57),
from Eq. (14.29) we can write
e u' e x
a t + d n t + d u' x x u
T
- = S t + d LW d .
--------------------------T- -------------------------- (14.59)
x u
n t + d a t
This allows us to write Eq. (14.57) as
14-24
e
F F
--------------
u
- = --------------
u
-
a t a t
(14.60)
u' x x u T F
+ S t + d LW d ------------------------
u'
a t + d
-
u' U x X d DL x u
Many of the terms in the summation on the right hand side of Eq. (14.60)
will be zero and will not have to be computed. In order to provide a more
efficient implementation of Eq. (14.60), we define the following indicator
sets:
X x u
E LW u = x X LW d 0 d 0 , (14.61)
U u x
E S x = u U S 0 . (14.62)
The first set contains all of the input layers that have a connection from
output layer u with at least some nonzero delay. The second set contains
output layers that have a nonzero sensitivity with input layer x . When the
u x
sensitivity S is nonzero, there is a static connection from input layer x
to output layer u .
We can now rearrange the order of the summation in Eq. (14.60) and sum
only over the existing terms:
e
F F
--------------
u
- = --------------
u
-
a t a t
(14.63)
x u T u' x T F
+ LW d S t + d ------------------------
u'
a t + d
-
x E LW u d DL x u
X U
u' E S x
Summary
The total BPTT algorithm is summarized in the following pseudo code.
14-25
14 Dynamic Networks
Initialize: Backpropagation-Through-Time Gradient

F
- = 0 t Q , for all u U ,
--------------
u
a t
For t = Q to 1,
U
U = , E S u = , and E S u = for all u U .
b
D
u m u l l m
0 F n t
m

m
S t = S t LW
b
l ES u L m

U
add u to the set E S m
EndFor u
If m U
m m
S t = F n t
m m
U
add m to the sets U , E S m and E S m
EndIf m
EndFor m
For u U decremented through the BP order
e
F F x u T u' x T F
--------------
u
a t
- = --------------
u
a t
-+ LW d S t + d ------------------------
u'
a t + d
-
x E LW u d DL x u
X U
u' E S x
EndFor u
For all layers m
u m T F

m
d t = S t --------------
u
-
U a t
u ES m
EndFor m
EndFor t
Compute Gradients
Q
F T
d
m l
---------------------------
m l
- = t a t d
LW d t=1
Q
F T
d
m l
-------------------------
m l
- = t p t d
IW d
t=1
Q
F
d
m
--------m- = t
b t=1
14-26
Example BPTT Implementations (FIR and IIR)

2 To demonstrate the BPTT algorithm, we will use the same example net-
+2
works that we used for the RTRL algorithm. First, we use the feedforward
dynamic network of Figure 14.2. We defined the network architecture on
page 14-18.
Before the gradient can be computed using BPTT, the network response
must be computed for all time steps:
a 1 = n 1 = iw 1 1 0 p 1 + iw 1 1 1 p 0 + iw 1 1 2 p 1 ,
a 2 = n 2 = iw 1 1 0 p 2 + iw 1 1 1 p 1 + iw 1 1 2 p 0 ,
a 3 = n 3 = iw 1 1 0 p 3 + iw 1 1 2 p 0 + iw 1 1 2 p 1 .
The BPTT algorithm begins with some initialization:

U
U = , E S 1 = , E S 1 = .
The first step in the derivative calculation will be the sensitivity calcula-
tion. For BPTT, we start at the last time point ( t = 3 ):
1 1
3 = F n 3 = 1 ,
1 1
S
U
ES 1 = 1 , ES 1 = 1 .
The next step is the calculation of the following derivative using Eq.
(14.63):
e
F F
---------------
1
- = 2e 3 .
- = ---------------
1
a 3 a 3
The final step for t = 3 is Eq. (14.52):
1 1 1 T F
d 3 = S 3 ---------------
1
- = 2e 3 .
a 3
We repeat the previous steps for t = 2 and t = 1 , to obtain
1 1 1 T F
d 2 = S 2 ---------------
1
- = 2e 2 ,
a 2
1 1 1 T F
d 1 = S 1 ---------------
1
- = 2e 1 .
a 1
14-27
14 Dynamic Networks
Now, all time steps are combined in Eq. (14.55):

3 3
F F T
d 2e t p t ,
1 1
- = ----------------------- =
------------------------- t p t =
IW 0
1 1 iw 1 1 0
t=1 t=1
3 3
F F T
d 2e t p t 1 ,
1 1
- = ----------------------- =
------------------------- t p t 1 =
IW 1
1 1 iw 1 1 1
t=1 t=1
3 3
F F T
d 2e t p t 2 .
1 1
- = ----------------------- =
------------------------- t p t 2 =
IW 2
1 1 iw 1 1 2
t=1 t=1
Note that this is the same result we obtained for the RTRL algorithm ex-
ample on page 14-19. RTRL and BPTT should always produce the same
gradient. The only difference is in the implementation.
2 Lets now use our previous recurrent neural network example of Figure
+2
14.4. We defined the architecture of this network on page 14-19.
Unlike the RTRL algorithm, where initial conditions for the derivatives
must be provided, the BPTT algorithm requires final conditions for the de-
rivatives:
a 4 a 4
--------------- and ----------------------- ,
iw 1 1 lw 1 1 1
which are normally set to zero.

The network response is then computed for all time steps:
a 1 = lw 1 1 1 a 0 + iw 1 1 p 1
a 2 = lw 1 1 1 a 1 + iw 1 1 p 2
a 3 = lw 1 1 1 a 2 + iw 1 1 p 3
The derivative calculation begins with

1 1
3 = F n 3 = 1 ,
1 1
S
X
ES 1 = 1 , ES 1 = 1 .
Next we compute the following derivative using Eq. (14.63):
14-28
e
F F 1 1 T 1 1 T F
--------------
1
- + LW 1 S t + 1 -----------------------
- = --------------
1 1
-
a t a t a t + 1
For t = 3 , we find
0
e e
F F 1 1 T F F
---------------
1
- + lw 1 1 1 S 4 ---------------
- = ---------------
1 1 1
- = 2e 3
- = ---------------
a 3 a 3 a 4 a 3
and
1 1 1 T F
d 3 = S 3 ---------------
1
- = 2e 3
a 3
Continuing to t = 2 ,
1 1
2 = F n 2 = 1 ,
1 1
S
e
F F 1 1 T F
---------------
1 1
- + lw 1 1 1 S 3 ---------------
- = --------------- 1
-
a 2 a 2 a 3
= 2e 2 + lw 1 1 1 2e 3
and
1 1 1 T F
d 2 = S 2 ---------------
1
- = 2e 2 + lw 1 1 1 2e 3
a 2
Finally, for t = 1 ,
1 1
1 = F n 1 = 1 ,
1 1
S
e
F F 1 1 T F
---------------
1
- + lw 1 1 1 S 2 ---------------
- = ---------------
1 1
-
a 1 a 1 a 2
2
= 2e 1 + lw 1 1 1 2e 2 + lw 1 1 1 2e 3
and
1 1 1 T F 2
d 1 = S 1 ---------------
1
- = 2e 1 + lw 1 1 1 2e 2 + lw 1 1 1 2e 3
a 1
Now we can compute the total gradient, using Eq. (14.54) and Eq. (14.55):
14-29
14 Dynamic Networks
3
F F T
d
1 1
- = ----------------------- =
--------------------------- t a t 1
LW 1
1 1 lw 1 1 1
t=1
2
= a 0 2e 1 + lw 1 1 1 2e 2 + lw 1 1 1 2e 3
+ a 1 2e 2 + lw 1 1 1 2e 3 + a 0 2e 3
3
F F T
d
1 1
- = --------------- =
------------------------- t p t
IW 0
1 1 iw 1 1
t=1
2
= p 1 2e 1 + lw 1 1 1 2e 2 + lw 1 1 1 2e 3
+ p 2 2e 2 + lw 1 1 1 2e 3 + p 3 2e 3
This is the same result that we obtained with the RTRL algorithm on page
14-22.
Summary and Comments on Dynamic Training

The RTRL and BPTT algorithms represent two methods for computing the
gradients for dynamic networks. Both algorithms compute the exact gradi-
ent, and therefore they produce the same final results. The RTRL algo-
rithm performs the calculations from the first time point forward, which is
suitable for on-line (real-time) implementation. The BPTT algorithm starts
from the last time point and works backward in time. The BPTT algorithm
generally requires fewer computations for the gradient calculation than
RTRL, but BPTT usually requires more memory storage.
In addition to the gradient, versions of BPTT and RTRL can be used to com-
pute Jacobian matrices, as are needed in the Levenberg-Marquardt de-
scribed in Chapter 12. For Jacobian calculations, the RTRL algorithm is
generally more efficient that the BPTT algorithm. See [DeHa07] for details.
Once the gradients or Jacobians are computed, many standard optimiza-
tion algorithms can be used to train the networks. However, training dy-
namic networks is generally more difficult than training feedforward
networksfor a number of reasons. First, a recurrent net can be thought
of as a feedforward network, in which the recurrent network is unfolded in
time. For example, consider the simple single-layer recurrent network of
Figure 14.4. If this network were to be trained over five time steps, we
could unfold the network to create 5 layers - one for each time step. If a sig-
moid transfer function is used, then if the output of the network is near the
saturation point for any time point, the resulting gradient could be quite
small.
Another problem in training dynamic networks is the shape of the error
surface. It has been shown (see [PhHa13]) that the error surfaces of recur-
rent networks can have spurious valleys that are not related to the dynam-
14-30
ic system that is being approximated. The underlying cause of these valleys

is the fact that recurrent networks have the potential for instabilities. For
example, the network of Figure 14.4 will be unstable if lw 1 1 1 is greater
than one in magnitude. However, it is possible, for a particular input se-
quence, that the network output can be small for a particular value of
lw 1 1 1 greater than one in magnitude, or for certain combinations of val-
ues for lw 1 1 1 and iw 1 1 .
Finally, it is sometimes difficult to get adequate training data for dynamic
networks. This is because the inputs to some layers will come from tapped
delay lines. This means that the elements of the input vector cannot be se-
lected independently, since the time sequence from which they are sampled
is generally correlated in time. Unlike static networks, in which the net-
work response depends only on the input to the network at the current
time, dynamic network responses depend on the history of the input se-
quence. The data used to train the network must be representative of all
situations for which the network will be used, both in terms of the ranges
for each input, but also in terms of the variation of the inputs over time.
2 To illustrate the training of dynamic networks, consider again the simple
+2
recurrent network of Figure 14.4, but lets use a nonlinear sigmoid transfer
function, as shown in Figure 14.9.
Inputs Tan-Sigmoid Neuron
iw1,1 n(t) a(t)

p(t)
S
lw1,1(1)
D
a(t) = tansig( iw1,1 p(t) + lw1,1(1) a(t-1) )
Figure 14.9 Nonlinear Recurrent Network

Recall from Chapter 11 that static multilayer networks can be used to ap-
proximate functions. Dynamic networks can be used to approximate dy-
namic systems. A function maps from one vector space (the domain) to
another vector space (the range). A dynamic system maps from one set of
time sequences (the input sequences p t ) to another set of time sequences
(the output sequences a t ). For example, the network of Figure 14.9 is a
dynamic system. It maps from input sequences to output sequences.
In order to simplify our analysis, we will give the network a problem for
which we know the optimal solution. The dynamic system we will approxi-
mate is the same network, with the following values for the weights:
14-31
14 Dynamic Networks
lw 1 1 1 = 0.5 , iw 1 1 = 0.5 , (14.64)
The input sequence that we use to train a dynamic network must be repre-
sentative of all possible input sequences. Because this network is so simple,
it is not difficult to find an appropriate input sequence, but for many prac-
tical networks it can be difficult. We will use a standard form of input se-
quence (called the skyline function), which consists of a series of pulses of
varying height and width. The input and target sequences are shown in
Figure 14.10. The circles represent the input sequence and the dots repre-
sent the target sequence. The targets were created by applying the given
input sequence to the network of Figure 14.9, with the weights given by Eq.
(14.64).
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18 20
Figure 14.10 Input and Target Sequences

Figure 14.11 shows the squared error performance surface for this prob-
lem. Note that as the weight lw 1 1 1 becomes greater than one in magni-
tude, the squared error grows steeply. This effect would be even more
prominent, if the length of the training sequence were longer. However, we
can also see some narrow valleys in the surface in the regions where
lw 1 1 1 is greater than one. (This is a very common result, as discussed in
[PhHa13]. See Exercise E14.18 to investigate the cause of these valleys.)
The narrow valleys can have an effect on training, since the trajectory can
be trapped or misdirected by the spurious valleys. On the left of Figure
14.11 we see a steepest descent path. The path is misdirected at the begin-
ning of the trajectory, because of the narrow valley seen near the bottom of
the contour plot.
14-32
F
1.5
1.4
1 1.2
1
0.5
0.8
lw 1 1 1 0 0.6
0.4
0.5
0.2
0
1
14
2
1 2
1.5 1
0
lw 1 1 1
0
1
1
2
2 1.5 1 0.5 0 0.5 1 1.5 2 2 2 iw 1 1
iw 1 1
Figure 14.11 Performance Surface and Steepest Descent Trajectory

To experiment with the training of this recurrent network, use the Neural
Network Design Demonstration Recurrent Network Training (nnd14rnt).
14-33
14 Dynamic Networks
Summary of Results
Initialize: Real-Time Recurrent Learning Gradient
u
a t
- = 0 t 0 , for all u U ,
--------------
T
x
For t = 1 to Q,
X
U = , E S u = and E S u = for all u U .
D For all u U , if E S u L m
b
u m u l l m
0 F n t
m

m
S t = S t LW
b
l ES u Lm
X
if m X , add m to the set E S u
EndFor u
If m U
m m
S t = F n t
m m
add m to the sets U and E S m

X
if m X , add m to the set E S m
EndIf m
EndFor m
For u U incremented through the simulation order
For all weights and biases (x is a vector containing all weights and biases)
e u
a t l T u m
------------------------------------------T = p t d S t
m l
vec IW d
e u
a t l T u m
-------------------------------------------T- = a t d S t
m l
vec LW d
e u
a t u m
----------------T- = S t
m
b
EndFor weights and biases
u e u u'
--------------
x
T
- = ----------------
x
T
-+ S t LW d ------------------------
x
T
-
X d DL x u'
x ES u U
u' E LW x
EndFor u
EndFor t
Compute Gradients
Q T
u e
F a t F
------ =
x --------------
x
T
- --------------
u
a t
-
t = 1u U
14-34
Summary of Results
Initialize: Backpropagation-Through-Time Gradient

F
- = 0 t Q , for all u U ,
--------------
u
a t
For t = Q to 1,
U
U = , E S u = , and E S u = for all u U .
b
u m u l l m
14
0 F n t
m

m
S t = S t LW
b
l ES u L m

U
add u to the set E S m
EndFor u
If m U
m m
S t = F n t
m m
U
add m to the sets U , E S m and E S m
EndIf m
EndFor m
For u U decremented through the BP order
e
F F x u T u' x T F
--------------
u
a t
- = --------------
u
a t
-+ LW d S t + d ------------------------
u'
a t + d
-
x E LW u d DL x u
X U
u' E S x
EndFor u
For all layers m
u m T F

m
d t = S t --------------
u
-
U a t
u ES m
EndFor m
EndFor t
Compute Gradients
Q
F T
d
m l
---------------------------
m l
- = t a t d
LW d t=1
Q
F T
d
m l
-------------------------
m l
- = t p t d
IW d
t=1
Q
F
d
m
--------m- = t
b t=1
14-35
14 Dynamic Networks
Definitions/Notation
l
p t is the lth input vector to the network at time t.
m
n t is the net input for layer m.
m
f is the transfer function for layer m.
m
a t is the output for layer m.
m l
D IW is the input weight between input l and layer m.
m l
LW is the layer weight between layer l and layer m.
m
b is the bias vector for layer m.
DL m l is the set of all delays in the tapped delay line between Layer l and
Layer m.
DIm l is the set of all delays in the tapped delay line between Input l and
Layer m.
I m is the set of indices of input vectors that connect to layer m.
f
L m is the set of indices of layers that directly connect forward to layer m.
b
L m is the set of indices of layers that are directly connected backwards to
layer m (or to which layer m connects forward) and that contain no delays
in the connection.
A layer is an input layer if it has an input weight, or if it contains any de-
lays with any of its weight matrices. The set of input layers is X.
A layer is an output layer if its output will be compared to a target during
training, or if it is connected to an input layer through a matrix that has
any delays associated with it. The set of output layers is U.
e u
u m ak t
The sensitivity is defined s k i t ----------------
m
-.
n i t
U x u
E LW x = u U LW d 0 d 0
X u x
E S u = x X S 0
u x
E S u = x S 0
X x u
E LW u = x X LW d 0 d 0
U u x
E S x = u U S 0
14-36
Solved Problems
Solved Problems
P14.1 Before stating the problem, lets first introduce some notation that
will allow us to efficiently represent dynamic networks:
Layer #1 with 2 tansig neurons. Input vector with 3 elements.
2 3
Tapped delay from 1 to 9. Tapped delay from 0 to 9.
9 9
Figure P14.1 Blocks for Dynamic Network Schematics

Using this notation, consider the following network
1
1
2 3 1 4 5 1 6
1
3 2
3 2 1 2 4
7 8 1 9 10
2 3 2 1
Figure P14.2 Example Dynamic Network for Problem P14.1

Define the network architecture, by showing U , X , I m , DI m 1 , DL m l ,
f b U X
L m , L m , E LW x , E LW u . Also, select a simulation order and indicate
the dimension of each weight matrix.
The input layers have input weights, or have delays with their weight ma-
trices. Therefore X = 1 2 4 6 9 . The output layers are compared to tar-
gets, or connect to an input layer through a matrix that has any delays. If
14-37
14 Dynamic Networks
we assume that only Layer 10 is compared to a target, then

U = 2 3 4 5 8 10 . Since the single input vector connects only to Layer
1, the only nonempty set of inputs will be I 1 = 1 . For the same reason,
there will only be one nonempty set of input delays: DI 1 1 = 0 . The con-
nections between layers are defined by
f f f f f
L 1 = , L 2 = 1 2 , L 3 = 2 , L 4 = 3 4 7 , L 5 = 4 ,
f f f f f
D L 6 = 2 5 , L 7 = 2 , L 8 = 7 , L 9 = 4 8 , L 10 = 6 9 .
b b b b b
L 1 = 2 , L 2 = 3 6 7 , L 3 = , L 4 = 5 9 , L 5 = ,
b b b b b
L 6 = 10 , L 7 = 4 8 , L 8 = , L 9 = 10 , L 10 = .
Associated with these connections are the following layer delays
DL 2 1 = 0 , DL 2 2 = 1 , DL 3 2 = 0 , DL 4 3 = 1 , DL 4 4 = 1 ,
DL 4 7 = 0 , DL 5 4 = 0 , DL 6 2 = 0 , DL 6 5 = 0 1 , DL 7 2 = 0 ,
DL 8 7 = 0 , DL 9 4 = 0 , DL 9 8 = 1 , DL 10 6 = 0 , DL 10 9 = 0 .
The layers that have connections from output layers are

U U
E LW 2 = 2 , E LW 4 = 3 4 ,
U U
E LW 6 = 5 , E LW 9 = 8 .
The layers that connect to input layers are

X X X
E LW 2 = 2 , E LW 3 = 4 , E LW 4 = 4 ,
X X
E LW 5 = 6 , E LW 8 = 9 .
The simulation order can be chosen to be 1 2 3 7 4 5 6 8 9 10 . The di-

mensions of the weight matrices are
1 1 2 1 2 2 3 2
IW 0 2 3 , LW 0 3 2 , LW 1 3 3 , LW 0 2 3 ,
4 3 4 4 4 7 5 4
LW 1 1 2 , LW 1 1 1 , LW 0 1 2 , LW 0 2 1 ,
6 2 6 5 7 2 8 7
LW 0 4 3 , LW 1 4 2 , LW 0 2 3 , LW 0 3 2 ,
9 4 9 8 10 6 10 9
LW 0 2 1 , LW 1 2 3 , LW 0 1 4 , LW d 1 2 .
14-38
Solved Problems
P14.2 Write out the BPTT equations for the network presented in Prob-
lem P14.1.
We will assume that the network response has been computed for all time
points, and we will demonstrate the process for one time step, since they all
will be similar. We proceed through the layers according to the backpropa-
gation order, which is the reverse of the simulation order:
10 9 8 6 5 4 7 3 2 1 .
10 10
t = F n t
10 10
S
e
F F
----------------
10
- = ----------------
10
-
a t a t
10 10 10 T F
d t = S t ----------------
10
-
a t
10 9 10 10 10 9
0 F n t
9 9
S t = S t LW
9 10 9 T F
d t = S t ----------------
10
-
a t
8 8
t = F n t
8 8
S
e
F F 9 8 T 10 9 T F
8 8
- + LW 1 S
- = --------------
-------------- t + 1 -------------------------
10
-
a t a t a t + 1
8 8 8 T F
d t = S t --------------
8
-
a t
10 6 10 10 10 6
0 F n t
6 6
S t = S t LW
6 10 6 T F
d t = S t ----------------
10
-
a t
5 5
t = F n t
5 5
S
e
F F 6 5 T 10 6 T F 6 5 T 10 6 T F
--------------
5
- + LW 1 S t + 1 -------------------------
- = --------------
5 10
- + LW 0 S t ----------------
10
-
a t a t a t + 1 a t
5 5 5 T F
d t = S t --------------
5
-
a t
14-39
14 Dynamic Networks
10 4 10 9 9 4
0 F n t
4 4
S t = S t LW
5 4 5 5 5 4
0 F n t
4 4
S t = S t LW
4 4
t = F n t
4 4
S
e
F F 4 4 T 4 4 T F 5 4 T F 10 4 T F
- = --------------
-------------- - + S t + 1 -----------------------
- + LW 1 S t + 1 ----------------------- - + S t + 1 -------------------------
-
D 4
a t
4
a t
4
a t + 1
5
a t + 1
10
a t + 1
4 4 4 T F 5 4 T F 10 4 T F
d t = S 4
- + S t --------------
t -------------- 5
- + S t ----------------
10
-
a t a t a t
10 7 10 4 4 7
0 F n t
7 7
S t = S t LW
8 7 8 8 8 7
0 F n t
7 7
S t = S t LW
5 7 5 4 4 7
0 F n t
7 7
S t = S t LW
4 7 4 4 4 7
0 F n t
7 7
S t = S t LW
7 10 7 T F 8 7 T F 5 7 T F 4 7 T F
d t = S 10
- + S t --------------
t ---------------- 8
- + S t --------------
5
- + S t --------------
4
-
a t a t a t a t
3 3
t = F n t
3 3
S
e
F F 4 3 T 4 4 T F 5 4 T F 10 4 T F
--------------
3
- + LW 1 S t + 1 -----------------------
- = --------------
3 4
- + S t + 1 -----------------------
5
- + S t + 1 -------------------------
10
-
a t a t a t + 1 a t + 1 a t + 1
3 3 3 T F
d t = S t --------------
3
-
a t
10 2 10 6 6 2 10 7 7 2
0 F n t + S t LW 0 F n t
2 2 2 2
S t = S t LW
8 2 8 7 7 2
0 F n t
2 2
S t = S t LW
5 2 5 7 7 2
0 F n t
2 2
S t = S t LW
4 2 4 7 7 2
0 F n t
2 2
S t = S t LW
3 2 3 3 3 2
0 F n t
2 2
S t = S t LW
14-40
Solved Problems
2 2
t = F n t
2 2
S
e
F F 2 2 T 2 2 T F 3 2 T F
2
- = --------------
-------------- 2 2
- + S t + 1 -----------------------
- + LW 1 S t + 1 ----------------------- 3
-
a t a t a t + 1 a t + 1
4 2 T F 5 2 T F
+S 4
- + S t + 1 -----------------------
t + 1 ----------------------- 5
-
a t + 1 a t + 1
F F
+S
8 2 T
-+S
t + 1 -----------------------
8
a t + 1
10 2 T
t + 1 -------------------------
10
a t + 1
- 14
2 10 2 T F 8 2 T F 5 2 T F
d t = S 10
- + S t --------------
t ---------------- 8
- + S t --------------
5
-
a t a t a t
4 2 T F 3 2 T F 2 2 T F
+ S 4
- + S t --------------
t -------------- 3
- + S t --------------
2
-
a t a t a t
10 1 10 2 2 1
0 F n t
1 1
S t = S t LW
8 1 8 2 2 1
0 F n t
1 1
S t = S t LW
5 1 5 2 2 1
0 F n t
1 1
S t = S t LW
4 1 4 2 2 1
0 F n t
1 1
S t = S t LW
3 1 3 2 2 1
0 F n t
1 1
S t = S t LW
2 1 2 2 2 1
0 F n t
1 1
S t = S t LW
1 10 1 T F 8 1 T F 5 1 T F
d t = S 10
- + S t --------------
t ---------------- 8
- + S t --------------
5
-
a t a t a t
4 1 T F 3 1 T F 2 1 T F
+ S 4
- + S t --------------
t -------------- 3
- + S t --------------
2
-
a t a t a t
After the preceding steps have been repeated for all time points, from the
last time point back to the first, then the gradient can be computed as fol-
lows:
Q
F T
d
m l
---------------------------
m l
- = t a t d
LW d t=1
14-41
14 Dynamic Networks
Q
F T
d
m l
-------------------------
m l
- = t p t d
IW d t=1
Q
F
d
m
--------m- = t
b t=1
P14.3 Write out the RTRL equations for the network presented in Prob-
lem P14.1.
As in the previous problem, we will demonstrate the process for one time
step, since each step is similar. We will proceed through the layers accord-
u m
ing to the backpropagation order. The sensitivity matrices S t are com-
puted in the same way for the RTRL algorithm as for the BPTT algorithm,
so we wont repeat those steps from Problem P14.2.
The explicit derivative calculations for the input weight will be
e u
a t 1 T u 1
------------------------------------------T = p t S t
1 1
vec IW 0
For the layer weights and the biases, the explicit derivatives are calculated
by
e u
a t l T u m
-------------------------------------------T- = a t d S t
m l
vec LW d
e u
a t u m
----------------T- = S t
m
b
For the total derivatives, we have
2 e 2 2
a t a t 2 2 2 2 a t 1
--------------
T T
- + S t LW 1 -----------------------
- = ---------------- T
-
x x x
3 e 3 2
a t a t 3 2 2 2 a t 1
--------------
T
- + S t LW 1 -----------------------
- = ----------------
T T
-
x x x
14-42
Solved Problems
4 e 4 4 3
a t a t 4 4 4 4 a t 1 4 3 a t 1
--------------
T
- + S t LW 1 -----------------------
- = ----------------
T T
- + LW 1 -----------------------
T
-
x x x x
2
4 2 2 2 a t 1
+S t LW 1 -----------------------
T
-
x
5
a t
--------------
x
T
e 5
a t
x
T
5 4 4 4 a t 1
- + S t LW 1 -----------------------
- = ----------------
x
T
4 3
4
a t 1
- + LW 1 -----------------------
x
T
-
3
14
2
5 2 2 2 a t 1
+S t LW 1 -----------------------
T
-
x
8 e 8 2
a t a t 8 2 2 2 a t 1
T T
- + S t LW 1 -----------------------
- = ----------------
-------------- T
-
x x x
10 e 10 8
a t a t 10 9 9 8 a t 1
----------------
T
- + S t LW 1 -----------------------
- = ------------------
T T
-
x x x
5 5
10 6 6 5 a t 6 5 a t 1
+S t LW - + LW 1 -----------------------
0 --------------
T T
-
x x
4 3 2
10 4 4 4 a t 1 4 3 a t 1 10 2 2 2 a t 1
+S t LW - + LW 1 -----------------------
1 -----------------------
T T
- +S t LW 1 -----------------------
T
-
x x x
After the above steps are iterated through all time points, we can compute
the gradient with
Q T T T
e 2 e 3 e 4
F F a t F a t F a t
-------T- =
x
--------------
2
a t
- --------------
x
T
- + --------------
3
a t
- --------------
x
T
- + --------------
4
a t
- --------------
x
T
-
t=1
e T 5 e T 8 e T 10
F a t F a t F a t
+ --------------
5
- --------------
T
- + --------------
8
- --------------
T
- + ----------------
10
- ----------------
T
-
a t x a t x a t x
P14.4 From the previous problem, show the detail of the calculations in-
volved in the explicit derivative term
e 2
a t 1 T 2 1
------------------------------------------T = p t S t
1 1
vec IW 0
14-43
14 Dynamic Networks
First, lets display the details of the individual vectors and matrices in this
expression:
1 1 1 1 1 1
1 1 iw 1 1 iw 1 2 iw 1 3
IW 0 =
1 1 1 1 1 1
iw 2 1 iw 2 2 iw 2 3
1 1 T
vec IW 0 = iw 1 1 iw 1 1 iw 1 1 iw 1 1 iw 1 1 iw 1 1
D 1 1 2 1 1 2 2 2 1 3 2 3
2 1 2 1
p1 e 2
s 1 1 s 1 2
1 2 1 a t
p t = p2 S t = -----------------T- = s 2 1 s 2 1
1 2 1 2 2
n t
p3 2 1 2 1
s 3 1 s 3 2
2 2 2 2 2 2
a 1 a 1 a 1 a 1 a 1 a 1
1 1 1 1 1 1 1 1 1 1 1 1
iw 1 1 iw 2 1 iw 1 2 iw 2 2 iw 1 3 iw 2 3
e 2
a t
2 2 2 2 2 2
a 2 a 2 a 2 a 2 a 2 a 2
------------------------------------------T =
1 1 1 1 1 1 1 1 1 1 1 1 1 1
vec IW 0 iw 1 1 iw 2 1 iw 1 2 iw 2 2 iw 1 3 iw 2 3
2 2 2 2 2 2
a 3 a 3 a 3 a 3 a 3 a 3
1 1 1 1 1 1 1 1 1 1 1 1
iw 1 1 iw 2 1 iw 1 2 iw 2 2 iw 1 3 iw 2 3
The Kronecker product is defined as
a 1 1 B a 1 m B
AB =

a n 1 B a n m B
,
therefore
2 1 2 1 2 1 2 1 2 1 2 1
p 1 s 1 1 p 1 s 1 2 p 2 s 1 1 p 2 s 1 2 p 3 s 1 1 p 3 s 1 2
1 T 2 1
p t S t = p s 2 1 p s 2 1 p s 2 1 p s 2 1 p s 2 1 p s 2 1 .
1 2 1 1 2 2 2 2 1 2 2 2 3 2 1 3 2 2
2 1 2 1 2 1 2 1 2 1 2 1
p 1 s 3 1 p 1 s 3 2 p 2 s 3 1 p 2 s 3 2 p 3 s 3 1 p 3 s 3 2
14-44
Solved Problems
P14.5 Find the computational complexity for the BPTT and RTRL algo-
rithms applied to the sample network in Figure P14.3 as a function
1
of the number of neurons in Layer 1 ( S ), the number of delays in
the tapped delay line ( D ) and the length of the training sequence
( Q ).
1
D
R
S
1
Figure P14.3 Sample Network for Exercise E14.1

The complexity of the BPTT gradient calculation is generally determined
by Eq. (14.54). For this network, the most important weights will be
1 1
LW d :
Q
F T
d
1 1
---------------------------
1 1
- = t a t d ,
LW d t=1
1 2
The outer product calculation involves S operations, which must be
done for Q time steps and for D delays, so the BPTT gradient calculation is
1 2
O S DQ .
The complexity of the RTRL gradient is based generally on Eq. (14.34). For
this sample network, we can consider the equation for u = 2 :
D
2 e 2 1
a t a t 2 1 1 1 a t d
--------------
x
T
- + S t
- = ----------------
x
T LW d -----------------------
x
T
-
d=1
1 1
Inside the summation we have a matrix multiplication involving an S S
1 1 1
matrix times an S DS + 3 S + 1 matrix. This multiplication will be
1 4
O S D . It must be done for every d and every t, therefore the RTRL gra-
1 4 2
dient calculations are O S D Q . The multiplication by the sensitivity
matrix does not change the order of the complexity.
14-45
14 Dynamic Networks
Epilogue
Dynamic networks can be trained using the same optimization procedures
that we described in Chapter 12 for static multilayer networks. However,
the calculation of the gradient for dynamic networks is more complex than
for static networks. There are two basic approaches to the gradient calcu-
lation in dynamic networks. One approach, backpropagation through time
(BPTT), starts at the last time point, and works backward in time to com-
pute the gradient. The second approach, real-time recurrent learning
(RTRL), starts at the first time point, and then works forward through
time.
RTRL requires more calculations than BPTT for calculating the gradient,
but RTRL allows a convenient framework for on-line implementation. Also,
RTRL generally requires less storage than BPTT. For Jacobian calcula-
tions, RTRL is often more efficient than the BPTT algorithm.
Chapter 27 presents a real-world case study for using dynamic networks to
solve a prediction problem.
14-46
Further Reading
Further Reading
[DeHa07] O. De Jess and M. Hagan, Backpropagation Algorithms
for a Broad Class of Dynamic Networks, IEEE Transac-
tions on Neural Networks, vol. 18, no. 1, pp., 2007.
This paper presents a general development of BPTT and
RTRL algorithms for gradient and Jacobian calculations.
Experimental results are presented that compare the com-
putational complexities of the two algorithms for a variety
14
of network architectures.
[MaNe99] J.R. Magnus and H. Neudecker, Matrix Differential Calcu-
lus, John Wiley & Sons, Ltd., Chichester, 1999.
This textbook provides a very clear and complete descrip-
tion of matrix theory and matrix differential calculus.
[PhHa13] M. Phan and M. Hagan, Error Surface of Recurrent Net-
works, IEEE Transactions on Neural Networks and
Learning Systems, Vol. 24, No. 11, pp. 1709 - 1721, October,
2013.
This paper describes spurious valleys that appear in the er-
ror surfaces of recurrent networks. It also describes some
procedures that can be used to improve training of recur-
rent networks.
[Werb90] P. J. Werbos, Backpropagation through time: What it is
and how to do it, Proceedings of the IEEE, vol. 78, pp.
15501560, 1990.
The backpropagation through time algorithm is one of the
two principal approaches to computing the gradients in re-
current neural network. This paper describes the general
framework for backpropagation through time.
[WiZi89] R. J. Williams and D. Zipser, A learning algorithm for con-
tinually running fully recurrent neural networks, Neural
Computation, vol. 1, pp. 270280, 1989.
This paper introduced the real-time, recurrent learning al-
gorithm for computing the gradients of dynamic networks.
Using this method, the gradients are computed by starting
at the first time point, and then working forward through
time. The algorithm is suitable for on-line, or real-time, im-
plementation.
14-47
14 Dynamic Networks
Exercises
E14.1 Put the network of Figure 14.1 into the schematic form, which we intro-
duced in Problem P14.1.
E14.2 Consider the network in Figure 14.4, with weight values iw 1 1 = 2 and
lw 1 1 1 = 0.5 . If a 0 = 4 , and p 1 = 2 , p 2 = 3 , p 3 = 2 , find a 1 ,
a2 , a3 .
1
E14.3 Consider the network in Figure P14.3, with D = 2 , S = 1 , R = 1 ,
1 1 1 1 1 1 1
IW = 1 , LW 1 = 0.5 , LW 2 = 0.2 , a 0 = 2 and
1 1 1 1 1
a 1 = 1 . If p 1 = 1 , p 2 = 2 and p 3 = 1 , find a 1 ,
1 1
a 2 and a 3 .
E14.4 Consider the network in Figure E14.1. Define the network architecture, by
f b U X
showing U , X , I m , DI m 1 , DL m l , L m , L m , E LW x , E LW u . Also, select a
simulation order and indicate the dimension of each weight matrix.
2
1 2 3 2 4
2
3 2 4 5 3
Figure E14.1 Dynamic Network for Exercise E14.4
E14.5 Write out the RTRL equations for the network in Figure E14.2
1 1 2 3
1
2
3
2 2 2
E14.6 Write out the BPTT equations for the network in Figure E14.2.
14-48
Exercises
E14.7 Write out the equations of operation for the network in Figure E14.3. As-
sume that all weights have a value of 0.5, and that all biases have a value
of 0.
i. Assume that the initial network output is a 0 = 0.5 , and that the
initial network input is p 1 = 1 . Solve for a 1 .
ii. Describe any problems that you would have in simulating this net-
work. Will you be able to apply the BPTT and RTRL algorithms to
compute the gradients for this network? What test should you apply
to recurrent networks to determine if they can be simulated and
trained?
1
1
1
1
E14.8 Consider the network in Figure E14.4.
1 1 2
1
1
3
2 5

i. Write out the equations for computing the network response.
ii. Write out the BPTT equations for the network.
iii. Write out the RTRL equations for the network.
E14.9 Repeat E14.8 for the following networks.
14-49
14 Dynamic Networks
1
i. 1 2
1
3
2 5
ii. 1 2
1
3
2 5
E14.10 Consider the network in Figure E14.5.
1 2 1 3
1
2
3
1 5 4
Figure E14.5 Recurrent Network for Exercise E14.10

i. Define the network architecture, by showing U , X , I m , DI m 1 ,
f b U X
DL m l , L m , L m , E LW x , E LW u .
ii. Select a simulation order and write out the equations needed to de-
termine the network response.
u x
iii. Which S t will need to be calculated (i.e., for which u and which
x )?
3
iv. Write out Eq. (14.63) specifically for the term F a t and Eq.
3 T
(14.34) specifically for the term a t x . (Expand the summation
to show exactly which terms are included.)
14-50
Exercises
3
E14.11 Repeat E14.10 for the following networks, except, in part iv., change a to
the indicated layer.
2
1 3 4
4
i. 3 2 5 ,a .
3
2 5
1
6 3
1
ii. 1 1 2 3 1 4 4
,a .
1
3 2 2 5 5
2
1 2 4
1
2
iii. 2 5 3 ,a .
3
14-51
14 Dynamic Networks
2
1 3 4
iv. 3 2 5 2
,a .
3
2 5
1
6 3
1 2 3 4 6
3 3 2 2 5 1
v. ,a .
1
2 5
2
6 3
E14.12 One of the advantages of the RTRL algorithm is that the gradient can be
computed at the same time as the network response, since it is computed
by starting at the first time point, and then working forward through time.
Therefore, RTRL allows a convenient framework for on-line implementa-
tion. Suppose that you are implementing the RTRL algorithm, and you up-
date the weights of the network at each time step.
i. Discuss the accuracy of the gradient calculation if the weights are
updated at each time step of a steepest descent algorithm.
2+2
ans = ii. Implement the RTRL algorithm in MATLAB for the network in Fig-
4 ure 14.4. Generate training data by using the network with the two
weights set equal to 0.5. Use the same input sequence shown in Fig-
ure 14.10, and use the network responses a t as the targets. Using
this data set, train the network using the steepest descent algo-
rithm, with a learning rate of = 0.1 . Set the initial weights to ze-
ro. Update the weights at each time step. Save the gradient at the
end of the eighth time step.
14-52
Exercises
iii. Repeat part ii., but do not update the weights. Compute the gradi-
ent at the end of the eighth time step. Compare this gradient with
the one you obtained in part ii. Explain any differences
E14.13 Consider again the recurrent network of Figure 14.4. Assume that F x is
the sum squared error for the first two time points.
i. Let a 0 = 0 . Find a 1 and a 2 as a functions of p 1 , p 2 , and
the weights in the network.
ii. Find the sum squared error for the first two time steps as an explicit
function of p 1 , p 2 , t 1 , t 2 , and the network weights.
F
iii. Using part (ii), find ----------------------- .
lw 1 1 1
iv. Compare the results of part (iii) with the results determined by
RTRL on page 14-22 and by BPTT on page 14-30.
E14.14 In the process of deriving the RTRL algorithm in Exercise E14.5, you
should have produced the following expression
e 3
a t 1 T 3 2
-------------------------------------------T- = a t 1 S t
2 1
vec LW 1
If
1
a 1 = 1 and S 3 2 2 = 2 3 ,
1 4 5
e 3
e 3
a 2 a 1 2
Find -------------------------------------------T- and indicate ---------------------------------------
-.
2 1 2 1 T
vec LW 1 vec lw 1 2 1
E14.15 Each layer of a standard LDDN network has a summing junction, which
combines contributions from inputs, other layers, and the bias, as in Eq.
(14.1), which is repeated here:
m l m l

m l l m
n t = LW d a t d + IW d p t d + b .
l L m d DL m l l I m d DI m l
f
If, instead of summing, the net input was computed as a product of the con-
tributions, how would the RTRL and BPTT algorithms change?
E14.16 As discussed in Exercise E14.15, the contribution to the net input from oth-
er layers is computed as product of a layer matrix with a layer output, as in
m l l
LW d a t d .
14-53
14 Dynamic Networks
If, instead of multiplying the layer matrix times the layer output, we were
to compute the distance between each row of the weight matrix and the lay-
er output, as in
n i = iw a .
How would the RTRL and BPTT algorithms change?
E14.17 Find and compare the computational complexity for the BPTT and RTRL
algorithms applied to the sample network in Figure E14.6 as a function of
2
the number of neurons in Layer 2 ( S ), the number of delays in the tapped
delay line ( D ) and the length of the training sequence ( Q ).
1 D 2
3
2 S
2
Figure E14.6 Recurrent Network for Exercise E14.17
E14.18 Consider again the network of Figure 14.4. Let the input weight of the net-
work iw 1 1 = 1 . Assume that the initial network output is a 0 = 0 .
i. Write the network output at time t as a function only of the layer

weight lw 1 1 1 , and the input sequence. (The result should be a
polynomial in lw 1 1 1 .)
2+2
ans = ii. Find the network output at time t = 8 , using lw 1 1 1 = 1.4 and
4 the following input sequence:
p t = 3 1 1 6 3 5 1 6 .
iii. With lw 1 1 1 = 1.4 , the network should be unstable, since this

weight in the feedback loop is greater than one in magnitude. The
output would generally grow with time for an unstable network.
(This applies to linear networks.) In ii., you should have found a
small value for a 8 . Can you explain this result? (Hint: Investigate
the roots of the polynomial you found in part i. You can use the
MATLAB command roots.) How might this result be related to the
spurious valleys in the error surface discussed on page 14-32?
14-54
Objectives
15 Associative Learning
Objectives 15-1
Simple Associative Network 15-3
Unsupervised Hebb Rule 15-5
Hebb Rule with Decay 15-7
Simple Recognition Network 15-9
Instar Rule 15-11
Kohonen Rule 15-15
Simple Recall Network 15-16
Outstar Rule 15-17
Epilogue 15-34
Exercises 15-37
Objectives
The neural networks we have discussed so far (in Chapters 4, 7, 1014)
have all been trained in a supervised manner. Each network required a tar-
get signal to define correct network behavior.
In contrast, this chapter introduces a collection of simple rules that allow
unsupervised learning. These rules give networks the ability to learn asso-
ciations between patterns that occur together frequently. Once learned, as-
sociations allow networks to perform useful tasks such as pattern
recognition and recall.
Despite the simplicity of the rules in this chapter, they will form the foun-
dation for powerful networks in Chapters 16, 18, 19.
15-1
Theory and Examples

This chapter is all about associations: how associations can be represented
by a network, how a network can learn new associations.
What is an association? An association is any link between a systems input
and output such that when a pattern A is presented to the system it will
respond with pattern B. When two patterns are linked by an association,
Stimulus the input pattern is often referred to as the stimulus. Likewise, the output
Response pattern is referred to as the response.
Associations are so fundamental that they formed the foundation of the be-
haviorist school of psychology. This branch of psychology attempted to ex-
plain much of animal and human behavior by using associations and rules
for learning associations. (This approach has since been largely discredit-
ed.)
One of the earliest influences on the behaviorist school of psychology was
the classic experiment of Ivan Pavlov, in which he trained a dog to salivate
at the sound of a bell, by ringing the bell whenever food was presented. This
is an example of what is now called classical conditioning. B. F. Skinner
was one of the most influential proponents of the behaviorist school. His
classic experiment involved training a rat to press a bar in order to obtain
a food pellet. This is an example of instrumental conditioning.
It was to provide a biological explanation for some of this behavior that led
Donald Hebb to his postulate, previously quoted in Chapter 7 [Hebb49]:
When an axon of cell A is near enough to excite a cell B and repeatedly or
In Chapter 7 we analyzed the performance of a supervised learning rule
based on Hebbs postulate. In this chapter we will discuss unsupervised
forms of Hebbian learning, as well as other related associative learning
rules.
A number of researchers have contributed to the development of associa-
tive learning. In particular, Tuevo Kohonen, James Anderson and Stephen
Grossberg have been very influential. Anderson and Kohonen indepen-
dently developed the linear associator network in the late 1960s and early
1970s ([Ande72], [Koho72]). Grossberg introduced nonlinear continuous-
time associative networks during the same time period (e.g., [Gross68]). All
of these researchers, in addition to many others, have continued the devel-
opment of associative learning up to the present time.
In this chapter we will discuss the elemental associative learning rules.
Then, in Chapters 1416 we will present more complex networks that use
15-2
Simple Associative Network
associative learning as a primary component. Chapter 16 will describe Ko-

honen networks, and Chapters 18 and 19 will discuss Grossberg networks.
Simple Associative Network

Lets take a look at the simplest network capable of implementing an asso-
ciation. An example is the single-input hard limit neuron shown in Figure
15.1.
w n a
p
b = -0.5
1
a = hardlim (wp + b)
Figure 15.1 Single-Input Hard Limit Associator

The neurons output a is determined from its input p according to
a = hardlim wp + b = hardlim wp 0.5 . (15.1)
For simplicity, we will restrict the value of p to be either 0 or 1, indicating

whether a stimulus is absent or present. Note that a is limited to the same
values by the hard limit transfer function. It indicates the presence or ab-
sence of the networks response.
1 stimulus 1 response
p = a = (15.2)
0 no stimulus 0 no response
The presence of an association between the stimulus p = 1, and the re-

sponse a = 1 is dictated by the value of w . The network will respond to the
stimulus only if w is greater than b (in this case 0.5).
The learning rules discussed in this chapter are normally used in the
framework of a larger network, such as the competitive networks of Chap-
ters 16, 18 and 19. In order to demonstrate the operation of the associative
learning rules, without using complex networks, we will use simple net-
works that have two types of inputs.
Unconditioned Stimulus One set of inputs will represent the unconditioned stimulus. This is analo-
gous to the food presented to the dog in Pavlovs experiment. Another set
Conditioned Stimulus of inputs will represent the conditioned stimulus. This is analogous to the
bell in Pavlovs experiment. Initially the dog salivates only when food is
15-3
presented. This is an innate characteristic that does not have to be learned.

However, when the bell is repeatedly paired with the food, the dog is con-
ditioned to salivate at the sound of the bell, even when no food is present.
We will represent the unconditioned stimulus as p 0 and the conditioned
stimulus simply as p . For our purposes we will assume that the weights
associated with p 0 are fixed, but that the weights associated with p are ad-
justed according to the relevant learning rule.
2 Figure 15.2 shows a network for recognizing bananas. The network has
+2
both an unconditioned stimulus (banana shape) and a conditioned stimulus
(banana smell). We dont mean to imply here that smell is more condition-
able than sight. In our examples in this chapter the choices of conditioned
and unconditioned stimuli are arbitrary and are used simply to demon-
strate the performance of the learning rules. We will use this network to
demonstrate the operation of the Hebb rule in the following section.
Sight of banana p0 w0 = 1
n a Banana?

Smell of banana p w = 0 b = -0.5
1
a = hardlim (w0p0 + w p + b)
Figure 15.2 Banana Associator

The definitions of the unconditioned and conditioned inputs for this net-
Fruit work are
0 1 shape detected 1 smell detected

p = p = . (15.3)
Shape Smell 0 shape not detected 0 smell not detected
Network At this time we would like the network to associate the shape of a banana,
but the not the smell, with a response indicating the fruit is a banana. The
0
problem is solved by assigning a value greater than b to w and assigning
Banana? a value less than b to w . The following values satisfy these requirements:
0
w = 1, w = 0. (15.4)
The banana associators input/output function now simplifies to
15-4
Unsupervised Hebb Rule
0
a = hardlim p 0.5 . (15.5)
0
Thus, the network will only respond if a banana is sighted ( p = 1 ), wheth-
er a banana is smelled ( p = 1 ) or not ( p = 0 ).
We will use this network in later sections to illustrate the performance of
several associative learning rules.

For simple problems it is not difficult to design a network with a fixed set
of associations. On the other hand, a more useful network would be able to
learn associations.
When should an association be learned? It is generally accepted that both
animals and humans tend to associate things that occur simultaneously.
To paraphrase Hebb: if a banana smell stimulus occurs simultaneously
with a banana concept response (activated by some other stimulus such as
the sight of a banana shape), the network should strengthen the connection
between them so that later it can activate its banana concept in response
to the banana smell alone.
The unsupervised Hebb rule does just that by increasing the weight w ij be-
tween a neurons input p j and output a i in proportion to their product:
w ij q = w ij q 1 + a i q p j q . (15.6)
(See also Eq. (7.5).) The learning rate dictates how many times a stimu-
lus and response must occur together before an association is made. In the
network in Figure 15.2, an association will be made when w b = 0.5 ,
since then p = 1 will produce the response a = 1 , regardless of the value
0
of p .
Note that Eq. (15.6) uses only signals available within the layer containing
the weights being updated. Rules that satisfy this condition are called local
Local Learning learning rules. This is in contrast to the backpropagation rule, for example,
in which the sensitivity must be propagated back from the final layer. The
rules introduced in this chapter will all be local learning rules.
The unsupervised Hebb rule can also be written in vector form:
W q = W q 1 + a q p T q . (15.7)
As with all unsupervised rules, learning is performed in response to a se-

Training Sequence ries of inputs presented in time (the training sequence):
p 1 p 2 p Q . (15.8)
15-5
(Note that we are using the notation p q , instead of p q , in order to empha-

size the time-sequence nature of the inputs.) At each iteration, the output
a is calculated in response to the input p , and then the weights W are up-
dated with the Hebb rule.
2 Lets apply the unsupervised Hebb rule to the banana associator. The asso-
+2
ciator will start with the weight values determined in our previous exam-
ple, so that it will initially respond to the sight, but not the smell, of a
banana.
0
w = 1 w 0 = 0 (15.9)
The associator will be repeatedly exposed to a banana. However, while the

networks smell sensor will work reliably, the shape sensor will operate
only intermittently (on even time steps). Thus the training sequence will
consist of repetitions of the following two sets of inputs:
0 0
p 1 = 0 p 1 = 1 p 2 = 1 p 2 = 1 . (15.10)
0
The first weight w , representing the weight for the unconditioned stimu-
0
lus p , will remain constant, while w will be updated at each iteration, us-
ing the unsupervised Hebb rule with a learning rate of 1:
w q = w q 1 + a q p q . (15.11)
The output for the first iteration ( q = 1 ) is

0 0
a 1 = hardlim w p 1 + w 0 p 1 0.5 (15.12)
= hardlim 1 0 + 0 1 0.5 = 0 (no response) .
The smell alone did not generate a response. Without a response, the Hebb
rule does not alter w .
w 1 = w 0 + a 1 p 1 = 0 + 0 1 = 0 (15.13)
In the second iteration, both the bananas shape and smell are detected and
the network responds accordingly:
0 0
a 2 = hardlim w p 2 + w 1 p 2 0.5 (15.14)
= hardlim 1 1 + 0 1 0.5 = 1 (banana) .
Because the smell stimulus and the response have occurred simultaneous-
ly, the Hebb rule increases the weight between them.
w 2 = w 1 + a 2 p 2 = 0 + 1 1 = 1 (15.15)
When the sight detector fails again, in the third iteration, the network re-
sponds anyway. It has made a useful association between the smell of a ba-
nana and its response.
15-6
0 0
a 3 = hardlim w p 3 + w 2 p 3 0.5 (15.16)
= hardlim 1 0 + 1 1 0.5 = 1 (banana)
w 3 = w 2 + a 3 p 3 = 1 + 1 1 = 2 (15.17)
From now on, the network is capable of responding to bananas that are de-
tected either by sight or smell. Even if both detection systems suffer inter-
mittent faults, the network will be correct most of the time.
To experiment with the unsupervised Hebb rule, use the Neural Network De-
sign Demonstration Unsupervised Hebb Rule (nnd13uh).
We have seen that the unsupervised Hebb rule can learn useful associa-
tions. However, the Hebb rule, as defined in Eq. (15.6), has some practical
shortcomings. The first problem becomes evident if we continue to present
inputs and update w in the example above. The weight w will become ar-
bitrarily large. This is at odds with the biological systems that inspired the
Hebb rule. Synapses cannot grow without bound.
The second problem is that there is no mechanism for weights to decrease.
If the inputs or outputs of a Hebb network experience any noise, every
weight will grow (however slowly) until the network responds to any stim-
ulus.
Hebb Rule with Decay

One way to improve the Hebb rule is by adding a weight decay term (Eq.
(7.45)),
W q = W q 1 + a q p T q W q 1
= 1 W q 1 + a q p T q , (15.18)
Decay Rate where , the decay rate, is a positive constant less than one. As approach-
es zero, the learning law becomes the standard rule. As approaches one,
the learning law quickly forgets old inputs and remembers only the most
recent patterns. This keeps the weight matrix from growing without
bound. (The idea of filtering the weight changes was also discussed in
Chapter 12, where we called it momentum.)
max
The maximum weight value w ij is determined by . This value is found
by setting both a i and p j to a value of 1 for all q (to maximize learning) in
the scalar version of Eq. (15.18) and solving for the steady state weight (i.e.
when both new and old weights are equal).
15-7
w ij = 1 w ij + a i p j
w ij = 1 w ij + (15.19)

w ij = ---

2 Lets examine the operation of the Hebb rule with decay on our previous ba-
+2
nana associator problem. We will use a decay rate of 0.1. The first itera-
tion, where only the smell stimulus is presented, is the same:
a1 = 0 (no response) w 1 = 0 . (15.20)
The next iteration also produces identical results. Here both stimuli are
presented, and the network responds to the shape. Coincidence of the smell
stimulus and response create a new association:
a2 = 1 (banana) w2 = 1 . (15.21)
The results of the third iteration are not the same. The network has
learned to respond to the smell, and the weight continues to increase. How-
ever, this time the weight increases by only 0.9, instead of 1.0.
w 3 = w 2 + a 3 p 3 0.1w 2 = 1 + 1 1 0.1 1 = 1.9 (15.22)
The decay term limits the weights value, so that no matter how often the
max
association is reinforced, w will never increase beyond w ij .
max 1
w ij = --- = ------- = 10 (15.23)
0.1
The new rule also ensures that associations learned by the network will not
be artifacts of noise. Any small random increases will soon decay away.
Figure 15.3 displays the response of the Hebb rule, with and without decay,
for the banana recognition example. Without decay, the weight continues
to increase by the same amount each time the neuron is activated. When
decay is added, the weight exponentially approaches its maximum value
max
( w ij = 10 ).
To experiment with the Hebb rule with decay, use the Neural Network De-
sign Demonstrations Hebb with Decay(nnd13hd) and Effect of Decay Rate
(nnd13edr).
15-8
Simple Recognition Network
30 10
20
6
w
4
10 Hebb Rule Hebb with Decay
0 0
0 10 20 30 0 10 20 30
q q
Figure 15.3 Response of the Hebb Rule, With and Without Decay
The Hebb rule with decay does solve the problem of large weights. Howev-
er, it does so at a price. The environment must be counted on to occasionally
present all stimuli that have associations. Without reinforcement, associa-
tions will decay away.
To illustrate this fact, consider Eq. (15.18) if a i = 0 :
w ij q = 1 w ij q 1 . (15.24)
If = 0.1 , this reduces to
w ij q = 0.9 w ij q 1 . (15.25)
Therefore w ij will be decreased by 10% at each presentation for which

a i = 0 . Any association that was previously learned will eventually be lost.
We will discuss a solution to this problem in a later section.
Simple Recognition Network

So far we have considered only associations between scalar inputs and out-
puts. We will now examine a neuron that has a vector input. (See Figure
Instar 15.4.) This neuron, which is sometimes referred to as an instar, is the sim-
plest network that is capable of pattern recognition, as we will demonstrate
shortly.
15-9
p1 w1,1
w1,2 n a
p2
pR w1,R b
1
Figure 15.4 Instar

You will notice the similarities between the instar of Figure 15.4 and the
perceptron of Figure 4.2 (also the ADALINE of Figure 10.2 and the linear
associator of Figure 7.1). We give these networks different names, in part
for historical reasons (since they arose at different times and out of differ-
ent environments), and because they perform different functions and are
analyzed in different ways. For example, we will not directly consider the
decision boundary of the instar, although this was an important concept for
the perceptron. Instead, we will analyze the ability of the instar to recog-
nize a pattern, as with the neurons in the first layer of the Hamming net-
work. (See page 3-10.)
The input/output expression for the instar is
a = hardlim Wp + b = hardlim 1w T p + b . (15.26)
The instar will be active whenever the inner product between the weight
vector (row of the weight matrix) and the input is greater than or equal to
b :
1w
Tp b . (15.27)
From our discussion of the Hamming network on page 3-10, we know that
for two vectors of constant length, the inner product will be largest when
they point in the same direction. We can also show this using Eq. (5.15):
1w
Tp = 1w p cos b , (15.28)
where is the angle between the two vectors. Clearly the inner product is
maximized when the angle is 0. If p and 1w have the same length
( p = 1w ), then the inner product will be largest when p = 1w .
Based on these arguments, the instar of Figure 15.4 will be active when p
is close to 1w . By setting the bias b appropriately, we can select how close
the input vector must be to the weight vector in order to activate the instar.
15-10
Instar Rule
If we set
b = 1w p , (15.29)
then the instar will only be active when p points in exactly the same direc-
tion as 1w ( = 0 ). Thus, we will have a neuron that recognizes only the
pattern 1w .
If we would like the instar to respond to any pattern near 1w ( small),
then we can increase b to some value larger than 1w p . The larger the
value of b , the more patterns there will be that can activate the instar, thus
making it the less discriminatory.
We should note that this analysis assumes that all input vectors have the
same length (norm). We will revisit the question of normalization in Chap-
ters 16, 18 and 19.
We can now design a vector recognition network if we know which vector
we want to recognize. However, if the network is to learn a vector without
supervision, we need a new rule, since neither version of the Hebb rule pro-
duces normalized weights.
Instar Rule
One problem of the Hebb rule with decay was that it required stimuli to be
repeated or associations would be lost. A better rule might allow weight de-
cay only when the instar is active ( a 0 ). Weight values would still be lim-
ited, but forgetting would be minimized. Consider again the original Hebb
rule:
w ij q = w ij q 1 + a i q p j q . (15.30)
To get the benefits of weight decay, while limiting the forgetting problem,
a decay term can be added that is proportional to a i q :
w ij q = w ij q 1 + a i q p j q a i q w ij q 1 (15.31)
We can simplify Eq. (15.31) by setting equal to (so new weight values
are learned at the same rate old values decay) and gathering terms.
w ij q = w ij q 1 + a i q p j q w ij q 1 (15.32)
Instar Rule This equation, called the instar rule, can also be rewritten in vector form:
iw q = iw q 1 + a i q p q iw q 1 . (15.33)
The performance of the instar rule can be best understood if we consider

the case where the instar is active ( a i = 1 ). Eq. (15.33) can then be written
15-11
iw q = iw q 1 + p q iw q 1
= 1 iw q 1 + p q . (15.34)
This operation is displayed graphically in Figure 15.5.
p(q)
w(q)
i
w(q - 1)
i
Figure 15.5 Graphical Representation of the Instar Rule

When the instar is active, the weight vector is moved toward the input vec-
tor along a line between the old weight vector and the input vector. The dis-
tance the weight vector moves depends on the value of . When = 0 , the
new weight vector is equal to the old weight vector (no movement). When
= 1 , the new weight vector is equal to the input vector (maximum move-
ment). If = 0.5 , the new weight vector will be halfway between the old
weight vector and the input vector.
One useful feature of the instar rule is that if the input vectors are normal-
ized, then iw will also be normalized once it has learned a particular vector
p . We have found a rule that not only minimizes forgetting, but results in
normalized weight vectors, if the input vectors are normalized.
2 Lets apply the instar rule to the network in Figure 15.6. It has two inputs:
+2
one indicating whether a fruit has been visually identified as an orange
(unconditioned stimulus) and another consisting of the three measure-
ments taken of the fruit (conditioned stimulus).
The output of this network is
Fruit 0 0
a = hardlim w p + Wp + b . (15.35)
The elements of input p will be constrained to 1 values, as defined in

Sight Measure
Chapter 3 (Eq. (3.2)). This constraint ensures that p is a normalized vector
0
with a length of p = 3 . The definitions of p and p are
Network
1 orange detected visually shape

0
p = p = texture . (15.36)
Orange? 0 orange not detected
weight
15-12
Instar Rule
2
The bias b is 2 , a value slightly more positive than p = 3 . (See Eq.
(15.29).)
Sight of orange p0 w0 = 3
w1,1 n a Orange?
Measured shape p1
Measured texture p2 b = -2
Measured weight p3 w1,3
1
a = hardlim (w0 p0 + W p + b)
Figure 15.6 Orange Recognizer

We would like the network to have a constant association between the sight
0
of an orange and its response, so w will be set greater than b . But ini-
tially, the network should not respond to any combination of fruit measure-
ments, so the measurement weights will start with values of 0.
0 T
w = 3 , W 0 = 1w 0 = 0 0 0 (15.37)
The measurement weights will be updated with the instar rule, using a
learning rate of = 1 .
1w q = 1w q 1 + a q p q 1w q 1 (15.38)
The training sequence will consist of repeated presentations of an orange.

The measurements will be given every time. However, in order to demon-
strate the operation of the instar rule, we will assume that the visual sys-
tem only operates correctly on even time steps, due to a fault in its
construction.
1 1
0 0
p 1 = 0 p 1 = 1 p 2 = 1 p 2 = 1 (15.39)

1 1
Because W initially contains all zeros, the instar does not respond to the
measurements of an orange in the first iteration.
15-13
0 0
a 1 = hardlim w p 1 + Wp 1 2
1
(15.40)
a 1 = hardlim 3 0 + 0 0 0 1 2 = 0 (no response)

1
Since the neuron did not respond, its weights 1w are not altered by the in-
star rule.
1w 1 = 1w 0 + a 1 p 1 1w 0 (15.41)
0 1 0 0

= 0 + 0 1 0 = 0

0 1 0 0
However, the neuron does respond when the orange is identified visually,
in addition to being measured, in the second iteration.
0 0
a 2 = hardlim w p 2 + Wp 2 2 (15.42)
1

= hardlim 3 1 + 0 0 0 1 2 = 1 (orange)

1
The result is that the neuron learns to associate the oranges measurement
vector with its response. The weight vector 1w becomes a copy of the orange
measurement vector.
1w 2 = 1w 1 + a 2 p 2 1w 1 (15.43)
0 1 0 1

= 0 + 1 1 0 = 1

0 1 0 1
The network can now recognize the orange by its measurements. The neu-
ron responds in the third iteration, even though the visual detection system
failed again.
15-14
Instar Rule
0 0
1
(15.44)
a 3 = hardlim 3 0 + 1 1 1 1 2 = 1 (orange)

1
Having completely learned the measurements, the weights stop changing.

(A lower learning rate would have required more iterations.)
1w 3 = 1w 2 + a 3 p 3 1w 2 (15.45)
15
1 1 1 1

= 1 + 1 1 1 = 1

1 1 1 1
The network has learned to recognize an orange by its measurements, even

when its visual detection system fails.
To experiment with the instar rule, use the Neural Network Design Demon-
strations Instar (nnd13is) and Graphical Instar (nnd13gis).
Kohonen Rule
At this point it is appropriate to introduce another associative learning
Kohonen Rule rule, which is related to the instar rule. It is the Kohonen rule:
iw q = iw q 1 + p q iw q 1 for i X q . (15.46)
Like the instar rule, the Kohonen rule allows the weights of a neuron to
learn an input vector and is therefore suitable for recognition applications.
Unlike the instar rule, learning is not proportional to the neurons output
a i q . Instead, learning occurs when the neurons index i is a member of
the set X q .
If the instar rule is applied to a layer of neurons whose transfer function
only returns values of 0 or 1 (such as hardlim), then the Kohonen rule can
be made equivalent to the instar rule by defining X q as the set of all i
such that a i q = 1 . The advantage of the Kohonen rule is that it can also
be used with other definitions. It is useful for training networks such as the
self-organizing feature map, which will be introduced in Chapter 16.
15-15
Simple Recall Network

We have seen that the instar network (with a vector input and a scalar out-
put) can perform pattern recognition by associating a particular vector
Outstar stimulus with a response. The outstar network, shown in Figure 15.7, has
a scalar input and a vector output. It can perform pattern recall by associ-
ating a stimulus with a vector response.
The input-output expression for this network is
a = satlins Wp . (15.47)
The symmetric saturating function satlins was chosen because this net-
work will be used to recall a vector containing values of -1 or 1.
Input Linear Layer
n1 a1
w1,1
w2,1 n2 a2
p
wS,1 nS aS

a = satlins (Wp)
Figure 15.7 Outstar Network
If we would like the network to associate a stimulus (an input of 1) with a
particular output vector a , we can simply set W (which contains only a
single column vector) equal to a . Then, if p is 1, the output will be a :
a = satlins Wp = satlins a 1 = a . (15.48)
(This assumes that the elements of a are less than or equal to 1 in mag-
nitude.)
Note that we have created a recall network by setting a column of the
weight matrix to the desired vector. Earlier we designed a recognition net-
work by setting a row of the weight matrix to the desired vector.
We can now design a network that can recall a known vector a , but we
need a learning rule if the network is to learn a vector without supervision.
We will describe such a learning rule in the next section.
15-16
Outstar Rule
Outstar Rule
To derive the instar rule, forgetting was limited by making the weight de-
cay term of the Hebb rule proportional to the output of the network, a i .
Conversely, to obtain the outstar learning rule, we make the weight decay
term proportional to the input of the network, p j :
w ij q = w ij q 1 + a i q p j q p j q w ij q 1 . (15.49)
If we set the decay rate equal to the learning rate and collect terms,
we get
w ij q = w ij q 1 + a i q w ij q 1 p j q . (15.50)
The outstar rule has properties complimentary to the instar rule. Learning
occurs whenever p j is nonzero (instead of a i ). When learning occurs, col-
umn w j moves toward the output vector.
Outstar Rule As with the instar rule, the outstar rule can be written in vector form
w j q = w j q 1 + a q w j q 1 p j q , (15.51)
where w j is the jth column of the matrix W .

2 To test the outstar rule we will train the network shown in Figure 15.8.
+2
Inputs Linear Layer
a1 Recalled shape
w1,10 = 1 n1
Measured shape p 0
1
Measured texture p20 w2,2
0
=1
Measured weight p30 =1 n2 a2 Recalled texture
w3,3
0

w1,1
a3 Recalled weight
n3
Identified Pineapple p
w3,1
a = satlins (W0p0 + Wp)
Figure 15.8 Pineapple Recaller
15-17
The outputs of the network are calculated as follows:

0 0
a = satlins W p + Wp , (15.52)
where
Fruit 0
1 0 0
W = 0 1 0 . (15.53)
0 0 1
Sight Measure
0
The networks two inputs provide it with measurements p taken on a fruit
Network (unconditioned stimulus), as well as a signal p indicating a pineapple has
been identified visually (conditioned stimulus).
Measurements? 0
shape 1 if a pineapple can be seen
p = texture p = (15.54)
0 otherwise
weight
The networks output is to reflect the measurements of the fruit currently

being examined, using whatever inputs are available.
0
The weight matrix for the unconditioned stimulus, W , is set to the identi-
0
ty matrix, so that any set of measurements p (with 1 values) will be cop-
ied to the output a . The weight matrix for the conditioned stimulus, W , is
set to zero initially, so that a 1 on p will not generate a response. W will
be updated with the outstar rule using a learning rate of 1:
w j q = w j q 1 + a q w j q 1 p q . (15.55)
The training sequence will consist of repeated presentations of the sight

and measurements of a pineapple. The pineapple measurements are
1
pineapple
p = 1 . (15.56)
1
However, due to a fault in the measuring system, measured values will

only be available on even iterations.
0 1
0 0
p 1 = 0 p 1 = 1 p 2 = 1 p 2 = 1 (15.57)

0 1
15-18
Outstar Rule
In the first iteration, the pineapple is seen, but the measurements are un-
available.
0 0
a 1 = satlins W p 1 + Wp 1 , (15.58)
0 0 0

a 1 = satlins 0 + 0 1 = 0 (no response) (15.59)

0 0 0
The network sees the pineapple but cannot output proper measurements,
because it has not learned them and the measurement system is not work-
ing. The weights remain unchanged after being updated.
15
0 0 0
0
w 1 1 = w 1 0 + a 1 w 1 0 p 1 = 0 + 0 0 1 = 0 (15.60)

0 0 0 0
In the second iteration the pineapple is seen, and the measurements are
taken properly.
1 0 1

a 2 = satlins 1 + 0 1 = 1 (measurements given) (15.61)

1 0 1
The measurements are available, so the network outputs them correctly.

The weights are then updated as follows:
w 1 2 = w 1 1 + a 2 w 1 1 p 2
0 1 0
1
= 0 + 1 0 1 = 1 .

0 1 0 1
(15.62)
Since the sight of the pineapple and the measurements were both avail-
able, the network forms an association between them. The weight matrix
is now a copy of the measurements, so they can be recalled later.
15-19
In iteration three, measurements are unavailable once again, but the out-
put is
0 1 1

a 3 = satlins 0 + 1 1 = 1 (measurements recalled) . (15.63)

0 1 1
The network is now able to recall the measurements of the pineapple when
it sees it, even though the measurement system fails. From now on, the
weights will no longer change values unless a pineapple is seen with differ-
ent measurements.
w 1 3 = w 1 2 + a 2 w 1 2 p 2
1 1 1 1
= 1 + 1 1 1 = 1

1 1 1 1
(15.64)
To experiment with the outstar rule with decay, use the Neural Network De-
sign Demonstration Outstar Rule (nnd13os).
In Chapter 19 we will investigate the ART networks, which use both the
instar and the outstar rules.
15-20
Summary of Results
Summary of Results
Association
An association is a link between the inputs and outputs of a network so
that when a stimulus A is presented to the network, it will output a re-
sponse B.
Associative Learning Rules

W q = W q 1 + a q p T q
15
Hebb Rule with Decay
W q = 1 W q 1 + a q p T q
Instar
p1 w1,1
w1,2 n a
p2
pR w1,R b
1
a = hardlim 1w T p + b
The instar is activated for 1w T p = 1w p cos b ,
where is the angle between p and 1w .
Instar Rule
iw q = iw q 1 + a i q p q iw q 1
iw q = 1 w q 1 + p q if a i q = 1
15-21
p(q)
iw(q)
iw(q - 1)
Graphical Representation of the Instar Rule ( a i q = 1 )
Kohonen Rule
iw q = iw q 1 + p q iw q 1 for i X q
Outstar
Input Linear Layer
n1 a1
w1,1
w2,1 n2 a2
p
wS,1 nS aS

a = satlins (Wp)
Outstar Rule
w j q = w j q 1 + a q w j q 1 p j q
15-22
Solved Problems
Solved Problems
P15.1 In Eq. (15.19) the maximum weight for the Hebb rule with decay
was calculated, assuming that p j and a i were 1 at every time step.
Calculate the maximum weight resulting if p j and a i alternate to-
gether between values of 0 and 1.
We begin with the scalar version of the Hebb rule with decay:
w ij q = 1 w ij q 1 + a i q p j q .
We can rewrite this expression twice using q to index the weight values as
the weight is updated over two time steps.
w ij q + 1 = 1 w ij q + a i q p j q
w ij q + 2 = 1 w ij q + 1 + a i q + 1 p j q + 1
By substituting the first equation into the second, we get a single expres-
sion showing how w ij is updated over two time steps.
w ij q + 2 = 1 1 w ij q + a i q p j q + a i q + 1 p j q + 1
At this point we can substitute values for p j and a i . Because we are looking
for a maximum weight, we will set p j q and a i q to 0, and p j q + 1 and
a i q + 1 to 1. This will mean that the weight decreases in the first time
step, and increases in the second, ensuring that w ij q + 2 is the maximum
of the two weights. If we solve for w ij q + 2 , we obtain
2
w ij q + 2 = 1 w ij q + .
Assuming that w ij will eventually reach a steady state value, we can find
max
it by setting both w ij q + 2 and w ij q equal to w ij and solving
max 2 max
w ij = 1 w ij +,
max
w ij = ---------------2- .
2
2+2
ans = We can use MATLAB to make a plot of this relationship. The plot will show
4 learning rates and decay rates at intervals of 0.025.
lr = 0:0.025:1;
dr = 0.025:0.025:1;
Here are the commands for creating a mesh plot of the maximum weight,
as a function of the learning and decay rate values.
15-23
[LR,DR] = meshgrid(dr,lr);
MW = LR ./ (DR .* (2 - DR));
mesh(DR,LR,MW);
max
The plot shows that w ij approaches infinity as the decay rate becomes
small with respect to the learning rate (see Figure P15.1).
20
15
Maximum weight
10
0
0
0.2 1
0.4 0.8
0.6 0.6
0.4
0.8 0.2
1 0
Learning Rate
Decay Rate
max
Figure P15.1 Maximum Weight w ij
P15.2 Retrain the orange recognition network on page 13-13 using the in-
star rule with a learning rate of 0.4. Use the same training se-
quence. How many time steps are required for the network to
learn to recognize an orange by its measurements?
Here is the training sequence. It is to be repeated until the network can re-
T
spond to the orange measurements ( p = 1 1 1 ), even when the visual
0
detection system fails ( p = 0 ).
1 1
0 0
p 1 = 0 p 1 = 1 p 2 = 1 p 2 = 1

1 1
2+2
ans = We will use MATLAB to make the calculations. These two lines of code set
4 the weights to their initial values.
w0 = 3;
W = [0 0 0];
15-24
Solved Problems
We can then simulate the first time step of the network.

p0 = 0;
p = [1; -1; -1];
a = hardlim(w0*p0+W*p-2)
a =
0
The neuron does not yet recognize the orange, so its output is 0. The
weights do not change when updated with the instar rule.
W = W + 0.4*a*(p-W)
W =
0 0 0
The neuron begins learning the measurements in the second iteration.

15
p0 = 1;
p = [1; -1; -1];
a =
1
W = W + 0.4*a*(p-W)
W =
0.4000 -0.4000 -0.4000
But the association is still not strong enough for a response in the third it-
eration.
p0 = 0;
p = [1; -1; -1];
a =
0
W = W + 0.4*a*(p-W)
W =
0.4000 -0.4000 -0.4000
Here are the results of the fourth iteration:

a =
1
W =
0.6400 -0.6400 -0.6400
the fifth iteration:
15-25
a =
0
W =
0.6400 -0.6400 -0.6400
and the sixth iteration:

a =
1
W =
0.7840 -0.7840 -0.7840 .
By the seventh iteration the network is able to recognize the orange by its
measurements alone.
p0 = 0;
p = [1; -1; -1];
a =
1
W = W + 0.4*a*(p-W)
W =
0.8704 -0.8704 -0.8704
Due to the lower learning rate, the network had to experience the measure-
ments paired with its response three times (the even numbered iterations)
before it developed a strong association between them.
P15.3 Both the recognition and recall networks used in this chapters ex-
amples could only learn a single vector. Draw the diagram and de-
termine the parameters of a network capable of recognizing and
responding to the following two vectors:
5 5
p1 = 5 p2 = 5 .
5 5
The network should only respond when an input vector is identi-

cal to one of these vectors.
We know the network must have three inputs, because it must recognize
three-element vectors. We also know that it will have two outputs, one out-
put for each response.
Such a network can be obtained by combining two instars into a single lay-
er, as in Figure P15.2.
15-26
Solved Problems
Inputs Hard Limit Layer
w1,1 n1 a1
p1
b1
p2
1
n2 a2
p3
w2,3
b2
1
Figure P15.2 Two-Vector Recognition Network

We now set the weights 1w of the first neuron equal to p 1 , so that its net
input will be at a maximum when an input vector points in the same direc-
tion as p 1 . Likewise, we will set 2w to p 2 so that the second neuron is most
sensitive to vectors in the direction of p 2 .
Combining the weight vectors gives us the weight matrix
T T
1w p1 5 5 5 .
W = = =
T T 5 5 5
2w p2
(Note that this is the same manner in which we determined the weight ma-
trix for the first layer of the Hamming network. In fact, the first layer of
the Hamming network is a layer of instars. More about that in the next
chapter.)
The lengths of p 1 and p 2 are the same:
2 2 2
p1 = p2 = 5 + 5 + 5 = 75 .
To ensure that only an exact match between an input vector and a stored
vector results in a response, both biases are set as follows (Eq. (15.29)):
2
b1 = b2 = p1 = 75 .
2+2
ans = We can use MATLAB to test that the network does indeed respond to p 1 .
4
W = [5 -5 5; -5 5 5];
b = [-75; -75];
15-27
p1 = [5; -5; 5];

a = hardlim(W*p1+b)
a =
1
0
The first neuron responded, indicating that the input vector was p 1 . The
second neuron did not respond, because the input was not p 2 .
We can also check that the network does not respond to a third vector p 3
that is not equal to either of the stored vectors.
p3 = [-5; 5; -5];
a = hardlim(W*p3+b)
a =
0
Neither neuron recognizes this new vector, so they both output 0.
P15.4 A single instar is being used for pattern recognition. Its weights
and bias have the following values:
T
W = 1w = 1 1 1 b = 2 .
How close must an input vector (with a magnitude of 3 ) be to the

weight vector for the neuron to output a 1? Find a vector that oc-
curs on the border between those vectors that are recognized and
those vectors that are not.
We begin by writing the expression for the neurons output.
T
According to the definition of hardlim , a will be 1 if and only if the inner

T
product between 1w and p is greater than or equal to b (Eq. (15.28)):
1w
Tp = 1w p cos b .
We can find the maximum angle between 1w and p that meets this condi-
tion by substituting for the norms and solving
3 3 cos 2
cos --- = 48.19 .

1 2
3
15-28
Solved Problems
To find a borderline vector with magnitude 3 , we need a vector p that

meets the following conditions:
2 2 2
p1 = p1 + p2 + p3 = 3,
T
1w p = w1 p1 + w2 p2 + w3 p3 b = p1 p2 p3 2 = 0 .
Since we have three variables and only two constraints, we can set the
third variable p 1 to 0 and solve
2 2 2 2 2
p1 + p2 + p3 = 3 p2 + p3 = 3 ,
p1 p2 p3 2 p2 + p3 = 2 ,
15
2 2 2 2
p 2 + p 3 = p 2 + p 3 + 2p 2 p 3 = 2 = 4 ,
3 + 2p 2 p 3 = 4 p 2 p 3 = 0.5 ,
2 2
p 2 p 2 + p 3 = p 2 + p 2 p 3 = p 2 + 0.5 = p 2 2 = 2p 2 .
After a little work we find that there are two possible values for p 2 :
2
p 2 + 2p 2 + 0.5 = 0 ,
p 2 = 1 0.5 .
It turns out that if we pick p 2 to be one of these values, then p 3 will take
on the other value.
p 2 + p 3 = 1 0.5 + p 3 = 2
p3 = 1
0.5
Therefore, the following vector p is just the right distance from w to be rec-
ognized.
0
p = 1 + 0.5
1 0.5
We can test it by presenting it to the network.
15-29
T
0

a = hardlim 1 1 1 1 + 0.5 2

1 0.5
a = hardlim 0 = 1
The vector p does indeed result in a net input of 0 and is therefore on the
boundary of the instars active region.
P15.5 Consider the instar network shown in Figure P15.3. The training
sequence for this network will consist of the following inputs:
0 1 0 1
p 1 = 0 p 1 = p 2 = 1 p 2 = .
1 1
These two sets of inputs are repeatedly presented to the network

until the weight matrix W converges.
i. Perform the first four iterations of the instar rule, with
learning rate = 0.5 . Assume that the initial W matrix is set
to all zeros.
ii. Display the results of each iteration of the instar rule in
graphical form (as in Figure 15.5).
p0 w0 = 3
w1,1 n a
p1
b = -2
p2 w1,2
1
Figure P15.3 Instar Network for Problem P15.5

i. Because W initially contains all zeros, the instar does not respond to
the measurements in the first iteration.
15-30
Solved Problems
0 0

a 1 = hardlim 3 0 + 0 0 1 2 = 0
1
The neuron did not respond. Therefore its weights 1w are not altered by the
instar rule.
1w 1 = 1w 0 + 0.5a 1 p 1 1w 0

= 0 + 0 1 0 = 0
0 1 0 0
15
Because the unconditioned stimulus appears on the second iteration, the
instar does respond.
0 0

a 2 = hardlim 3 1 + 0 0 1 2 = 1
1
The neuron did respond, and its weights 1w are updated by the instar rule.
1w 2 = 1w 1 + 0.5a 2 p 2 1w 1

= 0 + 0.5 1 0 = 0.5
0 1 0 0.5
On the third iteration, the unconditioned stimulus is not presented, and

the weights have not yet converged close enough to the input pattern.
Therefore, the instar does not respond.
0 0

a 3 = hardlim 3 0 + 0.5 0.5 1 2 = 0
1
Since the neuron did not respond, its weights are not updated.
15-31
1w 3 = 1w 2 + 0.5a 3 p 3 1w 2

= 0.5 + 0 1 0.5 = 0.5
0.5 1 0.5 0.5
Because the unconditioned stimulus again appears on the fourth iteration,

the instar does respond.
0 0

a 4 = hardlim 3 1 + 0.5 0.5 1 2 = 1
1
Since the instar was activated, its weights are updated.
1w 4 = 1w 3 + 0.5a 4 p 4 1w 3

= 0.5 + 0.5 1 0.5 = 0.75
0.5 1 0.5 0.75
This completes the fourth iteration. If we continue this process, 1w will

converge to p .
ii. Note that the weights are only updated (instar active) on iterations 2
and 4. Recall from Eq. (15.34) that when the instar is active, the learning
rule can be written
1w q = 1w q 1 + p q 1w q 1 = 1 1w q 1 + p q .
When the instar is active, the weight vector is moved toward the input vec-
tor along a line between the old weight vector and the input vector. Figure
P15.4 displays the movement of the weight vector for this problem. The
weights were updated on iterations 2 and 4. Because = 0.5 , whenever
the instar is active the weight vector moves halfway from its current posi-
tion toward the input vector.
1w q = 0.5 1w q 1 + 0.5 p q
15-32
Solved Problems
p
1 w(4)
1 w(2)
1 w(0)
Figure P15.4 Instar Rule Example
15
15-33
Epilogue
In this chapter we introduced some simple networks capable of forming
associations. We also developed and studied several learning rules that
allowed networks to create new associations. Each rule operated by
strengthening an association between any stimulus and response that
occurred simultaneously.
The simple associative networks and learning rules developed in this chap-
ter are useful in themselves, but they are also important building blocks for
more powerful networks. In this chapter we introduced two networks, and
associated learning rules, that will be fundamental for the development of
important networks in the next three chapters: the instar and the outstar.
The instar is a network that is trained to recognize a pattern. The outstar
is a network that is trained to recall a pattern. We will use layers of instars
in Chapters 16 and 18 to perform pattern recognition. These networks are
very similar to the Hamming network of Chapter 3, whose first layer was,
in fact, a layer of instars. In Chapter 19 we will introduce a more complex
network, which combines both instars and outstars in order to produce sta-
ble learning.
15-34
Further Reading
Further Reading
197220, 1972.
Anderson has proposed a linear associator model for asso-
ciative memory. The model was trained, using a generali-
zation of the Hebb postulate, to learn an association
between input and output vectors. The physiological plau-
sibility of the network was emphasized. Kohonen published
a closely related paper at the same time [Koho72], although
the two researchers were working independently.
[Gros68] S. Grossberg, Some physiological and biochemical conse-
quences of psychological postulates, Proceedings of the Na-
15
tional Academy of Sciences, vol. 60, pp. 758765, 1968.
This article describes early mathematical models (nonlin-
ear differential equations) of associative learning. It syn-
thesizes psychological, mathematical and physiological
ideas.
[Gros82] S. Grossberg, Studies of Mind and Brain, Boston: D. Reidel
Publishing Co., 1982.
This book is a collection of Stephen Grossberg papers from
the period 1968 through 1980. It covers many of the funda-
mental concepts which are used in later Grossberg net-
works, such as the Adaptive Resonance Theory networks.
[Hebb49] D. O. Hebb, The Organization of Behavior, New York:
Wiley, 1949.
The main premise of this seminal book was that behavior
could be explained by the action of neurons. In it, Hebb pro-
posed one of the first learning laws, which postulated a
mechanism for learning at the cellular level.
pendently.
15-35
[Koho87] T. Kohonen, Self-Organization and Associative Memory,

2nd Ed., Berlin: Springer-Verlag, 1987.
This book introduces the Kohonen rule and several net-
works that use it. It provides a complete analysis of linear
associative models and gives many extensions and exam-
ples.
[Leib90] D. Lieberman, Learning, Behavior and Cognition, Bel-
mont, CA: Wadsworth, 1990.
Leibermans text forms an excellent introduction to behav-
ioral psychology. This field is of interest to anyone looking
to model human (or animal) learning with neural net-
works.
15-36
Exercises
Exercises
E15.1 The network shown in Figure E15.1 is to be trained using the Hebb rule
with decay, using a learning rate of 0.3 and a decay rate of 0.1.
p0 w0 = 1
n a

p w b = -0.8
1
a = hardlim (w0p0 + wp + b)
Figure E15.1 Associative Network

0
i. If w is initially set to 0, and w and b remain constant (with the val-
ues shown in Figure E15.1), how many consecutive presentations of
the following training set are required before the neuron will re-
spond to the test set? Make a plot of w versus iteration number.
0 0
Training set: p = 1 p = 1 Test set: p = 0 p = 1
ii. Assume that w has an initial value of 1. How many consecutive pre-
sentations of the following training set are required before the neu-
ron will no longer be able to respond to the test set? Make a plot of
w versus iteration number.
0 0
Training set: p = 0 p = 0 Test set: p = 0 p = 1
E15.2 For Exercise E15.1 part (i), use Eq. (15.19) to determine the steady state
value of w . Verify that this answer agrees with your plot from Exercise
E15.1 part (i).
E15.3 Repeat Exercise E15.1, but this time use the Hebb rule without decay
( = 0 ).
E15.4 The following rule looks similar to the instar rule, but it behaves quite dif-
ferently:
old
w ij = a i p j + w ij
15-37
i. Determine the conditions under which the w ij is nonzero.

ii. What value does the weight approach when w ij is nonzero?
iii. Can you think of a use for this rule?
E15.5 The instar shown in Figure E15.2 is to be used to recognize a vector.
p0 w0 = 1
w1,1 n a
p1
w1,2 b = -0.5
p2
1
a = hardlim (w0p0 + Wp + b)
Figure E15.2 Vector Recognizer

i. Train the network with the instar rule on the following training se-
quence. Apply the instar rule to the second inputs weights only
2+2
(which should be initialized to zeros), using a learning rate of 0.6.
ans = The other weight and the bias are to remain constant at the values
4 in the figure. (You may wish to use MATLAB to perform the calcu-
lations.)
0 0.174 p 0 2 = 0 p 2 = 0.174
p 1 = 1 p 1 =
0.985 0.985
0 0.174 p 0 4 = 0 p 4 = 0.174
p 3 = 1 p 3 =
0.985 0.985
0 0.174 p 0 6 = 0 p 6 = 0.174
p 5 = 1 p 5 =
0.985 0.985
ii. What were your final values for W ?

iii. How do these final values compare with the vectors in the training
sequence?
15-38
Exercises
iv. What magnitude would you expect the weights to have after train-
ing, if the network were trained for many more iterations of the
same training sequence?
E15.6 Consider the instar network shown in Figure E15.3. The training sequence
for this network will consist of the following inputs:
0 1 p 0 2 = 1 p 2 = 1 .
p 1 = 0 p 1 =
1 1
These two sets of inputs are repeatedly presented to the network until the
weight matrix W converges.
i. Perform the first eight iterations of the instar rule, with learning
rate = 0.25 . Assume that the initial W matrix is set to
W = 1 0 .
ii. Display the results of each iteration of the instar rule in graphical
form (as in Figure 15.5).
p0 w0 = 3
w1,1 n a
p1
b = -2
p2 w1,2
1
Figure E15.3 Instar Network for Exercise E15.6
E15.7 Draw a diagram of a network capable of recognizing three different four-

element vectors (of 1 values) when given different stimuli (of value 1).
i. How many inputs does your network have? How many outputs?
What transfer function did you use?
ii. Choose values for the networks weights so that it can recognize
each of the following vectors:
15-39
1 1 1
p1 = 1
p2 = 1 p3 = 1

1 1 1
1 1 1
iii. Choose an appropriate value for the biases. Explain your choice.
iv. Test the network with one of the vectors above. Was its response
correct?
v. Test the network with the following vector.
1
p1 = 1
1
1
Why did it respond the way it did?
E15.8 This chapter included an example of a recognition network that initially

used a visual system to identify oranges. At first the network needed the
visual system to tell it when an orange was present, but eventually it
learned to recognize oranges from sensor measurements.
i. Let us replace the visual system with a person. Initially, the net-
work would depend on a person to tell it when an orange was
present. Would you consider the network to be learning in a super-
vised or unsupervised manner?
ii. In what ways would the input from a person resemble the targets
used to train supervised networks in earlier chapters?
iii. In what ways would it differ?
E15.9 The network shown in Figure E15.4 is installed in an elevator used by

G
three senior executives in a plush high-security corporate building. It has
Scan buttons marked 1 through 4 for four floors above the ground floor. When
an executive enters the elevator on the ground floor, it determines which
person it is with a retinal scan, and then uses the network to select the floor
Push where that person is most likely to go to. If the guess is incorrect, the per-
1 2 son can push a different button at any time, but if the network is correct, it
3 4
will save an important executive the effort of pushing a button.
15-40
Exercises
Input Sym. Hard Limit Layer
p0
Floor Code 2x1
W0
2x2
a Floor?
2 2x1
p n
Retinal Scan W
3x1 2x1
2x3
3
1 b 2
2x1
a = hardlims (W0 p0 + W p + b)
Figure E15.4 Elevator Network

15
Push Scan The networks input/output function is
1 2 0 0
3 4 a = hardlims W p + Wp + b .
0
The first input p provides the network with a floor code, if a button has
been pushed.
Network
p 1 = 1 (1st floor) 1 (2nd floor)
0 0
p2 =
1 1
Floor?
p 3 = 1 (3rd floor) p 4 = 1 (4th floor)
0 0
1 1
If no button is pushed, then no code is given.
p 0 = 0 (no button pushed)

0
The first input is weighted with an identity matrix, and the biases are set
to -0.5, so that if a button is pushed the network will respond with its code.
W = I , b = 0.5
0
0.5
The second input is always available. It consists of three elements that rep-
resent the three executives:
15-41
1 0 0
p 1 = 0 (President) , p 2 = 1 (Vice-President) , p 3 = 0 (Chairman) .
0 0 1
The network learns to recall the executives favorite floors by updating the
second set of weights using the outstar rule (using a learning rate of 0.6).
Initially those weights are set to zero:
W = 0 0 0 .
0 0 0
2+2
ans = i. Use MATLAB to simulate the network for the following sequence of
4 events:
President pushes 4, Vice-President pushes 3,
Chairman pushes 1, Vice-President pushes 3,
Chairman pushes 2, President pushes 4.
In other words, train the network on the following sequence:
0 0 0 0 0 0
p = p 4 p = p 1 , p = p 3 p = p 2 , p = p 1 p = p 3 ,
0 0 0 0 0 0
p = p 3 p = p 2 , p = p 2 p = p 3 , p = p 4 p = p 1 .
ii. What are the final weights?

iii. Now continue simulating the network on these events:
President does not push a button,
Vice-President does not push a button,
Chairman does not push a button.
iv. Which floors did the network take each executive to?
v. If the executives were to push the following buttons many times,
what would you expect the resulting weight matrix to look like?
President pushes 3,
Vice-President pushes 2,
Chairman pushes 4.
15-42
Objectives
Objectives 16-1
Hamming Network 16-3
Layer 1 16-3
Layer 2 16-4
Competitive Layer 16-5
Competitive Learning 16-7
Problems with Competitive Layers 16-9
Competitive Layers in Biology 16-10
Self-Organizing Feature Maps 16-12
Improving Feature Maps 16-15
Learning Vector Quantization 16-16
LVQ Learning 16-18
Improving LVQ Networks (LVQ2) 16-21
Epilogue 16-37
Exercises 16-39
Objectives
The Hamming network, introduced in Chapter 3, demonstrated one tech-
nique for using a neural network for pattern recognition. It required that
the prototype patterns be known beforehand and incorporated into the net-
work as rows of a weight matrix.
In this chapter we will discuss networks that are very similar in structure
and operation to the Hamming network. Unlike the Hamming network,
however, they use the associative learning rules of Chapter 15 to adaptive-
ly learn to classify patterns. Three such networks are introduced in this
chapter: the competitive network, the feature map and the learning vector
quantization (LVQ) network.
16-1
Theory and Examples

The Hamming network is one of the simplest examples of a competitive
network. The neurons in the output layer of the Hamming network com-
pete with each other to determine a winner. The winner indicates which
prototype pattern is most representative of the input pattern. The compe-
tition is implemented by lateral inhibition a set of negative connections
between the neurons in the output layer. In this chapter we will illustrate
how this competition can be combined with the associative learning rules
of Chapter 15 to produce powerful self-organizing (unsupervised) net-
works.
As early as 1959, Frank Rosenblatt created a simple spontaneous classi-
fier, an unsupervised network based on the perceptron, which learned to
classify input vectors into two classes with roughly equal members.
In the late 1960s and early 1970s, Stephen Grossberg introduced many
competitive networks that used lateral inhibition to good effect. Some of
the useful behaviors he obtained were noise suppression, contrast-en-
hancement and vector normalization. His networks will be examined in
Chapters 18 and 19.
In 1973, Christoph von der Malsburg introduced a self-organizing learning
rule that allowed a network to classify inputs in such a way that neighbor-
ing neurons responded to similar inputs. The topology of his network mim-
icked, in some ways, the structures previously found in the visual cortex of
cats by David Hubel and Torten Wiesel. His learning rule generated a great
deal of interest, but it used a nonlocal calculation to ensure that weights
were normalized. This made it less biologically plausible.
Grossberg extended von der Malsburg's work by rediscovering the instar
rule, examined in Chapter 15. (The instar rule had previously been intro-
duced by Nils Nilsson in his 1965 book Learning Machines.) Grossberg
showed that the instar rule removed the necessity of re-normalizing
weights, since weight vectors that learn to recognize normalized input vec-
tors will automatically be normalized themselves.
The work of Grossberg and von der Malsburg emphasizes the biological
plausibility of their networks. Another influential researcher, Teuvo Ko-
honen, has also been a strong proponent of competitive networks. However,
his emphasis has been on engineering applications and efficient mathe-
matical descriptions of the networks. During the 1970s he developed a sim-
plified version of the instar rule and also, inspired by the work of von der
Malsburg and Grossberg, found an efficient way to incorporate topology
into a competitive network.
In this chapter we will concentrate on the Kohonen framework for compet-
itive networks. His models illustrate the major features of competitive net-
16-2
Hamming Network
works, and yet they are mathematically more tractable than the Grossberg
networks. They provide a good introduction to competitive learning.
We will begin with the simple competitive network. Next we will present
the self-organizing feature map, which incorporates a network topology. Fi-
nally, we will discuss learning vector quantization, which incorporates
competition within a supervised learning framework.
Hamming Network
Since the competitive networks discussed in this chapter are closely related
to the Hamming network (shown in Figure 16.1), it is worth reviewing the
key concepts of that network first.
Feedforward Layer Recurrent Layer
p
Rx1
W1 n2(t + 1) a2(t + 1) a2(t)
n1 a1
SxR
Sx1 Sx1
W2 Sx1 Sx1 D Sx1
SxS
1 b1
Sx1
R S S
- Exp 1(W
a1 = purelin - 1 p + b 1) a2(0) = a1 a2(t + 1) = poslin (W2a2(t))
Figure 16.1 Hamming Network

The Hamming network consists of two layers. The first layer (which is a
layer of instars) performs a correlation between the input vector and the
prototype vectors. The second layer performs a competition to determine
which of the prototype vectors is closest to the input vector.
Layer 1
Recall from Chapter 15 (see page 15-9 and following) that a single instar is
able to recognize only one pattern. In order to allow multiple patterns to be
classified, we need to have multiple instars. This is accomplished in the
Hamming network.
Suppose that we want the network to recognize the following prototype vec-
tors:
p 1 p 2 p Q . (16.1)
16-3
Then the weight matrix, W 1 , and the bias vector, b 1 , for Layer 1 will be:
1w
T p 1T
R
2w
T p 2T
W1 = = , b1 = R , (16.2)

Sw p QT
T
R
where each row of W 1 represents a prototype vector which we want to rec-

ognize, and each element of b 1 is set equal to the number of elements in
each input vector ( R ). (The number of neurons, S , is equal to the number
of prototype vectors which are to be recognized, Q .)
Thus, the output of the first layer is
p 1T p + R
1 p 2T p + R
a = W1p + b1 = . (16.3)

p QT p + R
Note that the outputs of Layer 1 are equal to the inner products of the pro-
totype vectors with the input, plus R . As we discussed in Chapter 3 (page
3-9), these inner products indicate how close each of the prototype patterns
is to the input vector. (This was also discussed in our presentation of the
instar on page 15-10.)
Layer 2
In the instar of Chapter 15, a hardlim transfer function was used to decide
if the input vector was close enough to the prototype vector. In Layer 2 of
the Hamming network we have multiple instars, therefore we want to de-
cide which prototype vector is closest to the input. Instead of the hardlim
transfer function, we will use a competitive layer to choose the closest pro-
totype.
Layer 2 is a competitive layer. The neurons in this layer are initialized with
the outputs of the feedforward layer, which indicate the correlation be-
tween the prototype patterns and the input vector. Then the neurons com-
pete with each other to determine a winner. After the competition, only one
neuron will have a nonzero output. The winning neuron indicates which
category of input was presented to the network (each prototype vector rep-
resents a category).
1
The first-layer output a is used to initialize the second layer.
2 1
a 0 = a (16.4)
16-4
Competitive Layer
Then the second-layer output is updated according to the following recur-

rence relation:
2 2
a t + 1 = poslin W 2 a t . (16.5)
The second-layer weights W 2 are set so that the diagonal elements are 1,
and the off-diagonal elements have a small negative value.

w ij = 1 if i = j
2 1
, where 0 ------------ (16.6)
otherwise S 1
Lateral Inhibition This matrix produces lateral inhibition, in which the output of each neuron
has an inhibitory effect on all of the other neurons. To illustrate this effect,
substitute weight values of 1 and for the appropriate elements of W 2 ,
and rewrite Eq. (16.5) for a single neuron.
a i t + 1 = poslin a i t a j t
2 2 2
(16.7)

ji
At each iteration, each neurons output will decrease in proportion to the

sum of the other neurons outputs (with a minimum output of 0). The out-
put of the neuron with the largest initial condition will decrease more slow-
ly than the outputs of the other neurons. Eventually that neuron will be the
only one with a positive output. At this point the network has reached
steady state. The index of the second-layer neuron with a stable positive
output is the index of the prototype vector that best matched the input.
Winner-Take-All This is called a winner-take-all competition, since only one neuron will have
a nonzero output. In Chapter 18 we will discuss other types of competition.
You may wish to experiment with the Hamming network and the apple/or-
ange classification problem. The Neural Network Design Demonstration
Hamming Classification (nnd3hamc) was previously introduced in Chapter 3.
Competitive Layer
Competition The second-layer neurons in the Hamming network are said to be in com-
petition because each neuron excites itself and inhibits all the other neu-
rons. To simplify our discussions in the remainder of this chapter, we will
define a transfer function that does the job of a recurrent competitive layer:
a = compet n . (16.8)
It works by finding the index i of the neuron with the largest net input,
and setting its output to 1 (with ties going to the neuron with the lowest
index). All other outputs are set to 0.
16-5
1 i = i
ai = , where n i n i i , and i i n i = n i (16.9)
0 i i
Replacing the recurrent layer of the Hamming network with a competitive

transfer function on the first layer will simplify our presentations in this
chapter. (We will study the competition process in more detail in Chapter
18.) A competitive layer is displayed in Figure 16.2.
Input Competitive Layer
p n a
Rx1
W Sx1 C Sx1
SxR
R S
a = compet (Wp)
Figure 16.2 Competitive Layer

As with the Hamming network, the prototype vectors are stored in the rows
of W . The net input n calculates the distance between the input vector p
and each prototype iw (assuming vectors have normalized lengths of L ).
The net input n i of each neuron i is proportional to the angle i between
p and the prototype vector iw :
2
1w
T
1w
Tp L cos 1
2w 2w
Tp 2
= L cos 2 .
T
n = Wp = p = (16.10)

Sw Sw
T Tp 2
L cos S
The competitive transfer function assigns an output of 1 to the neuron

whose weight vector points in the direction closest to the input vector:
a = compet Wp . (16.11)
To experiment with the competitive network and the apple/orange classifi-

cation problem, use the Neural Network Design Demonstration Competitive
Classification (nnd14cc).
16-6
Competitive Layer
Competitive Learning
We can now design a competitive network classifier by setting the rows of
W to the desired prototype vectors. However, we would like to have a
learning rule that could be used to train the weights in a competitive net-
work, without knowing the prototype vectors. One such learning rule is the
instar rule from Chapter 15:
iw q = iw q 1 + a i q p q iw q 1 . (16.12)
For the competitive network, a is only nonzero for the winning neuron
( i = i ). Therefore, we can get the same results using the Kohonen rule.
(16.13)
= 1 iw q 1 + p q
and
iw q = iw q 1 i i (16.14)
Thus, the row of the weight matrix that is closest to the input vector (or has
the largest inner product with the input vector) moves toward the input
vector. It moves along a line between the old row of the weight matrix and
the input vector, as shown in Figure 16.3.
p(q)
w(q)
i
w(q - 1)
i
Figure 16.3 Graphical Representation of the Kohonen Rule

2 Lets use the six vectors in Figure 16.4 to demonstrate how a competitive
+2
layer learns to classify vectors. Here are the six vectors:
p 1 = 0.1961 p 2 = 0.1961 p 3 = 0.9806 (16.15)

0.9806 0.9806 0.1961
p4 = 0.9806 p = 0.5812 p = 0.8137 .

5 6
0.1961 0.8137 0.5812
16-7
Figure 16.4 Sample Input Vectors

Our competitive network will have three neurons, and therefore it can clas-
sify vectors into three classes. Here are the randomly chosen normalized
initial weights:
T
1w
0.7071 w = 0.7071 w = 1.0000 , W =
1w = . (16.16)
T
2w
2 3
p1 p2 0.7071 0.7071 0.0000
2w 3w
T
p3 The data vectors are shown at left, with the weight vectors displayed as ar-
3w rows. Lets present the vector p 2 to the network:
p4
p6 0.7071 0.7071
1w
p5 0.1961
a = compet Wp 2 = compet 0.7071 0.7071 (16.17)
0.9806
1.0000 0.0000
0.5547 0

= compet 0.8321 = 1 .

0.1961 0
The second neurons weight vector was closest to p 2 , so it won the compe-
tition ( i = 2 ) and output a 1. We now apply the Kohonen learning rule to
the winning neuron with a learning rate of = 0.5.
new old old

2w = 2w + p 2 2w (16.18)
p1 p2
2w
= 0.7071 + 0.5 0.1961 0.7071 = 0.4516
0.7071 0.9806 0.7071 0.8438
p3
3w
p4 The Kohonen rule moves 2w closer to p 2 , as can be seen in the diagram at
left. If we continue choosing input vectors at random and presenting them
p6 to the network, then at each iteration the weight vector closest to the input
1w
p5 vector will move toward that vector. Eventually, each weight vector will
16-8
Competitive Layer
point at a different cluster of input vectors. Each weight vector becomes a

prototype for a different cluster.
This problem is simple enough that we can predict which weight vector will
point at which cluster. The final weights will look something like those
p1 p2
2w
p3
1w
p4
3w
p6
p5
2w Figure 16.5 Final Weights
1w Once the network has learned to cluster the input vectors, it will classify
new vectors accordingly.The diagram in the left margin uses shading to
3w
show which region each neuron will respond to. The competitive layer as-
signs each input vector p to one of these classes by producing an output of
1 for the neuron whose weight vector is closest to p .
To experiment with the competitive learning use the Neural Network Design
Demonstration Competitive Learning (nnd14cl).
Problems with Competitive Layers

stable Competitive layers make efficient adaptive classifiers, but they do suffer
slow from a few problems. The first problem is that the choice of learning rate
forces a trade-off between the speed of learning and the stability of the final
weight vectors. A learning rate near zero results in slow learning. However,
once a weight vector reaches the center of a cluster it will tend to stay close
0
to the center.
unstable In contrast, a learning rate near 1.0 results in fast learning. However, once
fast
the weight vector has reached a cluster, it will continue to oscillate as dif-
ferent vectors in the cluster are presented.
1 Sometimes this trade-off between fast learning and stability can be used to
advantage. Initial training can be done with a large learning rate for fast
learning. Then the learning rate can be decreased as training progresses,
to achieve stable prototype vectors. Unfortunately, this technique will not
work if the network needs to continuously adapt to new arrangements of
input vectors.
A more serious stability problem occurs when clusters are close together.
In certain cases, a weight vector forming a prototype of one cluster may in-
16-9
vade the territory of another weight vector, and therefore upset the cur-
rent classification scheme.
The series of four diagrams in Figure 16.6 illustrate this problem. Two in-
put vectors (shown with blue circles in diagram (a)) are presented several
times. The result is that the weight vectors representing the middle and
right clusters shift to the right. Eventually one of the right cluster vectors
is reclassified by the center weight vector. Further presentations move the
middle vector over to the right until it loses some of its vectors, which
then become part of the class associated with the left weight vector.
(a) (b) (c) (d)
Figure 16.6 Example of Unstable Learning

A third problem with competitive learning is that occasionally a neurons
initial weight vector is located so far from any input vectors that it never
wins the competition, and therefore never learns. The result is a dead
neuron, which does nothing useful. For example, the downward-pointing
weight vector in the diagram to the left will never learn, regardless of the
order in which vectors are presented. One solution to this problem consists
of adding a negative bias to the net input of each neuron and then decreas-
ing the bias each time the neuron wins. This will make it harder for a neu-
ron to win the competition if it has won often. This mechanism is
sometimes called a conscience. (See Exercise E16.4.)
Finally, a competitive layer always has as many classes as it has neurons.
This may not be acceptable for some applications, especially when the num-
ber of clusters is not known in advance. In addition, for competitive layers,
each class consists of a convex region of the input space. Competitive layers
cannot form classes with nonconvex regions or classes that are the union of
unconnected regions.
Some of the problems discussed in this section are solved by the feature
map and LVQ networks, which are introduced in later sections of this chap-
ter, and the ART networks, which are presented in Chapter 19.
Competitive Layers in Biology

In previous chapters we have made no mention of how neurons are physi-
cally organized within a layer (the topology of the network). In biological
neural networks, neurons are typically arranged in two-dimensional lay-
ers, in which they are densely interconnected through lateral feedback. The
diagram to the left shows a layer of twenty-five neurons arranged in a two-
dimensional grid.
16-10
Competitive Layers in Biology
Often weights vary as a function of the distance between the neurons they
connect. For example, the weights for Layer 2 of the Hamming network are
assigned as follows:
1 if i = j
w ij = . (16.19)
if i j
Eq. (16.20) assigns the same values as Eq. (16.19), but in terms of the dis-
tances d ij between neurons:
neuron j
1 if d ij = 0
-e -e -e -e -e w ij = . (16.20)
if d ij 0
-e -e -e -e -e
-e -e +1 -e -e
Either Eq. (16.19) or Eq. (16.20) will assign the weight values shown in the
-e -e -e -e -e diagram at left. Each neuron i is labeled with the value of the weight w ij ,
-e -e -e -e -e which comes from it to the neuron marked j .
On-center/off-surround The term on-center/off-surround is often used to describe such a connection
pattern between neurons. Each neuron reinforces itself (center), while in-
hibiting all other neurons (surround).
It turns out that this is a crude approximation of biological competitive lay-
ers. In biology, a neuron reinforces not only itself, but also those neurons
close to it. Typically, the transition from reinforcement to inhibition occurs
smoothly as the distance between neurons increases.
This transition is illustrated on the left side of Figure 16.7. This is a func-
tion that relates the distance between neurons to the weight connecting
them. Those neurons that are close provide excitatory (reinforcing) connec-
tions, and the magnitude of the excitation decreases as the distance in-
creases. Beyond a certain distance, the neurons begin to have inhibitory
connections, and the inhibition increases as the distance increases. Be-
Mexican-Hat Function cause of its shape, the function is referred to as the Mexican-hat function.
On the right side of Figure 16.7 is a two-dimensional illustration of the
Mexican-hat (on-center/off-surround) function. Each neuron i is marked to
show the sign and relative strength of its weight w ij going to neuron j .
wij neuron j
- - - - -
+ - + + + -
dij - + + + -
- + + + -
-
- - - - -
Figure 16.7 On-Center/Off-Surround Layer in Biology
16-11
Biological competitive systems, in addition to having a gradual transition

between excitatory and inhibitory regions of the on-center/off-surround
connection pattern, also have a weaker form of competition than the win-
ner-take-all competition of the Hamming network. Instead of a single ac-
tive neuron (winner), biological networks generally have bubbles of
activity that are centered around the most active neuron. This is caused in
part by the form of the on-center/off-surround connectivity pattern and also
by nonlinear feedback connections. (See the discussion on contour enhance-
ment in Chapter 18.)
Self-Organizing Feature Maps

In order to emulate the activity bubbles of biological systems, without hav-
ing to implement the nonlinear on-center/off-surround feedback connec-
tions, Kohonen designed the following simplification. His self-organizing
SOFM feature map (SOFM) network first determines the winning neuron i us-
ing the same procedure as the competitive layer. Next, the weight vectors
for all neurons within a certain neighborhood of the winning neuron are up-
dated using the Kohonen rule,
i N i d , (16.21)
= 1 iw q 1 + p q
Neighborhood where the neighborhood N i d contains the indices for all of the neurons
that lie within a radius d of the winning neuron i :
N i d = j d ij d . (16.22)
When a vector p is presented, the weights of the winning neuron and its
neighbors will move toward p . The result is that, after many presenta-
tions, neighboring neurons will have learned vectors similar to each other.
2 To demonstrate the concept of a neighborhood, consider the two diagrams
+2
shown in Figure 16.8. The left diagram illustrates a two-dimensional
neighborhood of radius d = 1 around neuron 13 . The right diagram shows
a neighborhood of radius d = 2 .
The definition of these neighborhoods would be
N 13 1 = 8 12 13 14 18 , (16.23)
N 13 2 = 3 7 8 9 11 12 13 14 15 17 18 19 23 . (16.24)
16-12
1 2 3 4 5 1 2 3 4 5
6 7 8 9 10 6 7 8 9 10
11 12 13 14 15 11 12 13 14 15
16 17 18 19 20 16 17 18 19 20
21 22 23 24 25 21 22 23 24 25
N13(1) N13(2)
Figure 16.8 Neighborhoods

We should mention that the neurons in an SOFM do not have to be ar-
ranged in a two-dimensional pattern. It is possible to use a one-dimension-
al arrangement, or even three or more dimensions. For a one-dimensional
SOFM, a neuron will only have two neighbors within a radius of 1 (or a sin-
gle neighbor if the neuron is at the end of the line). It is also possible to de-
fine distance in different ways. For instance, Kohonen has suggested
rectangular and hexagonal neighborhoods for efficient implementation.
The performance of the network is not sensitive to the exact shape of the
neighborhoods.
2 Now lets demonstrate the performance of an SOFM network. Figure 16.9
+2
shows a feature map and the two-dimensional topology of its neurons.
Input Feature Map Feature Map
1 2 3 4 5
6 7 8 9 10
p n a
3x1
W 25 x 1 C 25 x 1
11 12 13 14 15
25 x 3 16 17 18 19 20
21 22 23 24 25
3 25
1 a = compet (Wp)
0.8
0.6
Figure 16.9 Self-Organizing Feature Map
0.4
0.2 The diagram in the left margin shows the initial weight vectors for the fea-
0
ture map. Each three-element weight vector is represented by a dot on the
-0.2
sphere. (The weights are normalized, therefore they will fall on the surface
-0.4
of a sphere.) Dots of neighboring neurons are connected by lines so you can
-0.6
see how the physical topology of the network is arranged in the input space.
-0.8
-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
16-13
1
The diagram to the left shows a square region on the surface of the sphere.
0.8
We will randomly pick vectors in this region and present them to the fea-
0.6
ture map.
0.4
0.2
Each time a vector is presented, the neuron with the closest weight vector
0
will win the competition. The winning neuron and its neighbors move their
-0.2
-0.4
weight vectors closer to the input vector (and therefore to each other). For
-0.6
this example we are using a neighborhood with a radius of 1.
-0.8
-1
The weight vectors have two tendencies: first, they spread out over the in-
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
put space as more vectors are presented; second, they move toward the
weight vectors of neighboring neurons. These two tendencies work together
to rearrange the neurons in the layer so that they evenly classify the input
space.
The series of diagrams in Figure 16.10 shows how the weights of the twen-
ty-five neurons spread out over the active input space and organize them-
selves to match its topology.
1 1 1 1
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
-0.2 -0.2 -0.2 -0.2
-0.4 -0.4 -0.4 -0.4
-0.6 -0.6 -0.6 -0.6
-0.8 -0.8 -0.8 -0.8
-1 -1 -1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
1 1 1 1
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
-0.2 -0.2 -0.2 -0.2
-0.4 -0.4 -0.4 -0.4
-0.6 -0.6 -0.6 -0.6
-0.8 -0.8 -0.8 -0.8
-1 -1 -1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Figure 16.10 Self-Organization, 250 Iterations per Diagram

In this example, the input vectors were generated with equal probability
from any point in the input space. Therefore, the neurons classify roughly
equal areas of the input space.
Figure 16.11 provides more examples of input regions and the resulting
feature maps after self-organization.
16-14
1 1 1 1
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
-0.2 -0.2 -0.2 -0.2
-0.4 -0.4 -0.4 -0.4
-0.6 -0.6 -0.6 -0.6
-0.8 -0.8 -0.8 -0.8
-1 -1 -1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Figure 16.11 Other Examples of Feature Map Training

Occasionally feature maps can fail to properly fit the topology of their input
space. This usually occurs when two parts of the net fit the topology of sep-
arate parts of the input space, but the net forms a twist between them. An
example is given in Figure 16.12.
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Figure 16.12 Feature Map with a Twist

It is unlikely that this twist will ever be removed, because the two ends of
the net have formed stable classifications of different regions.
Improving Feature Maps

So far, we have described only the most basic algorithm for training feature
maps. Now lets consider several techniques that can be used to speed up
the self-organizing process and to make it more reliable.
One method to improve the performance of the feature map is to vary the
size of the neighborhoods during training. Initially, the neighborhood size,
d , is set large. As training progresses, d is gradually reduced, until it only
includes the winning neuron. This speeds up self-organizing and makes
twists in the map very unlikely.
The learning rate can also be varied over time. An initial rate of 1 allows
neurons to quickly learn presented vectors. During training, the learning
rate is decreased asymptotically toward 0, so that learning becomes stable.
16-15
(We discussed the use of this technique for competitive layers earlier in the
chapter.)
Another alteration that speeds self-organization is to have the winning
neuron use a larger learning rate than the neighboring neurons.
Finally, both competitive layers and feature maps often use an alternative
expression for net input. Instead of using the inner product, they can di-
rectly compute the distance between the input vector and the prototype
vectors. The advantage of using the distance is that input vectors do not
need to be normalized. This alternative net input expression is introduced
in the next section on LVQ networks.
Other enhancements to the SOFM are described in Chapter 26, including
a batch version of the SOFM learning rule. That chapter is a case study of
using the SOFM for clustering.
To experiment with feature maps use the Neural Network Design Demon-
strations 1-D Feature Maps (nnd14fm1) and 2-D Feature Maps (nnd14fm2).
Learning Vector Quantization

The final network we will introduce in this chapter is the learning vector
quantization (LVQ) network, which is shown in Figure 16.13. The LVQ net-
work is a hybrid network. It uses both unsupervised and supervised learn-
ing to form classifications.
In the LVQ network, each neuron in the first layer is assigned to a class,
with several neurons often assigned to the same class. Each class is then
assigned to one neuron in the second layer. The number of neurons in the
1
first layer, S , will therefore always be at least as large as the number of
2
neurons in the second layer, S , and will usually be larger.
Inputs Competitive Layer Linear Layer
W1 2
1
S xR a
2
S x1
p n1 a1 n2
Rx1
||dist||
S1x1 C 1
S x1
W 2
2
S x1
S2xS1
1 2
R S S
1 1
n i = ||iw -p|| a2 = W2a1
1 1
a = compet(n )
Figure 16.13 LVQ Network
16-16
As with the competitive network, each neuron in the first layer of the LVQ
network learns a prototype vector, which allows it to classify a region of the
input space. However, instead of computing the proximity of the input and
weight vectors by using the inner product, we will simulate the LVQ net-
works by calculating the distance directly. One advantage of calculating
the distance directly is that vectors need not be normalized. When the vec-
tors are normalized, the response of the network will be the same, whether
the inner product is used or the distance is directly calculated.
The net input of the first layer of the LVQ will be
1 1
n i = iw p , (16.25)
or, in vector form,
1
1w p
1
1
n = 2w p
16

1
w p
1
S
. (16.26)
The output of the first layer of the LVQ is
1 1
a = compet n . (16.27)
Therefore the neuron whose weight vector is closest to the input vector will
output a 1, and the other neurons will output 0.
Thus far, the LVQ network behaves exactly like the competitive network
(at least for normalized vectors). There is a difference in interpretation,
however. In the competitive network, the neuron with the nonzero output
indicates which class the input vector belongs to. For the LVQ network, the
Subclass winning neuron indicates a subclass, rather than a class. There may be sev-
eral different neurons (subclasses) that make up each class.
The second layer of the LVQ network is used to combine subclasses into a
2 2
single class. This is done with the W matrix. The columns of W repre-
2
sent subclasses, and the rows represent classes. W has a single 1 in each
column, with the other elements set to zero. The row in which the 1 occurs
indicates which class the appropriate subclass belongs to.
2
w ki = 1 subclass i is a part of class k (16.28)
The process of combining subclasses to form a class allows the LVQ net-
work to create complex class boundaries. A standard competitive layer has
16-17
the limitation that it can only create decision regions that are convex. The
LVQ network overcomes this limitation.
LVQ Learning
The learning in the LVQ network combines competitive learning with su-
pervision. As with all supervised learning algorithms, it requires a set of
examples of proper network behavior:
p 1 t 1 p 2 t 2 p Q t Q .
Each target vector must contain only zeros, except for a single 1. The row
in which the 1 appears indicates the class to which the input vector be-
longs. For example, if we have a problem where we would like to classify a
particular three-element vector into the second of four classes, we can ex-
press this as

0
12
1 .
p1 = 0 t1 = (16.29)
0
12 0

Before learning can occur, each neuron in the first layer is assigned to an
2
output neuron. This generates the matrix W . Typically, equal numbers of
hidden neurons are connected to each output neuron, so that each class can
2
be made up of the same number of convex regions. All elements of W are
set to zero, except for the following:
2
If hidden neuron i is to be assigned to class k, then set w ki = 1 . (16.30)
2 1
Once W is defined, it will never be altered. The hidden weights W are
trained with a variation of the Kohonen rule.
The LVQ learning rule proceeds as follows. At each iteration, an input vec-
tor p is presented to the network, and the distance from p to each proto-
type vector is computed. The hidden neurons compete, neuron i wins the
competition, and the i th element of a is set to 1. Next, a is multiplied
1 1
2 2
by W to get the final output a , which also has only one nonzero element,
k , indicating that p is being assigned to class k .
The Kohonen rule is used to improve the hidden layer of the LVQ network
in two ways. First, if p is classified correctly, then we move the weights
1
i
w of the winning hidden neuron toward p .
1 1 1 2
i
w q = i
w q 1 + p q iw q 1 , if a k = t k = 1 (16.31)
16-18
Second, if p was classified incorrectly, then we know that the wrong hidden
1
neuron won the competition, and therefore we move its weights iw away
from p .
1 1 1 2
i
w q = i
w q 1 p q iw q 1 , if a k = 1 t k = 0 (16.32)
The result will be that each hidden neuron moves toward vectors that fall
into the class for which it forms a subclass and away from vectors that fall
into other classes.
2 Lets take a look at an example of LVQ training. We would like to train an
+2
LVQ network to solve the following classification problem:

class 1: p 1 = 1 p 2 = 1 , class 2: p 3 = 1 p 4 = 1 , (16.33)
Class 1 1 1 1 1
p4 p2 as illustrated by the figure in the left margin. We begin by assigning target
vectors to each input:
16
1 t = 1 , p = 1 t = 1 ,
p1 = 1 2 2 (16.34)
p1 p3 1 0 1 0
Class 2
1 t = 0 , p = 1 t = 0 .
p3 = 3 4 4 (16.35)
1 1 1 1
We now must choose how many subclasses will make up each of the two
classes. If we let each class be the union of two subclasses, we will end up
with four neurons in the hidden layer. The output layer weight matrix will
be
W = 1 1 0 0 .
2
(16.36)
0 0 1 1
2
W connects hidden neurons 1 and 2 to output neuron 1. It connects hidden
neurons 3 and 4 to output neuron 2. Each class will be made up of two con-
vex regions.
1
The row vectors in W are initially set to random values. They can be seen
4w
1
1w
1
p4 p2 in the diagram at left. The weights belonging to the two hidden neurons
that define class 1 are marked with hollow circles. The weights defining
3w
1
class 2 are marked with solid circles. The values for these weights are
2w
1
= 0.543 , 2w = 0.969 , 3w = 0.997 , 4w = 0.456 .

1 1 1 1
1w
p1 p3 (16.37)
0.840 0.249 0.094 0.954
16-19
At each iteration of the training process, we present an input vector, find

its response, and then adjust the weights. In this case we will begin by pre-
senting p 3 .
1
1w p 3

1 1 2w p 3
1
a = compet n = compet (16.38)
3w p 3
1

1
4w p 3
T T
0.543 0.840 1 1
2.40
T T
0
0.969 0.249 1 1
= compet = 0
= compet 2.11
T T
0.997 0.094 1 1 1.09 1
2.03 0
0.456 0.954 T 1 1 T

The third hidden neuron has the closest weight vector to p 3 . In order to de-
1 2
termine which class this neuron belongs to, we multiply a by W .
0
a = W a = 1 1 0 0 0 = 0
2 2 1
(16.39)
0 0 1 1 1 1
0
1
This output indicates that p 3 is a member of class 2. This is correct, so 3w
is updated by moving it toward p 3 .
1 1 1
3w 1 = 3w 0 + p 3 3w 0 (16.40)

= 0.997 + 0.5 1 0.997 = 0.998
0.094 1 0.094 0.453
1
The diagram on the left side of Figure 16.14 shows the weights after 3w
was updated on the first iteration. The diagram on the right side of Figure
16.14 shows the weights after the algorithm has converged.
16-20
The diagram on the right side of Figure 16.14 also indicates how the re-
gions of the input space will be classified. The regions that will be classified
as class 1 are shown in gray, and the regions that will be classified as class
2 are shown in blue.
1w
1
4w
1
4w
1
1w
1 p4 p2
p4 p2
2w
1
3w
1
p1 p3 p1 p3
2w 3w
1 1
Figure 16.14 After First and Many Iterations
Improving LVQ Networks (LVQ2)

The LVQ network described above works well for many problems, but it
does suffer from a couple of limitations. First, as with competitive layers,
occasionally a hidden neuron in an LVQ network can have initial weight
values that stop it from ever winning the competition. The result is a dead
neuron that never does anything useful. This problem is solved with the
use of a conscience mechanism, a technique discussed earlier for compet-
itive layers, and also presented in Exercise E16.4.
Secondly, depending on how the initial weight vectors are arranged, a neu-
rons weight vector may have to travel through a region of a class that it
doesnt represent, to get to a region that it does represent. Because the
weights of such a neuron will be repulsed by vectors in the region it must
cross, it may not be able to cross, and so it may never properly classify the
region it is being attracted to. This is usually solved by applying the follow-
ing modification to the Kohonen rule.
If the winning neuron in the hidden layer incorrectly classifies the current
input, we move its weight vector away from the input vector, as before.
However, we also adjust the weights of the closest neuron to the input vec-
tor that does classify it properly. The weights for this second neuron should
be moved toward the input vector.
When the network correctly classifies an input vector, the weights of only
one neuron are moved toward the input vector. However, if the input vector
is incorrectly classified, the weights of two neurons are updated, one weight
vector is moved away from the input vector, and the other one is moved to-
LVQ2 ward the input vector. The resulting algorithm is called LVQ2.
To experiment with LVQ networks use the Neural Network Design Demon-
strations LVQ1 Networks (nnd14lv1) and LVQ2 Networks (nnd14lv2).
16-21
Summary of Results
Competitive Layer
p n a
Rx1
W Sx1 C Sx1
SxR
R S
a = compet (Wp)
Competitive Learning with the Kohonen Rule

i
wq = i
w q 1 + p q iw q 1 = 1 iw q 1 + p q
i
wq = i
wq 1 i i ,
where i is the winning neuron.
p(q)
w(q)
i
w(q - 1)
i
16-22
Summary of Results
Self-Organizing Feature Map

Input Feature Map Feature Map
1 2 3 4 5
6 7 8 9 10
p n a
3x1
W 25 x 1 C 25 x 1
11 12 13 14 15
25 x 3 16 17 18 19 20
21 22 23 24 25
3 25
a = compet (Wp)
Self-Organizing with the Kohonen Rule

= 1 iw q 1 + p q
i N i d 16
N i d = j d ij d
LVQ Network
1
W 2
1
S xR a
2
S x1
p n
1
a1 n
2
Rx1
||dist||
S1x1 C 1
S x1
W 2
S2x1
S2xS1
R S1 S2
1 1
n i = ||iw -p|| a2 = W2a1
1 1
a = compet(n )
2
w ki = 1 subclass i is a part of class k
LVQ Network Learning with the Kohonen Rule

1 1 1 2
i
w q = i
w q 1 + p q iw q 1 , if a k = t k = 1
1 1 1 2
i
w q = i
w q 1 p q iw q 1 , if a k = 1 t k = 0
16-23
Solved Problems
P16.1 Figure P16.1 shows several clusters of normalized vectors.
class 1 class 2
class 4
class 3
Figure P16.1 Clusters of Input Vectors for Problem P16.1

Design the weights of the competitive network shown in Figure
P16.2, so that it classifies the vectors according to the classes indi-
cated in the diagram and with the minimum number of neurons.
p n a
Rx1
W Sx1 C Sx1
SxR
R S
a = compet (Wp)
Figure P16.2 Competitive Network for Problem P16.1

Redraw the diagram showing the weights you chose and the deci-
sion boundaries that separate the region of each class.
Since there are four classes to be defined, the competitive layer will need
four neurons. The weights of each neuron act as prototypes for the class
that neuron represents. Therefore, for each neuron we will choose a proto-
type vector that appears to be approximately at the center of a cluster.
Classes 1, 2 and 3 each appear to be roughly centered at a multiple of 45 .
Given this, the following three vectors are normalized (as is required for
the competitive layer) and point in the proper directions.
1w = 1 2 , 2w = 1 2 , 3w = 1 2
1 2 1 2 1 2
16-24
Solved Problems
The center of the fourth cluster appears to be about twice as far from the
vertical axis as it is from the horizontal axis. The resulting normalized
weight vector is
4w = 2 5 .
1 5
The weight matrix W for the competitive layer is simply the matrix of the
transposed prototype vectors:
T
1w 1 2 1 2
T
2w 1 2 1 2 .
W = =
3w
T
1 2 1 2
T 2 5 1 5
4w
We get Figure P16.3 by drawing these weight vectors with arrows and bi-
secting the circle between each adjacent weight vector to get the class re-
gions.
class 1 class 2
class 4
class 3
Figure P16.3 Final Classifications for Problem P16.1
P16.2 Figure P16.4 shows three input vectors and three initial weight
vectors for a three-neuron competitive layer. Here are the values
of the input vectors:
p1 = 1 , p2 = 0 , p3 = 1 2 .
0 1 1 2
The initial values of the three weight vectors are
0 , w = 2 5 , w = 1 5 .
1w = 2 3
1 1 5 2 5
16-25
Calculate the resulting weights found after training the competi-

tive layer with the Kohonen rule and a learning rate of 0.5, on
the following series of inputs:
p1 , p2 , p3 , p1 , p2 , p3 .
p2
3w
p3
2w
p1
1w
Figure P16.4 Input Vectors and Initial Weights for Problem P16.2
First we combine the weight vectors into the weight matrix W .
0 1
W = 2 5 1 5
1 5 2 5
Then we present the first vector p 1 .
0 1 0 0
1
a = compet Wp 1 = compet 2 5 1 5 = compet 0.894 = 1
0
1 5 2 5 0.447 0
The second neuron responded, since 2w was closest to p 1 . Therefore, we

will update 2w with the Kohonen rule.

= 2 5 + 0.5 1 2 5 = 0.947
new old old
2w = 2w + p 1 2w
p2
3w 0
p3 1 5 1 5 0.224

2w
p1
The diagram at left shows that the new 2w moved closer to p 1 .
We will now repeat this process for p 2 .
1w
16-26
Solved Problems
0 1 1 0
0
a = compet Wp 2 = compet 0.947 0.224 = compet 0.224 = 0
1
1 5 2 5 0.894 1
The third neuron won, so its weights move closer to p 2 .

= 1 5 + 0.5 0 1 5 = 0.224
new old old
3w = 3w + p 2 3w
1
2 5 2 5 0.947
We now present p 3 .
0 1
1 2
a = compet Wp 3 = compet 0.947 0.224
1 2
0.224 0.947
0.707 0

= compet 0.512 = 0

0.512 1
The third neuron wins again.

= 0.224 + 0.5 1 2 0.224 = 0.2417
new old old
3w = 3w + p 2 3w
0.947 1 2 0.947 0.8272
After presenting p 1 through p 3 again, neuron 2 will again win once and
neuron 3 twice. The final weights are
p2
3w p3 0 1
W = 0.974 0.118 .
2w 0.414 0.8103
p1
The final weights are also shown in the diagram at left.
Note that 2w has almost learned p 1 , and 3w is directly between p 2 and p 3 .
1w The other weight vector, 1w , was never updated. The first neuron, which
never won the competition, is a dead neuron.
P16.3 Consider the configuration of input vectors and initial weights

shown in Figure P16.5. Train a competitive network to cluster
16-27
these vectors using the Kohonen rule with learning rate = 0.5 .
Find graphically the position of the weights after all of the input
vectors (in the order shown) have been presented once.
p4
2 w(0) w(0)
1
p3 p1
p2
Figure P16.5 Input Vectors and Initial Weights for Problem P16.3
This problem can be solved graphically, without any computations. The re-
sults are displayed in Figure P16.6.
p4
w(0)
2 w(0)
1
2w(4)
1 w(1)
w(3)
2
p3 p1
w(2)
1
p2
Figure P16.6 Solution for Problem P16.3

The input vector p 1 is presented first. The weight vector 1w is closest to
p 1 , therefore neuron 1 wins the competition and 1w is moved halfway to
16-28
Solved Problems
p 1 , since = 0.5 . Next, p 2 is presented, and again neuron 1 wins the com-
petition and 1w is moved halfway to p 2 . During these first two iterations,
2w is not changed.
On the third iteration, p 3 is presented. This time 2w wins the competition

and is moved halfway to p 3 . On the fourth iteration, p 4 is presented, and
neuron 2 again wins. The weight vector 2w is moved halfway to p 4 .
If we continue to train the network, neuron 1 will classify the input vectors
p 1 and p 2 , and neuron 2 will classify the input vectors p 3 and p 4 . If the
input vectors were presented in a different order, would the final classifi-
cation be different?
P16.4 So far in this chapter we have only talked about feature maps
whose neurons are arranged in two dimensions. The feature map
shown in Figure P16.7 contains nine neurons arranged in one di-
mension.
Inputs Feature Map Feature Map

1
p n a 4
3x1
W 9x1 C 9x1 5
9x3 6
7
3 9 8
a = compet(Wp)
Figure P16.7 Nine-Neuron Feature Map

Given the following initial weights, draw a diagram of the weight
vectors, with lines connecting weight vectors of neighboring neu-
rons.
T
0.41 0.45 0.41 0 0 0 0.41 0.45 0.41
W = 0.41 0 0.41 0.45 0 0.45 0.41 0 0.41
0.82 0.89 0.82 0.89 1 0.89 0.82 0.89 0.82
Train the feature map for one iteration, on the vector below, using
a learning rate of 0.1 and a neighborhood of radius 1. Redraw the
diagram for the new weight matrix.
16-29
0.67
p = 0.07
0.74
The feature map diagram for the initial weights is given in Figure P16.8.
1
0.8
0.6
7 4
0.4 1
0.2
0 8 5 2
-0.2
-0.4 9 6 3
-0.6
-0.8
-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Figure P16.8 Original Feature Map

We start updating the network by presenting p to the network.
a = compet Wp
0.41 0.45 0.41 0 0 0 0.41 0.45 0.41 T

0.67

= compet 0.41 0 0.41 0.45 0 0.45 0.41 0 0.41 0.07

0.82 0.89 0.82 0.89 1 0.89 0.82 0.89 0.82 0.74
T
= compet 0.91 0.96 0.85 0.70 0.74 0.63 0.36 0.36 0.3
T
= 0 1 0 0 0 0 0 0 0
The second neuron won the competition. Looking at the network diagram,
we see that the second neurons neighbors, at a radius of 1, include neurons
1 and 3. We must update each of these neurons weights with the Kohonen
rule.
16-30
Solved Problems
0.41 0.67 0.41 0.43

1 w 1 = 1w 0 + p 1w 0 = 0.41 + 0.1 0.07 0.41 = 0.37

0.82 0.74 0.82 0.81
0.45 0.67 0.45 0.47

2 w 1 = 2w 0 + p 2w 0 = 0 + 0.1 0.07 0 = 0.01

0.89 0.74 0.89 0.88
0.41 0.67 0.41 0.43

3 w 1 = 3w 0 + p 3w 0 = 0.41 + 0.1 0.07 0.41 = 0.36

0.82 0.74 0.82 0.81
Figure P16.9 shows the feature map after the weights were updated.
1
0.8
0.6
7 4
0.4
1
0.2
0 8 5 2
-0.2
-0.4 9 6 3
-0.6
-0.8
-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Figure P16.9 Feature Map after Update
P16.5 Given the LVQ network shown in Figure P16.10 and the weight val-
ues shown below, draw the regions of the input space that make up
each class.
0 0
1 1 1 0 0 0 0
1 2
W = 1 1 , W = 0 1 0 0 0
1 1 0 0 1 1 1
1 1
16-31
W1 2
5x2 a
3x1
p n1 a1 n
2
2x1
-||dist||
5x1 C 5x1
W
2
3x1
3x5
2 5 3
1 1
n i = -||iw -p|| a2 = W2a1
1 1
a = compet(n )
Figure P16.10 LVQ Network for Problem P16.5

We create the diagram shown in Figure P16.11 by marking each vector iw
1
in W according to the index k of the corresponding nonzero element in the
2
ith column of W , which indicates the class.
Figure P16.11 Prototype Vectors Marked by Class

The decision boundaries separating each class are found by drawing lines
between each pair of prototype vectors, perpendicular to an imaginary line
connecting them and equidistant from each vector.
In Figure P16.12, each convex region is colored according to the weight vec-
tor it is closest to.
Figure P16.12 Class Regions and Decision Boundaries
16-32
Solved Problems
P16.6 Design an LVQ network to solve the classification problem shown

in Figure P16.13. The vectors in the diagram are to be classified
into one of three classes, according to their color.
Figure P16.13 Classification Problem

When the design is complete, draw a diagram showing the region
for each class.
We will begin by noting that since LVQ networks calculate the distance be-
tween vectors directly, instead of using the inner product, they can classify
vectors that are not normalized, such as those above.
Next we will identify each color with a class:
Class 1 will include all white dots.
Class 2 will include all black dots.
Class 3 will include all blue dots.
Now we can choose the dimensions of the LVQ network. Since there are
three classes, the network must have three neurons in its output layer.
There are nine subclasses (i.e., clusters). Therefore the hidden layer must
have nine neurons. This gives us the network shown in Figure P16.14.
W1 2
9x2 a
3x1
p n1 a1 n2
2x1
-||dist||
9x1 C 9x1
W
2
3x1
3x9
2 9 3
1 1
n i = -||iw -p|| a2 = W2a1
1 1
a = compet(n )
Figure P16.14 LVQ Network for Problem P16.6
16-33
1
We can now design the weight matrix W of the first layer by setting each
row equal to a transposed prototype vector for one cluster. Picking proto-
type vectors at the center of each cluster gives us the following values:
T
W = 1 0 1 1 0 1 1 0 1 .
1
1 1 1 0 0 0 1 1 1
Now each neuron in the first layer will respond to a different cluster.
2
Next we choose W so that each subclass is connected to the appropriate
class. To do this we use the following rule:
2
If subclass i is to be assigned to class k, then set w ki = 1 .
For example, the first subclass is the top-left cluster in the vector diagram.
The vectors in this cluster are white, so they belong in the first class. There-
2
fore we should set w 1 1 to one.
Once we have done this for all nine subclasses we end up with these values:
1 0 0 0 0 1 0 1 0
2
W = 0 1 0 0 1 0 1 0 0 .
0 0 1 1 0 0 0 0 1
We can test the network by presenting a vector to it. Here we calculate the
T
output of the first layer for p = 1 0 :

5 0

2 0
1 0

2 0
a = compet n = compet
1 1
1 = 0 .
0 1

5 0

2 0
0
1
The network says that the vector we presented is in the sixth subclass.
Lets see what the second layer says.
16-34
Solved Problems
0
0
0
1 0 0 0 0 1 0 1 0 0 1
2 2 1
a = W a = 0 1 0 0 1 0 1 0 0 0 = 0
0 0 1 1 0 0 0 0 1 1 0
0
0
0
The second layer indicates that the vector is in class 1, as indeed it is. The
diagram of class regions and decision boundaries is shown in Figure
P16.15.
Figure P16.15 Class Regions and Decision Boundaries
P16.7 Competitive layers and feature maps require that input vectors be
normalized. But what if the available data is not normalized?
One way to handle such data is simply to normalize it before giving
it to the network. This has the disadvantage that the vector magni-
tude information, which may be important, is lost.
Another solution is to replace the inner product expression usual-
ly used to calculate net input,
a = compet Wp ,
with a direct calculation of distance,
n i = iw p and a = compet n ,
as is done with the LVQ network. This works and saves the magni-
tude information.
However, a third solution is to append a constant of 1 to each input
vector before normalizing it. Now the change in the added element
will preserve the vector magnitude information.
16-35
Normalize the following vectors using this last method:
p1 = 1 , p2 = 0 , p3 = 0 .
1 1 0
First we add an extra element with value 1 to each vector.
1 0 0
p' 1 = 1 , p' 2 = 1 , p' 3 = 0
1 1 1
Then we normalize each vector.
1 1 1 3
p'' 1 = 1 1 = 1 3
1 1 1 3
0 0 0
p'' 2 = 1 1 = 1 2
1 1 1 2
0 0 0
p'' 3 = 0 0 = 0
1 1 1
Now the third element of each vector contains magnitude information,

since it is equal to the inverse of the magnitude of the extended vectors.
16-36
Epilogue
Epilogue
In this chapter we have demonstrated how the associative instar learning
rule of Chapter 15 can be combined with competitive networks, similar to
the Hamming network of Chapter 3, to produce powerful self-organizing
networks. By combining competition with the instar rule, each of the pro-
totype vectors that are learned by the network become representative of a
particular class of input vector. Thus the competitive networks learn to di-
vide their input space into distinct classes. Each class is represented by one
of the prototype vectors (rows of the weight matrix).
Three types of networks, all developed by Tuevo Kohonen, were discussed
in this chapter. The first is the standard competitive layer. Its simple oper-
ation makes it a practical network for many problems.
The self-organizing feature map is very similar to the competitive layer,
but more closely models biological on-center/off-surround networks. The re-
sult is a network that not only learns to classify input vectors, but also
learns the topology of the input space.
The third network, the LVQ network, uses both unsupervised and super-
vised learning to recognize clusters. It uses a second layer to combine mul-
tiple convex regions into classes that can have any shape. LVQ networks
can even be trained to recognize classes made up of multiple unconnected
regions.
Chapters 18 and 19 will build on the networks presented in this chapter.
For example, Chapter 18 will carry out a more detailed examination of lat-
eral inhibition, on-center/off-surround networks and the biology that in-
spired them. In Chapter 19 we discuss a modification to the standard
competitive network (called adaptive resonance theory), which solves the
weight stability problem that we discussed in this chapter.
Chapter 22 presents practical tips for training competitive networks, and
Chapter 26 is a case study of using self organizing feature maps on a real-
world clustering problem.
16-37
Further Reading
[FrSk91] J. Freeman and D. Skapura, Neural Networks: Algorithms,
Applications, and Programming Techniques, Reading, MA:
Addison-Wesley, 1991.
This text contains code fragments for network algorithms,
making the details of the networks clear.
2nd Ed., Berlin: Springer-Verlag, 1987.
Kohonen introduces the Kohonen rule and several net-
works that use it. It provides a complete analysis of linear
associative models and gives many extensions and exam-
ples.
[Hech90] R. Hecht-Nielsen, Neurocomputing, Reading, MA: Addi-
son-Wesley, 1990.
This book contains a section on the history and mathemat-
ics of competitive learning.
[RuMc86] D. Rumelhart, J. McClelland et al., Parallel Distributed
Processing, vol. 1, Cambridge, MA: MIT Press, 1986.
Both volumes of this set are classics in neural network lit-
erature. The first volume contains a chapter describing
competitive layers and how they learn to detect features.
16-38
Exercises
Exercises
E16.1 Suppose that the weight matrix for layer 2 of the Hamming network is giv-
en by
3 3
1 --- ---
4 4
2 3
W = --- 1 3
--- .
4 4
3 3
--- --- 1
4 4
This matrix violates Eq. (16.6), since
3 1 1
= --- ------------ = --- .
4 S1 2
Give an example of an output from Layer 1 for which Layer 2 will fail to
operate correctly.
E16.2 Consider the input vectors and initial weights shown in Figure E16.1.
p4 p1
p2 p3
3w
2w
1w
Figure E16.1 Cluster Data Vectors

i. Draw the diagram of a competitive network that could classify the
data above so that each of the three clusters of vectors would have
its own class.
ii. Train the network graphically (using the initial weights shown) by
presenting the labeled vectors in the following order:
p1 , p2 , p3 , p4 .
Recall that the competitive transfer function chooses the neuron

with the lowest index to win if more than one neuron has the same
16-39
net input. The Kohonen rule is introduced graphically in Figure

16.3.
iii. Redraw the diagram in Figure E16.1, showing your final weight vec-
tors and the decision boundaries between each region that repre-
sents a class.
E16.3 Train a competitive network using the following input patterns:
p1 = 1 , p = 1 , p = 1 .
2 3
1 1 1
i. Use the Kohonen learning law with = 0.5 , and train for one pass
through the input patterns. (Present each input once, in the order
given.) Display the results graphically. Assume the initial weight
matrix is
W = 2 0 .
0 2
ii. After one pass through the input patterns, how are the patterns
clustered? (In other words, which patterns are grouped together in
the same class?) Would this change if the input patterns were pre-
sented in a different order? Explain.
iii. Repeat part (i) using = 0.25 . How does this change affect the
training?
E16.4 Earlier in this chapter the term conscience was used to refer to a tech-
nique for avoiding the dead neuron problem plaguing competitive layers
and LVQ networks.
Neurons that are too far from input vectors to ever win the competition can
be given a chance by using adaptive biases that get more negative each
time a neuron wins the competition. The result is that neurons that win
very often start to feel guilty until other neurons have a chance to win.
Figure E16.2 shows a competitive network with biases. A typical learning
rule for the bias b i of neuron i is

0.9b i , if i i
old
new
bi = .
b old 0.2 , if i = i
i
16-40
Exercises
p a
2x1 W 3x1
n
3x2
3x1 C
1 b
3x1
2 3
a = compet (Wp + b)
Figure E16.2 Competitive Layer with Biases

i. Examine the vectors in Figure E16.3. Is there any order in which
the vectors can be presented that will cause 1w to win the competi-
tion and move closer to one of the vectors? (Note: assume that adap-
tive biases are not being used.)
p2
3w
p3
2w
p1
1w
Figure E16.3 Input Vectors and Dead Neuron

ii. Given the input vectors and the initial weights and biases defined
below, calculate the weights (using the Kohonen rule) and the bias-
es (using the above bias rule). Repeat the sequence shown below un-
til neuron 1 wins the competition.
p1 = 1 , p2 = 0 , p3 = 1 2
0 1 1 2
0 , w = 2 5 , w = 1 5 , b 0 = b 0 = b 0 = 0
1w = 2 3 1 2 3
1 1 5 2 5
Sequence of input vectors: p 1 , p 2 , p 3 , p 1 , p 2 , p 3 ,
16-41
iii. How many presentations occur before 1w wins the competition?
E16.5 The net input expression for LVQ networks calculates the distance be-
tween the input and each weight vector directly, instead of using the inner
product. The result is that the LVQ network does not require normalized
input vectors. This technique can also be used to allow a competitive layer
to classify nonnormalized vectors. Such a network is shown in Figure
E16.4.
Inputs Competitive Layer
W
2x2
p n1 a1
2x1
-||dist||
2x1 C 2x1
2 2
1
n i = -||iw-p||
1 1
a = compet(n )
Figure E16.4 Competitive Layer with Alternate Net Input Expression

Use this technique to train a two-neuron competitive layer on the (nonnor-
malized) vectors below, using a learning rate, , of 0.5.
p1 = 1 , p2 = 1 , p3 = 2
1 2 2
Present the vectors in the following order:
p1 , p2 , p3 , p2 , p3 , p1 .
Here are the initial weights of the network:
1w = 0 , 2w = 1 .
1 0
E16.6 Repeat E16.5 for the following inputs and initial weights. Show the move-
ments of the weights graphically for each step. If the network is trained for
a large number of iterations, how will the three vectors be clustered in the
final configuration?
16-42
Exercises
p1 = 2 , p2 = 0 , p3 = 2
0 1 2
1w = 1 , 2w = 1 .
0 0
E16.7 We have a competitive learning problem, where the input vectors are
p1 = 0 , p2 = 0 , p3 = 1 , p4 = 2 ,
1 2 1 2
and the initial weight matrix is
W = 1 1 .
1 1
i. Use the Kohonen learning law to train a competitive network using

a learning rate of = 0.5 . (Present each vector once, in the order
shown.) Use the modified competitive network of Figure E16.4,
which uses negative distance, instead of inner product.
ii. Display the results of part i graphically, as in Figure 16.3. (Show all
four iterations.)
iii. Where will the weights eventually converge (approximately)? Ex-
plain. Sketch the approximate final decision boundaries.
E16.8 Show that the modified competitive network of Figure E16.4, which com-
putes distance directly, will produce the same results as the standard com-
petitive network, which uses the inner product, when the input vectors are
normalized.
E16.9 We would like a classifier that divides the interval of the input space de-
fined below into five classes.
0 p1 1
2+2
ans = i. Use MATLAB to randomly generate 100 values in the interval
4 shown above with a uniform distribution.
ii. Square each number so that the distribution is no longer uniform.
iii. Write a MATLAB M-file to implement a competitive layer. Use the
M-file to train a five-neuron competitive layer on the squared values
16-43
until its weights are fairly stable.

iv. How are the weight values of the competitive layer distributed? Is
there some relationship between how the weights are distributed
and how the squared input values are distributed?
E16.10 We would like a classifier that divides the square region defined below into
sixteen classes of roughly equal size.
0 p1 1 , 2 p2 3
2+2
ans = i. Use MATLAB to randomly generate 200 vectors in the region shown
4 above.
ii. Write a MATLAB M-file to implement a competitive layer with Ko-
honen learning. Calculate the net input by finding the distance be-
tween the input and weight vectors directly, as is done by the LVQ
network, so the vectors do not need to be normalized. Use the M-file
to train a competitive layer to classify the 200 vectors. Try different
learning rates and compare performance.
iii. Write a MATLAB M-file to implement a four-neuron by four-neuron
(two-dimensional) feature map. Use the feature map to classify the
same vectors. Use different learning rates and neighborhood sizes,
then compare performance.
E16.11 We want to train the following 1-D feature map (which uses distance in-
stead of inner product to compute the net input):
W 1
4x2
p n a 2
2x1
-||dist|| 4x1 C 4x1
3
2 4 4
ni = -||iw-p||
a = compet(n)
Figure E16.5 1-D Feature Map for Exercise E16.11

T
The initial weight matrix is W 0 = 2 1 1 1 .
2 1 2 0
16-44
Exercises
i. Plot the initial weight vectors as dots, and connect the neighboring
weight vectors as lines (as in Figure 16.10, except that this is a 1-D
feature map).
ii. The following input vector is applied to the network. Perform one it-
eration of the feature map learning rule. (You can do this graphical-
ly.) Use a neighborhood size of 1 and a learning rate of = 0.5 .
T
p1 = 2 0
iii. Plot the new weight vectors as dots, and connect the neighboring
weight vectors as lines.
E16.12 Consider the following feature map, where distance is used instead of inner
product to compute the net input.
W
4x2 a 1 2
4x1
p n
2x1
-||dist|| 4x1 C
3 4
2 4
ni = -||iw - p||
a = compet(n)
Figure E16.6 2-D Feature Map for Exercise E16.12

The initial weight matrix is
T
W = 0 1 1 0
0 0 1 1
i. Plot the initial weights, and show their topological connections, as

in Figure 16.10.
T
ii. Apply the input p = 1 1 , and perform one iteration of the fea-
ture map learning rule, with learning rate of = 0.5 , and neighbor-
hood radius of 1.
iii. Plot the weights after the first iteration, and show their topological
connections.
16-45
E16.13 An LVQ network has the following weights:
0 0
1 0 1 0 0 0 0
1 2
W = 1 0 , W = 0 1 1 0 0 .
0 1 0 0 0 1 1
0 1
i. How many classes does this LVQ network have? How many sub-
classes?
ii. Draw a diagram showing the first-layer weight vectors and the de-
cision boundaries that separate the input space into subclasses.
iii. Label each subclass region to indicate which class it belongs to.
E16.14 We would like an LVQ network that classifies the following vectors accord-
ing to the classes indicated:
1 1 1 1 1 1 1

class 1: 1 1 , class 2:
1 1 1 , class 3: 1 1 .

1 1 1 1 1 1 1
i. How many neurons are required in each layer of the LVQ network?
ii. Define the weights for the first layer.
iii. Define the weights for the second layer.
iv. Test your network for at least one vector from each class.
E16.15 We would like an LVQ network that classifies the following vectors accord-
ing to the classes indicated:
1 p = 1
class 1: p 1 = 1 p 2 = 0 , class 2: p3 = 4
1 2 1 2
i. Could this classification problem be solved by a perceptron? Explain

your answer.
ii. How many neurons must be in each layer of an LVQ network that
can classify the above data, given that each class is made up of two
convex-shaped subclasses?
iii. Define the second-layer weights for such a network.
16-46
Exercises
iv. Initialize the first-layer weights of the network to all zeros and cal-
culate the changes made to the weights by the Kohonen rule (with
a learning rate of 0.5) for the following series of vectors:
p4 , p2 , p3 , p1 , p2 .
v. Draw a diagram showing the input vectors, the final weight vectors
and the decision boundaries between the two classes.
E16.16 An LVQ network has the following weights and training data.
1 0
W = 0 1 ,W = 1 1 0 ,
1 2
0 0 1
0 0
2 t = 1 , p = 2 t = 0 , p = 2 t = 1 ,
p1 = 1 2 2 3 3
2 0 0 1 2 0
2 t = 0
p4 = 4
0 1
i. Plot the training data input vectors and weight vectors (as in Figure
16.14).
ii. Perform four iterations of the LVQ learning rule, with learning rate
= 0.5 , as you present the following sequence of input vectors: p 1 ,
p 2 , p 3 , p 4 (one iteration for each input). Do this graphically, on a
separate diagram from part i.
iii. After completing the iterations in part ii, on a new diagram, sketch
the regions of the input space that make up each subclass and each
class. Label each region to indicate which class it belongs to.

T
W = 0 1 1 0 0 1 1 , W = 1 0 1 0 1 1 0 .
1 2
0 0 0 1 1 1 1 0 1 0 1 0 0 1
classes?
16-47
T
iv. Suppose that an input p = 1 0.5 from Class 1 is presented to the
network. Perform one iteration of the LVQ algorithm, with = 0.5 .

T
W = 0 0 2 1 1 1 , W = 1 1 1 0 0 0 .
1 2
0 2 2 1 1 1 0 0 0 1 1 1
classes?
iv. Perform one iteration of the LVQ algorithm, with the following in-
T T
put/target pair: p = 1 2 , t = 1 0 . Use learning rate
= 0.5 .
16-48
Objectives
17 Radial Basis Networks

Objectives 17-1
Radial Basis Network 17-2
Global vs. Local 17-9
Training RBF Networks 17-9
Linear Least Squares 17-11
Orthogonal Least Squares 17-17
Clustering 17-22
Nonlinear Optimization 17-23
Other Training Techniques 17-25
Epilogue 17-33
Further Reading 17-34 17
Exercises 17-36
Objectives
The multilayer networks discussed in Chapters 11 and 12 represent one
type of neural network structure for function approximation and pattern
recognition. As we saw in Chapter 11, multilayer networks with sigmoid
transfer functions in the hidden layers and linear transfer functions in the
output layer are universal function approximators. In this chapter we will
discuss another type of universal approximation network, the radial basis
function network. This network can be used for many of the same applica-
tions as multilayer networks.
This chapter will follow the structure of Chapter 11. We will begin by dem-
onstrating, in an intuitive way, the universal approximation capabilities of
the radial basis function network. Then we will describe three different
techniques for training these networks. They can be trained by the same
gradient-based algorithms discussed in Chapters 11 and 12, with deriva-
tives computed using a form of backpropagation. However, they can also be
trained using a two-stage process, in which the first layer weights are com-
puted independently from the weights in the second layer. Finally, these
networks can be built in an incremental way - one neuron at a time.
17-1
Theory and Examples

The radial basis function network is related to the multilayer perceptron
network of Chapter 11. It is also a universal approximator and can be used
for function approximation or pattern recognition. We will begin this chap-
ter with a description of the network and a demonstration of its abilities for
function approximation and pattern recognition.
The original work in radial basis functions was performed by Powell and
others during the 1980s [Powe87]. In this original work, radial basis func-
tions were used for exact interpolation in a multidimensional space. In oth-
er words, the function created by the radial basis interpolation was
required to pass exactly through all targets in the training set. The use of
radial basis functions for exact interpolation continues to be an important
application area, and it is also an active area of research.
For our purposes, however, we will not be considering exact interpolation.
Neural networks are often used on noisy data, and exact interpolation often
results in overfitting when the training data is noisy, as we discussed in
Chapter 13. Our interest is in the use of radial basis functions to provide
robust approximations to unknown functions based on generally limited
and noisy measurements. Broomhead and Lowe [BrLo88] were the first to
develop the radial basis function neural network model, which produces a
smooth interpolating function. No attempt is made to force the network re-
sponse to exactly match target outputs. The emphasis is on producing net-
works that will generalize well to new situations.
In the next section we will demonstrate the capabilities of the radial basis
function neural network. In the following sections we will describe proce-
dures for training these networks.
Radial Basis Network

The radial basis network is a two-layer network. There are two major dis-
RBF tinctions between the radial basis function (RBF) network and a two layer
perceptron network. First, in layer 1 of the RBF network, instead of per-
forming an inner product operation between the weights and the input
(matrix multiplication), we calculate the distance between the input vector
and the rows of the weight matrix. (This is similar to the LVQ network
shown in Figure 16.13.) Second, instead of adding the bias, we multiply by
the bias. Therefore, the net input for neuron i in the first layer is calculated
as follows:
1 1 1
n i = p iw b i . (17.1)
17-2
Each row of the weight matrix acts as a center point - a point where the net
input value will be zero. The bias performs a scaling operation on the trans-
fer (basis) function, causing it to stretch or compress.
We should note that most papers and texts on RBF networks use the terms
standard deviation, variance or spread constant, rather than bias. We have
used the bias in order to maintain a consistency with other networks in this
text. This is simply a matter of notation and pedagogy. The operation of the
network is not affected. When a Gaussian transfer function is used, the
bias is related to the standard deviation as follows: b = 1 2 .
The transfer functions used in the first layer of the RBF network are dif-
ferent than the sigmoid functions generally used in the hidden layers of
multilayer perceptrons (MLP). There are several different types of transfer
function that can be used (see [BrLo88]), but for clarity of presentation we
will consider only the Gaussian function, which is the one most commonly
used in the neural network community. It is defined as follows
2
n
a = e , (17.2)
and it is plotted in Figure 17.1.
1.0
0.8
2
0.6 n
a = e
a
0.4
0.2
3 2 1 0 1 2 3
Figure 17.1 Gaussian Basis Function

Local Function A key property of this function is that it is local. This means that the output
is close to zero if you move very far in either direction from the center point.
Global Function This is in contrast to the global sigmoid functions, whose output remains
close to 1 as the net input goes to infinity.
The second layer of the RBF network is a standard linear layer:
2 2 1 2
a = W a +b (17.3)
17-3
Figure 17.2 shows the complete RBF network.
Inputs Radial Basis Layer Linear Layer
W1
S1xR
1 1 2
p a a
1
||dist|| 1 W2 2
R x1 S x1 S x1
n1 S2xS1
n2
.* S1x1 S2x1
1 b 1
1 b 2
1 1
S2
1
R S x1 S S2x1
1 1 1 2 2 1 2
a i = radbas(||iw -p||b i) a = W a +b
Figure 17.2 Radial Basis Network
This RBF network has been shown to be a universal approximator
[PaSa93], just like the MLP network. To illustrate the capability of this
network, consider a network with two neurons in the hidden layer, one out-
put neuron, and with the following default parameters:
1 1 1 1
w 1 1 = 1 , w 2 1 = 1 , b 1 = 2 , b 2 = 2 ,
2 2 2
w 1 1 = 1 , w 1 2 = 1 , b = 0 .
The response of the network with the default parameters is shown in Fig-
2
ure 17.3, which plots the network output a as the input p is varied over
the range 2 2 .
Notice that the response consists of two hills, one for each of the Gaussian
neurons (basis functions) in the first layer. By adjusting the network pa-
rameters, we can change the shape and location of each hill, as we will see
in the following discussion. (As you proceed through this example, it may
be helpful to compare the response of this sample RBF network with the
response of the sample MLP network in Figure 11.5.)
17-4
2
a
1
2 1 0 1 2
Figure 17.3 Default Network Response

Figure 17.4 illustrates the effects of parameter changes on the network re-
sponse. The blue curve is the nominal response. The other curves corre-
spond to the network response when one parameter at a time is varied over
the following ranges:
1 2 1 2
0 w 2 1 2 , 1 w 1 1 1 , 0.5 b 2 8 , 1 b 1 . (17.4)
2 2
1 1
b2 w 2 1
1 1
0 0
1 1
2 1 0 1 2 2 1 0 1 2
(a) (b)
2 2
2 2
w 1 1 b
1 1
0 0
1 1
2 1 0 1 2 2 1 0 1 2
(c) (d)
Figure 17.4 Effect of Parameter Changes on Network Response

Figure 17.4 (a) shows how the network biases in the first layer can be used
to change the width of the hills - the larger the bias, the narrower the hill.
Figure 17.4 (b) illustrates how the weights in the first layer determine the
17-5
location of the hills; there will be a hill centered at each first layer weight.
For multidimensional inputs there will be a hill centered at each row of the
weight matrix. For this reason, each row of the first layer weight matrix is
Center often called the center for the corresponding neuron (basis function).
Notice that the effects of the weight and the bias in first layer of the RBF
network are much different than for the MLP network, which was shown
in Figure 11.6. In the MLP network, the sigmoid functions create steps.
The weights change the slopes of the steps, and the biases change the loca-
tions of the steps.
Figure 17.4 (c) illustrates how the weights in the second layer scale the
height of the hills. The bias in the second layer shifts the entire network
response up or down, as can be seen in Figure 17.4 (d). The second layer of
the RBF network is the same type of linear layer used in the MLP network
of Figure 11.6, and it performs a similar function, which is to create a
weighted sum of the outputs of the layer 1 neurons.
This example demonstrates the flexibility of the RBF network for function
approximation. As with the MLP, it seems clear that if we have enough
neurons in the first layer of the RBF network, we can approximate virtual-
ly any function of interest, and [PaSa93] provides a mathematical proof
that this is the case. However, although both MLP and RBF networks are
universal approximators, they perform their approximations in different
ways. For the RBF network, each transfer function is only active over a
small region of the input space - the response is local. If the input moves
far from a given center, the output of the corresponding neuron will be close
to zero. This has consequences for the design of RBF networks. We must
have centers adequately distributed throughout the range of the network
inputs, and we must select biases in such a way that all of the basis func-
tions overlap in a significant way. (Recall that the biases change the width
of each basis function.) We will discuss these design considerations in more
detail in later sections.
To experiment with the response of this RBF network, use the MATLAB
Neural Network Design Demonstration RBF Network Function (nnd17nf).
Pattern Classification
2 To illustrate the capabilities of the RBF network for pattern classification,
+2
consider again the classic exclusive-or (XOR) problem. The categories for
the XOR gate are

Category 1: p 2 = 1 p 3 = 1 , Category 2: p 1 = 1 p 4 = 1 .
1 1 1 1
17-6
The problem is illustrated graphically in the figure to the left. Because the
two categories are not linearly separable, a single-layer network cannot
p2 p4 perform the classification.
RBF networks can classify these patterns. In fact, there are many different
RBF solutions. We will consider one solution that demonstrates in a simple
p1 p3 way how to use RBF networks for pattern classification. The idea will be to
have the network produce outputs greater than zero when the input is near
patterns p 2 or p 3 , and outputs less than zero for all other inputs. (Note
that the procedures we will use to design this example network are not
suitable for complex problems, but they will help us illustrate the capabil-
ities of the RBF network.)
From the problem statement, we know that the network will need to have
two inputs and one output. For simplicity, we will use only two neurons in
the first layer (two basis functions), since this will be sufficient to solve the
XOR problem. As we discussed earlier, the rows of the first-layer weight
matrix will create centers for the two basis functions. We will choose the
centers to be equal to the patterns p 2 and p 3 . By centering a basis function
at each pattern, we can produce maximum network outputs there. The first
layer weight matrix is then
T
p2
= 1 1 .
1
W = (17.5)
T
p3 1 1
The choice of the bias in the first layer depends on the width that we want
for each basis function. For this problem, we would like the network func-
tion to have two distinct peaks at p 2 and p 3 . Therefore, we dont want the
basis functions to overlap too much. The centers of the basis functions are
each a distance of 2 from the origin. We want the basis function to drop
significantly from its peak in this distance. If we use a bias of 1, we would
get the following reduction in that distance:
2 2
n 1 2 2
a = e = e = e = 0.1353 . (17.6)
Therefore, each basis function will have a peak of 1 at the centers, and will
drop to 0.1353 at the origin. This will work for our problem, so we select the
first layer bias vector to be
b = 1 .
1
(17.7)
1
The original basis function response ranges from 0 to 1 (see Figure 17.1).
We want the output to be negative for inputs much different than p 2 and
p 3 , so we will use a bias of -1 for the second layer, and we will use a value
17-7
of 2 for the second layer weights, in order to bring the peaks back up to 1.
The second layer weights and biases then become
2 2
W = 2 2 , b = 1 . (17.8)
For the network parameter values given in (17.5), (17.7) and (17.8), the net-
work response is shown in Figure 17.5. This figure also shows where the
2
surface intersects the plane at a = 0 , which is where the decision takes
place. This is also indicated by the contours shown underneath the surface.
2
These are the function contours where a = 0 . They are almost circles that
surround the p 2 and p 3 vectors. This means that the network output will
be greater than 0 only when the input vector is near the p 2 and p 3 vectors.
0.5
2 0.5
a
1
1.5
2
3
2
3
1 2
0 1
1 0
p2 2 2
1
p1
3 3
Figure 17.5 Example 2-Input RBF Function Surface

Figure 17.6 illustrates more clearly the decision boundaries. Whenever the
input falls in the blue regions, the output of the network will be greater
than zero. Whenever the network input falls outside the blue regions, the
network output will be less than zero.
17-8
p2 0
3
3 2 1 0 1 2 3
p1
Figure 17.6 RBF Example Decision Regions

This network, therefore, classifies the patterns correctly. It is not the best
solution, in the sense that it does not always assign input patterns to the
closest prototype vector, unlike the MLP solution shown in Figure 11.2.
You will notice that the decision regions for this RBF network are circles,
unlike the linear boundaries that we see in single layer perceptrons. The
MLP can put linear boundaries together to create arbitrary decision re-
gions. The RBF network can put circular boundaries together to create ar-
bitrary decision regions. In this problem, the linear boundaries are more
efficient. Of course, when many neurons are used, and the centers are close
together, the elementary RBF boundaries are no longer purely circular,
and the elementary MLP boundaries are no longer purely linear. However,
associating circular boundaries with RBF networks and linear boundaries
with MLP networks can be helpful in understanding their operation as pat-
tern classifiers.
To experiment with the RBF network for pattern classification, use the
MATLAB Neural Network Design Demonstration RBF Pattern Classifica-
tion (nnd17pc).
Now that we see the power of RBF networks for function approximation
and pattern recognition, the next step is to develop general training algo-
rithms for these networks.
Global vs. Local

Before we discuss the training algorithms, we should say a final word about
the advantages and disadvantages of the global transfer functions used by
the MLP networks and the local transfer functions used by the RBF net-
works. The MLP creates a distributed representation, because all of the
transfer functions overlap in their activity. At any given input value, many
17-9
sigmoid functions in the first layer will have significant outputs. They must
sum or cancel in the second layer in order to produce the appropriate re-
sponse at each point. In the RBF network, each basis function is only active
over a small range of the input. For any given input, only a few basis func-
tions will be active.
There are advantages and disadvantages to each approach. The global ap-
proach tends to require fewer neurons in the hidden layer, since each neu-
ron contributes to the response over a large part of the input space. For the
RBF network, however, basis centers must be spread throughout the range
of the input space in order to provide an accurate approximation. This leads
to the problem of the curse of dimensionality, which we will discuss in the
next section. Also, if more neurons, and therefore more parameters, are
used, then there is a greater likelihood that the network will overfit the
training data and fail to generalize well to new situations.
On the other hand, the local approach generally leads to faster training, es-
pecially when the two-stage algorithms, which will be discussed in the next
section, are used. Also, the local approach can be very useful for adaptive
training, in which the network continues to be incrementally trained while
it is being used, as in adaptive filters (nonlinear versions of the filters in
Chapter 10) or controllers. If, for a period of time, training data only ap-
pears in a certain region of the input space, then a global representation
will tend to improve its accuracy in those regions at the expense of its rep-
resentation in other regions. Local representations will not have this prob-
lem to the same extent. Because each neuron is only active in a small region
of the input space, its weights will not be adjusted if the input falls outside
that region.
Training RBF Networks

Unlike the MLP network, which is almost always trained by some gradi-
ent-based algorithm (steepest descent, conjugate gradient, Levenberg-
Marquardt, etc.), the RBF network can be trained by a variety of approach-
es.
RBF networks can be trained using gradient-based algorithms. However,
because of the local nature of the transfer function and the way in which
the first layer weights and biases operate, there tend to be many more un-
satisfactory local minima in the error surfaces of RBF networks than in
those of MLP networks. For this reason, gradient-based algorithms are of-
ten unsatisfactory for the complete training of RBF networks. They are
used on occasion, however, for fine-tuning of the network after it has been
initially trained using some other method. Later in this chapter we will dis-
cuss the modifications to the backpropagation equations in Chapter 11 that
are needed to compute the gradients for RBF networks.
The most commonly used RBF training algorithms have two stages, which
treat the two layers of the RBF network separately. The algorithms differ
17-10
mainly in how the first layer weights and biases are selected. Once the first
layer weights and biases have been selected, the second layer weights can
be computed in one step, using a linear least-squares algorithm. We will
discuss linear least squares in the next section.
The simplest of the two-stage algorithms arranges the centers (first layer
weights) in a grid pattern throughout the input range and then chooses a
constant bias so that the basis functions have some degree of overlap. This
procedure is not optimal, because the most efficient approximation would
place more basis functions in regions of the input space where the function
to be approximated is most complex. Also, for many practical cases the full
range of the input space is not used, and therefore many basis functions
could be wasted. One of the drawbacks of the RBF network, especially
Curse of Dimensionality when the centers are selected on a grid, is that they suffer from the curse
of dimensionality. This means that as the dimension of the input space in-
creases, the number of basis functions required increases geometrically.
For example, suppose that we have one input variable, and we specify a
grid of 10 basis functions evenly spaced across the range of that input vari-
able. Now, increase the number of input variables to 2. To maintain the
same grid coverage for both input variables, we would need 102, or 100 ba-
sis functions.
Another method for selecting the centers is to select some random subset
of the input vectors in the training set. This ensures that basis centers will
be placed in areas where they will be useful to the network. However, due
to the randomness of the selection, this procedure is not optimal. A more
17
efficient approach is to use a method such as the Kohonen competitive lay-
er or the feature map, described in Chapter 16, to cluster the input space.
The cluster centers then become basis function centers. This ensures that
the basis functions are placed in regions with significant activity. We will
discuss this method in a later section.
A final procedure that we will discuss for RBF training is called orthogonal
least squares. It is based on a general method for building linear models
called subset selection. This method starts with a large number of possible
centerstypically all of the input vectors from the training data. At each
stage of the procedure, it selects one center to add to the first layer weight.
The selection is based on how much the new neuron will reduce the sum
squared error. Neurons are added until some criteria is met. The criteria is
typically chosen to maximize the generalization capability of the network.
Linear Least Squares

In this section we will assume that the first layer weights and biases of the
RBF network are fixed. This can be done by fixing the centers on a grid, or
by randomly selecting the centers from the input vectors in the training
data set (or by using the clustering method which is described in a later sec-
tion). When the centers are randomly selected, all of the biases can be com-
puted using the following formula [Lowe89]:
17-11
1
1 S
b i = ---------- , (17.9)
d max
where d max is the maximum distance between neighboring centers. This is

designed to ensure an appropriate degree of overlap between the basis
functions. Using this method, all of the biases have the same value. There
are other methods which use different values for each bias. We will discuss
one such method later, in the section on clustering.
Once the first layer parameters have been set, the training of the second
layer weights and biases is equivalent to training a linear network, as in
Chapter 10. For example, consider that we have the following training
points
{p 1, t 1} { p 2, t 2} {p Q, tQ} , (17.10)
where p q is an input to the network, and t q is the corresponding target

output. The output of the first layer for each input p q in the training set
can be computed as
1 1 1
n i q = p q iw b i , (17.11)
1 1
a q = radbas n q . (17.12)
Since the first layer weights and biases will not be adjusted, the training
data set for the second layer then becomes
1 1 1
a 1 t 1 a 2 t 2 a Q t Q . (17.13)
The second layer response is linear:

2 2 1 2
a = W a +b . (17.14)
We want to select the weights and biases in this layer to minimize the sum
square error performance index over the training set:
Q
2 T
tq aq
2
Fx = tq aq (17.15)
q=1
Our derivation of the solution to this linear least squares problem will fol-
low the linear network derivation starting with Eq. (10.6). To simplify the
discussion, we will assume a scalar target, and we will lump all of the pa-
rameters we are adjusting, including the bias, into one vector:
17-12
2
x = 1w . (17.16)
2
b
Similarly, we include the bias input 1 as a component of the input vector
1
zq = aq . (17.17)
1
Now the network output, which we usually write in the form
2 2 T 1 2
a q = 1w a q + b , (17.18)
can be written as
T
aq = x zq . (17.19)
This allows us to conveniently write out an expression for the sum square
error:
Q Q Q
F x = eq
2
= tq aq
2
= tq x
T 2
zq . (17.20) 17
q=1 q=1 q=1
To express this in matrix form, we define the following matrices:
T
1u
T
t1 z1 e1
T T
2u z2
t2 , U = ,e = e2 . (17.21)
t = =

tQ T
zQ
T eQ
Qu
The error can now be written
e = t Ux , (17.22)
and the performance index become

T
F x = t Ux t Ux . (17.23)
If we use regularization, as we discussed in Chapter 13, to help in prevent-

ing overfitting, we obtain the following form for the performance index:
17-13
F x = t Ux t Ux + x i = t Ux t Ux + x x , (17.24)
T 2 T T
i=1
where = from Eq. (13.4). Lets expand this expression to obtain

T T T T T T T
F x = t Ux t Ux + x x = t t 2t Ux + x U Ux + x x
(17.25)
T T T T
= t t 2t Ux + x U U + I x
Take a close look at Eq. (17.25), and compare it with the general form of the
quadratic function, given in Eq. (8.35) and repeated here:
T 1 T
F x = c + d x + --- x Ax . (17.26)
2
Our performance function is a quadratic function, where

T T T
c = t t , d = 2U t and A = 2 U U + I . (17.27)
From Chapter 8 we know that the characteristics of the quadratic function

depend primarily on the Hessian matrix A . For example, if the eigenvalues
of the Hessian are all positive, then the function will have one unique glo-
bal minimum.
T
In this case the Hessian matrix is 2 U U + I , and it can be shown that
this matrix is either positive definite or positive semidefinite (see Exercise
E17.4), which means that it can never have negative eigenvalues. We are
left with two possibilities. If the Hessian matrix has only positive eigenval-
ues, the performance index will have one unique global minimum (see Fig-
ure 8.7). If the Hessian matrix has some zero eigenvalues, the performance
index will either have a weak minimum (see Figure 8.9) or no minimum
(see Problem P8.7), depending on the vector d . In this case, it must have a
minimum, since F x is a sum square function, which cannot be negative.
Now lets locate the stationary point of the performance index. From our
previous discussion of quadratic functions in Chapter 8, we know that the
gradient is
F x = c + d x + --- x Ax = d + Ax = 2U t + 2 U U + I x . (17.28)
T 1 T T T
2
The stationary point of F x can be found by setting the gradient equal to

zero:
T T T T
2Z t + 2 U U + I x = 0 U U + I x = U t . (17.29)
Therefore, the optimum weights x can be computed from
17-14
U U + I x = U t .
T T
(17.30)
If the Hessian matrix is positive definite, there will be a unique stationary

point, which will be a strong minimum:
1
x = U U + I U t
T T
(17.31)
Lets demonstrate this procedure with a simple problem.
Example
2 To illustrate the least squares algorithm, lets choose a network and apply
+2
it to a particular problem. We will use an RBF network with three neurons
in the first layer to approximate the following function

g p = 1 + sin --- p for 2 p 2 . (17.32)
4
To obtain our training set we will evaluate this function at six values of p :
p = 2 1.2 0.4 0.4 1.2 2 . (17.33)
This produces the targets
t = 0 0.19 0.69 1.3 1.8 2 . (17.34) 17

We will choose the basis function centers to be spaced equally throughout
the input range: -2, 0 and 2. For simplicity, we will choose the bias to be the
reciprocal of the spacing between points. This produces the following first
layer weight and bias.
2 0.5
1 1
W = 0 , b = 0.5 . (17.35)
2 0.5
The next step is to compute the output of the first layer, using the following
equations.
1 1 1
n i q = p q iw b i , (17.36)
1 1
a q = radbas n q . (17.37)
1
This produces the following a vectors
17-15
1 0.852 0.527 0.237 0.077 0.018

1
a = 0.368 0.698 0.961 0.961 0.698 0.368 (17.38)

0.018 0.077 0.237 0.527 0.852 1
We can use Eq. (17.17) and Eq. (17.21) to create the U and t matrices
1 0.852 0.527 0.237 0.077 0.018

U = 0.368 0.698 0.961 0.961 0.698 0.368 ,
T
(17.39)
0.018 0.077 0.237 0.527 0.852 1
1 1 1 1 1 1
T
t = 0 0.19 0.69 1.3 1.8 2 . (17.40)
The next step is to solve for the weights and biases in the second layer us-
ing Eq. (17.30). We will begin with the regularization parameter set to zero.
1
x = U U + I U t
T T
1
2.07 1.76 0.42 2.71 1.01 1.03
(17.41)
= 1.76 3.09 1.76 4.05 4.05 = 0
0.42 1.76 2.07 2.71 4.41 1.03
2.71 4.05 2.71 6 6 1
The second layer weight and bias are therefore

2 2
W = 1.03 0 1.03 , b = 1 . (17.42)
Figure 17.7 illustrates the operation of this RBF network. The blue line
represents the RBF approximation, and the circles represent the six data
points. The dotted lines in the upper axis represent the individual basis
functions scaled by the corresponding weights in the second layer (includ-
ing the constant bias term). The sum of the dotted lines will produce the
blue line. In the small axis at the bottom, you can see the unscaled basis
functions, which are the outputs of the first layer.
17-16
2.5
1.5
2
a 1
0.5
0.5
1.5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
1
a 0.5
0
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Figure 17.7 RBF Sine Approximation

The RBF network design process can be sensitive to the choice of the center
locations and the bias. For example, if we select six basis functions and six
data points, and if we choose the first layer biases to be 8, instead of 0.5,
then the network response will be as shown in Figure 17.8. The spread of
the basis function decreases as the inverse of the bias. When the bias is this
large, there is not sufficient overlap in the basis functions to provide a
smooth approximation. We match each data point exactly. However, be-
cause of the local nature of the basis function, the approximation to the
true function is not accurate between the training data points.
1.5
2 1
a
0.5
0.5
1
2 1.5 1 0.5 0 0.5 1 1.5 2
1
a 0.5
0
2 1.5 1 0.5 0 0.5 1 1.5 2
Figure 17.8 RBF Response with Bias Too Large
17-17
To experiment with the linear least squares fitting, use the MATLAB Neu-
ral Network Design Demonstration RBF Linear Least Squares (nnd17lls).
Orthogonal Least Squares

In the previous section we assumed that the weights and biases in the first
layer were fixed. (For example, the centers could be fixed on a grid, or ran-
domly selected from the input vectors in the training set.) In this section
we consider a different approach for selecting the centers. We will assume
that there exists a number of potential centers. These centers could include
the entire set of input vectors in the training set, vectors chosen in a grid
pattern, or vectors chosen by any other procedure one might think of. We
will then select vectors one at a time from this set of potential centers, until
the network performance is satisfactory. We will build up the network one
neuron at a time.
The basic idea behind this method comes from statistics, and it is called
Subset Selection subset selection [Mill90]. The general objective of subset selection is to
choose an appropriate subset of independent variables to provide the most
efficient prediction of a target dependent variable. For example, suppose
that we have 10 independent variables, and we want to use them to predict
our target dependent variable. We want to create the simplest predictor
possible, so we want to use the minimum number of independent variables
for the prediction. Which subset of the 10 independent variables should we
use? The optimal approach, called an exhaustive search, tries all combina-
tions of subsets and finds the smallest one that provides satisfactory per-
formance. (We will define later what we mean by satisfactory
performance.)
Unfortunately, this strategy is not practical. If we have Q variables in our
original set, the following expression gives the number of distinct subsets:
Q
Q!
------------------------
q! Q q !
-. (17.43)
q=1
If Q = 10 , this number is 1023. If Q = 20 , the number is more than 1 mil-

lion. We need to have a less expensive strategy than the exhaustive search.
There are several suboptimal procedures. They are not guaranteed to find
the optimal subset, but they require significantly less computation. One
Forward Selection procedure is called forward selection. This method begins with an empty
model and then adds variables one at a time. At each stage, we add the in-
dependent variable that provides the largest reduction in squared error.
We stop adding variables when the performance is adequate. Another ap-
Backward Elimination proach, called backward elimination, starts with all independent variables
selected for the model. At each stage we eliminate the variable that would
cause the least increase in the squared error. The process continues until
the performance is inadequate. There are other approaches which combine
17-18
forward selection and backward elimination, so that variables can be added

and deleted at each iteration.
Any of the standard subset selection techniques can be used for selecting
RBF centers. For purposes of illustration, we will consider one specific form
of forward selection, called orthogonal least squares [ChCo91]. Its main
feature is that it efficiently calculates the error reduction provided by the
addition of each potential center to the RBF network.
To develop the orthogonal least squares algorithm, we begin with Eq.
(17.22), repeated here in slightly different form:
t = Ux + e . (17.44)
We will use our standard notation for matrix rows and columns to individ-
ually identify both the rows and the columns of the matrix U:
T
1u
T
z1
T T
U = 2u = z2 =
u1 u2 un (17.45)

T
zQ
T
Qu
Here each row of the matrix U represents the output of layer 1 of the RBF
network for one input vector from the training set. There will be a column
of the matrix U for each neuron (basis function) in layer 1 plus the bias term
1
( n = S + 1 ). Note that for the OLS algorithm, the potential centers for the
basis functions are often chosen to be all of the input vectors in the training
set. In this case, n will equal Q + 1 , since the constant 1 for the bias term
is included in z , as shown in Eq. (17.17).
Eq. (17.44) is in the form of a standard linear regression model. The matrix
Regression Matrix U is called the regression matrix, and the columns of U are called the re-
gressor vectors.
The objective of OLS is to determine how many columns of U (numbers of
neurons or basis functions) should be used. The first step is to calculate
how much each potential column would reduce the squared error. The prob-
lem is that the columns are generally correlated with each other, and so it
is difficult to determine how much each individual column would reduce
the error. For this reason, we need to first orthogonalize the columns. Or-
thogonalizing the columns means that we can decompose U as follows:
U = MR , (17.46)
where R is an upper triangular matrix, with ones on the diagonal:
17-19
1 r 1 2 r 1 3 r 1 n
0 1 r 2 3 r 2 n
R = , (17.47)

0 0 0 r n 1 n
0 0 0 1
and M is a matrix with orthogonal columns m i . This means that M has

the following properties:
T
v 1 1 0 0 m1 m1 0 0
T 0 v 2 2 0 0
T
m2 m2 0
M M = V = = . (17.48)

0 0 v n n 0 0
T
mn mn
Now Eq. (17.44) can be written
t = MRx + e = Mh + e , (17.49)
where
h = Rx . (17.50)
The least squares solution for Eq. (17.49) is

1
h = M M M t = V M t ,
T T 1 T
(17.51)
and because V is diagonal, the elements of h can be computed

T T
mi t mi t
h i = --------- -.
= -------------
T
(17.52)
v i i mi mi
From h we can compute x using Eq. (17.50). Since R is upper-triangu-

lar, Eq. (17.50) can be solved by back-substitution and does not require a
matrix inversion.
There are a number of ways to obtain the orthogonal vectors m i , but we
will use the Gram-Schmidt orthogonalization process of Eq. (5.20), starting
with the original columns of U .
m1 = u1 , (17.53)
17-20
k1
mk = uk ri k mi , (17.54)
i=1
where
T
mi uk
- , i = 1 k 1 .
r i k = -------------
T
(17.55)
mi mi
Now lets see how orthogonalizing the columns of U enables us to efficient-

ly calculate the squared error contribution of each basis vector. Using Eq.
(17.49), the total sum square value of the targets is given by
T T T T T T T T
t t = Mh + e Mh + e = h M Mh + e Mh + h M e + e e . (17.56)
Consider the second term in the sum:

T T T T T
e Mh = t Mh Mh = t Mh h M Mh . (17.57)
If we use the optimal h from Eq. (17.51), we find

T
e Mh = t Mh t MV M Mh = t Mh t Mh = 0 .
T T T 1 T T
(17.58)
Therefore the total sum square value from Eq. (17.56) becomes
n
hi mi mi + e
T T T T T T 2 T T
t t = h M Mh + e e = h Vh + e e = e. (17.59)
i=1
The first term on the right of Eq. (17.59) is the contribution to the sum
squared value explained by the regressors, and the second term is the re-
maining sum squared value that is not explained by the regressors. There-
fore, regressor (basis function) i contributes
2 T
hi mi mi (17.60)
to the squared value. This also represents how much the squared error can
be reduced by including the corresponding basis function in the network.
We will use this number, after normalizing by the total squared value, to
determine the next basis function to include at each iteration:
2 T
hi mi mi
-.
o i = -------------------
T
(17.61)
t t
This number always falls between zero and one.
17-21
Now lets put all these ideas together into an algorithm for selecting cen-
ters.
The OLS Algorithm

To begin the algorithm, we start with all potential basis functions included
in the regression matrix U . (As we explained below Eq. (17.45), if all input
vectors in the training set are to be considered potential basis function cen-
ters, then the U matrix will be Q by Q + 1 .) This matrix represents only
potential basis functions, since we start with no basis functions included in
the network.
The first stage of the OLS algorithm consists of the following three steps,
for i = 1 Q :
i
m1 = ui , (17.62)
iT
i m1 t
-,
h 1 = --------------------- (17.63)
i T i
m1 m1
i 2 iT i
i h1 m1 m1
o 1 = -------------------------------------
T
. (17.64)
t t
We then select the basis function that creates the largest reduction in er-
ror:
i i
o 1 = o 1 1 = max o 1 , (17.65)
i
m 1 = m 1 1 = u i1 . (17.66)
The remaining iterations of the algorithm continue as follows (for iteration

k):
For i = 1 Q , i i 1 , ..., i i k 1
T
i mj ui
- , j = 1 k 1 ,
r j k = -------------
T
(17.67)
mj mj
k1
i i
mk = ui rj k mj , (17.68)
j=1
17-22
iT
i mk t
-,
h k = --------------------- (17.69)
i T i
mk mk
i 2 iT i
i hk mk mk
ok = -------------------------------------
T
, (17.70)
t t
ik i
ok = ok = max o k , (17.71)
i
r j k = r j k , j = 1 k 1 .
k
(17.72)
i
mk = mk .
k
(17.73)
The iterations continue until some stopping criterion is met. One choice of
stopping criterion is
k
1 oj , (17.74)
j=1
where is some small number. However, if is chosen too small, we can

have overfitting, since the network will become too complex. An alternative
is to use a validation set, as we discussed in the chapter on generalization.
We would stop when the error on the validation set increased.
After the algorithm has converged, the original weights x can be computed
from the transformed weights h by using Eq. (17.50). This produces, by
back substitution,
n
xn = hn , xk = hk r j k x j , (17.75)
j = k+1
where n is the final number of weights and biases in the second layer (ad-
justable parameters).
To experiment with orthogonal least squares learning, use the MATLAB
Neural Network Design Demonstration RBF Orthogonal Least Squares
(nnd17ols).
Clustering
There is another approach [MoDa89] for selecting the weights and biases
in the first layer of the RBF network. This method uses the competitive net-
works described in Chapter 16. Recall that the competitive layer of Ko-
honen (see Figure 16.2) and the Self Organizing Feature Map (see Figure
17-23
16.9) perform a clustering operation on the input vectors of the training set.
After training, the rows of the competitive networks contain prototypes, or
cluster centers. This provides an approach for locating centers and select-
ing biases for the first layer of the RBF network. If we take the input vec-
tors from the training set and perform a clustering operation on them, the
resulting prototypes (cluster centers) could be used as centers for the RBF
network. In addition, we could compute the variance of each individual
cluster and use that number to calculate an appropriate bias to be used for
the corresponding neuron.
Consider again the following training set:
{p 1, t 1} { p 2, t 2} {p Q, tQ} . (17.76)
We want to perform a clustering of the input vectors from this training set:
p 1 p 2 p Q . (17.77)
We will train the first layer weights of the RBF network to perform a clus-
tering of these vectors, using the Kohonen learning rule of Eq. (16.13), and
repeated here:
1 1 1
i
w q = i
w q 1 + p q iw q 1 , (17.78)
1
where p q is one of the input vectors in the training set, and iw q 1 is
the weight vector that was closest to p q . (We could also use other cluster-
ing algorithms, such as the Self Organizing Feature Map, or the k-means
clustering algorithm, which was suggested in [MoDa89].) As described in
Chapter 16, Eq. (17.78) is repeated until the weights have converged. The
resulting converged weights will represent cluster centers of the training
set input vectors. This will insure that we will have basis functions located
in areas where input vectors are most likely to occur.
In addition to selecting the first layer weights, the clustering process can
provide us with a method for determining the first layer biases. For each
neuron (basis function), locate the n c input vectors from the training set
that are closest to the corresponding weight vector (center). Then compute
the average distance between the center and its neighbors.
1
nc ---
1 1 2
2
dist i = ----- p j iw
i
(17.79)
nc
j=1
i 1 i
where p1 is the input vector that closest to iw , and is p 2 the next closest
input vector. From these distances, [MoDa89] recommends setting the first
layer biases as follows:
17-24
1 1
b i = ------------------ . (17.80)
2dist i
Therefore, when a cluster is wide, the corresponding basis function will be

wide as well. Notice that in this case each bias in the first layer will be dif-
ferent. This should provide a network that is more efficient in its use of ba-
sis functions than a network with equal biases.
After the weights and biases of the first layer are determined, linear least
squares is used to find the second layer weights and biases.
There is a potential drawback to the clustering method for designing the
first layer of the RBF network. The method only takes into account the dis-
tribution of the input vectors; it does not consider the targets. It is possible
that the function we are trying to approximate is more complex in regions
for which there are fewer inputs. For this case, the clustering method will
not distribute the centers appropriately. On the other hand, one would
hope that the training data is located in regions where the network will be
most used, and therefore the function approximation will be most accurate
in those areas.
Nonlinear Optimization
It is also possible to train RBF networks in the same manner as MLP net-
worksusing nonlinear optimization techniques, in which all weights and
biases in the network are adjusted at the same time. These methods are not
generally used for the full training of RBF networks, because these net-
works tend to have many more unsatisfactory local minima in their error
surfaces. However, nonlinear optimization can be used for the fine-tuning
of the network parameters, after initial training by one of the two-stage
methods we presented in earlier sections.
We will not present the nonlinear optimization methods in their entirety
here, since they were treated extensively in Chapters 11 and 12. Instead,
we will simply indicate how the basic backpropagation algorithm for com-
puting the gradient in MLP networks can be modified for RBF networks.
The derivation of the gradient for RBF networks follows the same pattern
as the gradient development for MLP networks, starting with Eq. (11.9),
which you may wish to review at this time. Here we will only discuss the
one step where the two derivations differ. The difference occurs with Eq.
(11.20). The net input for the second layer of the RBF network has the same
form as its counterpart in the MLP network, but the first layer net input
has a different form (as given in Eq. (17.1) and repeated here):
1
S
2
pj w i j
1 1 1 1 1
n i = p iw b i = b i . (17.81)
j=1
17-25
If we take the derivative of this function with respect to the weights and
biases, we obtain
1 1 1
n i 1 12 1 b i w i j p j
1
- = b i --------------------------------------
---------- - 2 p j w i j 1 = -,
----------------------------
1
(17.82)
w i j p iw
1
S
1 2
pj wi j
j=1
1
n i 1
-------1- = p iw . (17.83)
b i
This produces the modified gradient equations (compare with Eq. (11.23)
and Eq. (11.24)) for Layer 1 of the RBF network
1 1
F 1 b i w i j p j
----------
1
- = s i -,
----------------------------
1
(17.84)
w i j p iw
F 1 1
-------1- = s i p iw . (17.85)
b i
Therefore, if we look at the summary of the gradient descent backpropaga-

tion algorithm for MLP networks, from Eq. (11.44) to Eq. (11.47), we find
that the only difference for RBF networks is that we substitute Eq. (17.84)
and Eq. (17.85) for Eq. (11.46) and Eq. (11.47) when m = 1 . When m = 2
the original equations remain the same.
To experiment with nonlinear optimization learning, use the MATLAB
Neural Network Design Demonstration RBF Nonlinear Optimization
(nnd17no).
Other Training Techniques

In this chapter we have only touched the surface of the variety of training
techniques that have been proposed for RBF networks. We have attempted
to present the principal concepts, but there are many variations. For exam-
ple, the OLS algorithm has been extended to handle multiple outputs
[ChCo92] and regularized performance indices [ChCh96]. It has also been
used in combination with a genetic algorithm [ChCo99], which was used to
select the first layer biases and the regularization parameter. The expecta-
tion maximization algorithm has also been suggested by several authors
for optimizing the center locations, starting with [Bish91]. [OrHa00] used
a regression tree approach for center selection. There have also been many
variations on the use of clustering and on the combination of clustering for
initialization and nonlinear optimization for fine-tuning. The architecture
of the RBF network lends itself to many training approaches.
17-26
Summary of Results
Summary of Results

Inputs Radial Basis Layer Linear Layer
W1
S1xR
1
p a1 a2
1
||dist|| W2
R x1 S1x1 S2x1
n1 S2xS1
n2
.* S1x1 S2x1
1 b
1
1 b
2
1 1
S2
1
R S x1 S S2x1
1 1 1 2 2 1 2
a i = radbas(||iw -p||b i) a = W a +b

Linear Least Squares 17
2
1w
1
x = , zq = aq
2
b 1
T
1u
T
t1 z1 e1
T T
2u z2
t2 e2
t = ,U = = ,e =

tQ T
zQ eQ
T
Qu
T T
F x = t Ux t Ux + x x
U U + I x = U t
T T
17-27
Orthogonal Least Squares

Step 1
i
m1 = ui ,
iT
i m1 t
-,
h 1 = ---------------------
i T i
m1 m1
i 2 iT i
i h1 m1 m1
o 1 = -------------------------------------
T
.
t t
i i
o 1 = o 1 1 = max o 1
i
m 1 = m 1 1 = u i1 .
Step k
For i = 1 Q , i i 1 , ..., i i k 1
T
i mj uk
- , j = 1 k ,
r j k = -------------
T
mj mj
k1
i i
mk = ui r j k mj ,
j=1
iT
i mk t
hk -,
= ---------------------
i T i
mk mk
i 2 iT i
i hk mk mk
ok = -------------------------------------
T
,
t t
i i
o k = o k k = max o k ,
i
mk = mk k .
17-28
Summary of Results
Clustering
Training the weights
1 1 1
i
w q = i
w q 1 + p q iw q 1
Selecting the bias

1
nc ---
1 1 2
2
dist i = ----- p j iw
i
nc
j=1
1 1
b i = ------------------
2dist i
Nonlinear Optimization
Replace Eq. (11.46) and Eq. (11.47) in standard backpropagation with
1 1
F 1 b i w i j p j
----------
1
- = s i -,
----------------------------
1
w i j p iw
F 1 1
-------1- = s i p iw .
b i
17-29
Solved Problems
P17.1 Use the OLS algorithm, to approximate the following function:
g p = cos p for 1 p 1 .
To obtain our training set we will evaluate this function at five val-
ues of p :
p = 1 0.5 0 0.5 1 .
This produces the targets
t = 1 0 1 0 1 .
Perform one iteration of the OLS algorithm. Assume that the in-
puts in the training set are the potential centers and that the bias-
es are all equal to 1.
First, we compute the outputs of the first layer:
1 1 1
n i q = p q iw b i ,
1 1
a q = radbas n q ,

1.000 0.779 0.368 0.105 0.018
0.779 1.000 0.779 0.368 0.105
1
a = 0.368 0.779 1.000 0.779 0.368 .

0.105 0.368 0.779 1.000 0.779
0.018 0.105 0.368 0.779 1.000

We can use Eq. (17.17) and Eq. (17.21) to create the U and t matrices:
1.000 0.779 0.368 0.105 0.018

0.779 1.000 0.779 0.368 0.105
U = 0.368 0.779 1.000 0.779 0.368 ,
T
0.105 0.368 0.779 1.000 0.779

0.018 0.105 0.368 0.779 1.000
1.000 1.000 1.000 1.000 1.000
T
t = 1 0 1 0 1 .
17-30
Solved Problems
Now we perform step one of the algorithm:

i
m1 = ui ,
1.000 0.779 0.368 0.105

0.779 1.000 0.779 0.368
1 2 3 4
m1 = 0.368 , m 1 = 0.779 , m 1 = 1.000 , m 1 = 0.779 ,
0.105 0.368 0.779 1.000
0.018 0.105 0.368 0.779
0.018 1.000
0.105 1.000
5 6
m1 = 0.368 , m 1 = 1.000 ,
0.779 1.000
1.000 1.000
iT
i m1 t
h1 -,
= ---------------------
i T i
m1 m1
1 2 3 4 5
h1 = 0.371 , h 1 = 0.045 , h 1 = 0.106 , h 1 = 0.045 , h 1 = 0.371 ,
6
h1 = 0.200 ,
i 2 iT i
i h1 m1 m1
o1 = -------------------------------------
T
,
t t
1 2 3 4 5
o1 = 0.0804 , o 1 = 0.0016 , o 1 = 0.0094 , o 1 = 0.0016 , o 1 = 0.0804 ,
6
o1 = 0.0667 .
We see that the first and fifth centers would produce a 0.0804 reduction in
the error. This means that the error would be reduced by 8.04%, if the first
or fifth center were used in a single-neuron first layer. We would typically
select the first center, since it has the smallest index.
If we stop at this point, we would add the first center to the hidden layer.
2 1
Using Eq. (17.75), we would find that w 1 1 = x 1 = h 1 = h 1 = 0.371 . Also,
6
b 2 = 0 , since the bias center, m 1 , was not selected on the first iteration.
Note that if we continue to add neurons in the hidden layer, the first weight
will change. This can be seen from Eq. (17.75). This equation to find x k is
only used after all of the h k are found. Only x n will exactly equal h n .
17-31
If we continue the algorithm, the first column would be removed from U .

We would then orthogonalize all remaining columns of U with respect to
m 1 , which was chosen on the first iteration, using Eq. (17.54). It is inter-
esting to note that the error reduction on the second iteration would be
much higher than the reduction on the first iteration. The sequence of re-
ductions would be 0.0804, 0.3526, 0.5074, 0.0448, 0.0147, 0, and the centers
would be chosen in the following order: 1, 2, 5, 3, 4, 6. The reason that re-
ductions in later iterations are higher is that it takes a combination of basis
functions to produce the best approximation. This is why forward selection
is not guaranteed to produce the optimal combination, which can be found
with an exhaustive search. Also, notice that the bias is selected last, and it
produces no reduction in the error.
P17.2 Figure P17.1 illustrates a classification problem, where Class I vec-

tors are represented by dark circles, and Class II vectors are rep-
resented by light circles. These categories are not linearly
separable. Design a radial basis function network to correctly clas-
sify these categories.
Figure P17.1 Classification Problem for Problem P17.2

From the problem statement, we know that the network will need to have
two inputs, and we can use one output to distinguish the two classes. We
will choose a positive output for Class I vectors, and a negative output for
Class II vectors. The Class I region is made up of two simple subregions,
and it appears that two neurons should be sufficient to perform the classi-
fication. The rows of the first-layer weight matrix will create centers for the
two basis functions, and we will choose each center to be located in the mid-
dle of one subregion. By centering a basis function in each subregion, we
can produce maximum network outputs there. The first layer weight ma-
trix is then
1
W = 1 1 .
2.5 2.5
The choice of the biases in the first layer depends on the width that we
want for each basis function. For this problem, the first basis function
should be wider than the second. Therefore, the first bias will be smaller
than the second bias. The boundary formed by the first basis function
17-32
Solved Problems
should have a radius of approximately 1, while the second basis function

boundary should have a radius of approximately 1 2 . We want the basis
functions to drop significantly from their peaks in these distances. If we use
a bias of 1 for the first neuron and a bias of 2 for the second neuron, we get
the following reductions within one radius of the centers:
2 2 2 2
n 1 1 1 n 2 0.5 1
a = e = e = e = 0.3679 , a = e = e = e = 0.3679
This will work for our problem, so we select the first layer bias vector to be
b = 1 .
1
The original basis function response ranges from 0 to 1 (see Figure 17.1).
We want the output to be negative for inputs outside the decision regions,
so we will use a bias of -1 for the second layer, and we will use a value of 2
for the second layer weights, in order to bring the peaks back up to 1. The
second layer weights and biases then become
2 2
W = 2 2 , b = 1 .
For these network parameter values, the network response is shown on the
right side of Figure P17.2. This figure also shows where the surface inter-
2
sects the plane at a = 0 , which is where the decision takes place. This is
also indicated by the contours shown underneath the surface. These are the
2
function contours where a = 0 . These decision regions are shown more
clearly on the left side of Figure P17.2.
3.5
3
0.5
2
2.5 a 0
p2 2
0.5
1.5
1.5
1 2
0
1 4
0.5 3
2
p1 3
1
2
p2
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4 0
p1
Figure P17.2 Decision Regions for Problem P17.2
17-33
P17.3 For an RBF network with one input and one neuron in the hidden
layer, the initial weights and biases are chosen to be
1 1 2 2
w 0 = 0 , b 0 = 1 , w 0 = 2 , b 0 = 1 .
p = 1 t = 1 .
Perform one iteration of steepest descent backpropagation with

= 1.
The first step is to propagate the input through the network.
1 1 1 2
n = p w b = 1 1 0 = 1
1 2
1 1 n 1
a = radbas n = e = e = 0.3679
2 2 1 2
n = w a + b = 2 0.3679 + 1 = 0.2642
2 2 2
a = purelin n = n = 0.2642
2
e = t a = 1 0.2642 = 0.7358
Now we backpropagate the sensitivities using Eq. (11.44) and Eq. (11.45).
s = 2F (n ) t a = 2 1 e = 2 1 0.7358 = 1.4716
2 2 2
1 2
1 n
s = F (n ) W s = 2 n e
1 1 1 2 T 2 2 2 1
w s = 2 1 e 2 1.4716 = 2.1655
Finally, the weights and biases are updated using Eq. (11.46) and Eq.
(11.47) for Layer 2, and Eq. (17.84) and Eq. (17.85) for Layer 1:
2 2 2 1 T
w 1 = w 0 s a = 2 1 1.4716 0.3679 = 1.4586 ,
1 b w p
1 1
1 0 1
= 0 1 2.1655 ---------------------------- = 2.1655 ,
1 1
w 1 = w 0 s -------------------------
pw
1 10
2 2 2
b 1 = b 0 s = 1 1 1.4716 = 2.4716 ,
1 1 1 1
b 1 = b 0 s p w = 1 1 2.1655 1 0 = 3.1655 .
17-34
Epilogue
Epilogue
The radial basis function network is an alternative to the multilayer per-
ceptron network for problems of function approximation and pattern recog-
nition. In this chapter we have demonstrated the operation of the RBF
network, and we have described several techniques for training the net-
work. Unlike MLP training, RBF training usually consists of two stages. In
the first stage, the weights and biases in the first layer are found. In the
second stage, which typically involves linear least squares, the second layer
weights and biases are calculated.
17
17-35
Further Reading
[Bish91] C. M. Bishop, Improving the generalization properties of
radial basis function neural networks, Neural Computa-
tion, Vol. 3, No. 4, pp. 579-588, 1991.
First published use of the expectation-maximization algo-
rithm for optimizing cluster centers for the radial basis net-
work.
[BrLo88] D.S. Broomhead and D. Lowe, Multivariable function in-
terpolation and adaptive networks, Complex Systems,
vol.2, pp. 321-355, 1988.
This seminal paper describes the first use of radial basis
functions in the context of neural networks.
[ChCo91] S. Chen, C.F.N. Cowan, and P.M. Grant, Orthogonal least
squares learning algorithm for radial basis function net-
works, IEEE Transactions on Neural Networks, Vol.2,
No.2, pp.302-309, 1991.
The first description of the use of the subset selection tech-
nique for selecting centers for radial basis function net-
works.
[ChCo92] S. Chen, P. M. Grant, and C. F. N. Cowan, Orthogonal
least squares algorithm for training multioutput radial ba-
sis function networks, Proceedings of the Institute of Elec-
trical Engineers, Vol. 139, Pt. F, No. 6, pp. 378384, 1992.
This paper extends the orthogonal least squares algorithm
to the case of multiple outputs.
[ChCh96] S. Chen, E. S. Chng, and K. Alkadhimi, Regularised or-
thogonal least squares algorithm for constructing radial
basis function networks, International Journal of Control,
Vol. 64, No. 5, pp. 829837, 1996.
Modifies the orthogonal least squares algorithm to handle
regularized performance indices.
[ChCo99] S. Chen, C.F.N. Cowan, and P.M. Grant, Combined Genet-
ic Algorithm Optimization and Regularized Orthogonal
Least Squares Learning for Radial Basis Function Net-
works, IEEE Transactions on Neural Networks, Vol.10,
No.5, pp.302-309, 1999.
Combines a genetic algorithm with orthogonal least
squares to compute the regularization parameter and basis
17-36
Further Reading
function spread, while also selecting centers and solving for

the optimal second layer weights of radial basis function
networks.
[Lowe89] D. Lowe, Adaptive radial basis function nonlinearities,
and the problem of generalization, Proceedings of the First
IEE International Conference on Artificial Neural Net-
works, pp. 171 - 175, 1989.
This paper describes the use of gradient-based algorithms
for training all of the parameters of an RBF network, in-
cluding basis function centers and widths. It also provides
a formula for setting the basis function widths, if the cen-
ters are randomly chosen from the training data set.
[Mill90] A.J. Miller, Subset Selection in Regression. Chapman and
Hall, N.Y., 1990.
This book provides a very complete and clear discussion of
the general problem of subset selection. This involves
choosing an appropriate subset from a large set of indepen-
dent input variables, in order to provide the most efficient
prediction of some dependent variable.
[MoDa89] J. Moody and C.J. Darken, Fast Learning in Networks of
Locally-Tuned Processing Units, Neural Computation,
Vol. 1, pp. 281294, 1989.
17
The first published use of clustering methods to find radial
basis function centers and variances.
[OrHa00] M. J. Orr, J. Hallam, A. Murray, and T. Leonard, Assess-
ing rbf networks using delve, IJNS, 2000.
This paper compares a variety of methods for training radi-
al basis function networks. The methods include forward
selection with regularization and also regression trees.
[PaSa93] J. Park and I.W. Sandberg, Universal approximation us-
ing radial-basis-function networks, Neural Computation,
vol. 5, pp. 305-316, 1993.
This paper proves the universal approximation capability
of radial basis function networks.
[Powe87] M.J.D. Powell, Radial basis functions for multivariable in-
terpolation: a review, Algorithms for Approximation, pp.
143-167, Oxford, 1987.
This paper provides the definitive survey of the original
work on radial basis functions. The original use of radial
basis functions was in exact multivariable interpolation.
17-37
Exercises
E17.1 Design an RBF network to perform the classification illustrated in Figure
E17.1. The network should produce a positive output whenever the input
vector is in the shaded region and a negative output otherwise.
Figure E17.1 Pattern Classification Regions
E17.2 Choose the weights and biases for an RBF network with two neurons in the
hidden layer and one output neuron, so that the network response passes
through the points indicated by the blue circles in Figure E17.2.
Use the MATLAB Neural Network Design Demonstration RBF Network
Function (nnd17nf) to check your result.
1
2 1 0 1 2
Figure E17.2 Function Approximation Exercise
17-38
Exercises
E17.3 Consider a 1-2-1 RBF network (two neurons in the hidden layer and one
output neuron). The first layer weights and biases are fixed as follows:
W = 1 , b = 0.5 .
1 1
1 0.5
2
Assume that the bias in the second layer is fixed at 0 ( b = 0 ). The training
set has the following input/target pairs:
p 1 = 1 t 1 = 1 , p 2 = 0 t 2 = 0 , p 3 = 1 t 3 = 1 .
i. Use linear least squares to solve for the second layer weights, as-
suming that the regularization parameter = 0 .
ii. Plot the contour plot for the sum squared error. Recall that it will be
a quadratic function. (See Chapter 8.)
2+2
ans = iii. Write a MATLAB M-file to check your answers to parts i. and ii.
4
iv. Repeat parts i. to iii., with = 4 . Plot regularized squared error.
E17.4 The Hessian matrix for the performance index of the RBF network, given
in Eq. (17.25), is
T
2 U U + I .
Show that this matrix is at least positive semidefinite for 0 , and show
that it is positive definite if 0 .
E17.5 Consider an RBF network with the weights and biases in the first layer
fixed. Show how the LMS algorithm of Chapter 10 could be modified for
learning the second layer weights and biases.
E17.6 Suppose that a Gaussian transfer function in the first layer of the RBF net-
work is replaced with a linear transfer function.
i. In Solved Problem P11.8, we showed that a multilayer perceptron
with linear transfer functions in each layer is equivalent to a single-
layer perceptron. If we use a linear transfer function in each layer
of an RBF network, is that equivalent to a single-layer network? Ex-
plain.
2+2
ans = ii. Work out an example, equivalent to Figure 17.4, to demonstrate the
4 operation of the RBF network with linear transfer function in the
first layer. Use MATLAB to plot your figures. Do you think that
the RBF network will be a universal approximator, if the first layer
transfer function is linear? Explain your answer.
17-39
E17.7 Consider a Radial Basis Network, as in Figure 17.2, but assume that there
is no bias in the second layer. There are two neurons in the first layer (two
basis functions). The first layer weights (centers) and biases are fixed, and
we have three inputs and targets. The outputs of the first layer and the tar-
gets are given by

a = 2 1 0 , t = 0 1 2 .
1
1 2 1
i. Use linear least squares to find the second layer weights of the net-
work.
ii. Assume now that the basis function centers in the first layer are
only potential centers. If orthogonal least squares is used to select
potential centers, which center will be selected first, what will be its
corresponding weight in the second layer, and how much will it re-
duce the squared error? Show all calculations clearly and in order.
iii. Is there a relationship between the two weights that you computed
in part i. and the single weight that you computed in part ii? Ex-
plain.
E17.8 Repeat E17.7 for the following data:

a = 1 2 1 , t = 1 2 1 .
1
i.
2 1 1

a = 2 1 0 , t = 3 1 2 .
1
ii.
1 1 1
E17.9 Consider the variation of the radial basis network shown in Figure E17.3.
The inputs and targets in the training set are p 1 = 1 t 1 = 1 ,
p 2 = 1 t 2 = 1 .
2
i. Find the linear least squares solution for the weight matrix W .
2
ii. For the weight matrix W that you found in part i., sketch the net-
work response as the input varies from -2 to 2.
17-40
Exercises
Inputs Linear Radial Basis Layer Linear Layer
W1
2x1
p a1 a2
||dist|| W2
1x1 2x1 1x1
n1 1x2
n2
.* 2x1 1x1
1 b
1
1 2x1 2 1
1 1 1 2 2 1
a i = ||iw -p||b i a =Wa
Figure E17.3 Radial Basis Network for Exercise E17.9
E17.10 Write a MATLAB program to implement the linear least squares algo-
1
rithm for the 1 S 1 RBF network with first layer weights and biases
fixed. Train the network to approximate the function

g p = 1 + sin --- p for 2 p 2 .
8
2+2
ans = i. Select 10 training points at random from the interval 2 p 2 .
4
ii. Select four basis function centers evenly spaced on the interval
2 p 2 . Then, use Eq. (17.9) to set the bias. Finally, use linear
least squares to find the second layer weights and biases, assuming
that there is no regularization. Plot the network response for
2 p 2 , and show the training points on the same plot. Compute
the sum squared error over the training set.
iii. Double the bias from part ii and repeat.
iv. Decrease the bias by half from part ii, and repeat.
v. Compare the final sum squared errors for all cases and explain your
results.
E17.11 Use the function described in Exercise E17.10, and use an RBF network
with 10 neurons in the hidden layer.
i. Repeat Exercise E17.10 ii. with regularization parameter = 0.2 .
Describe the changes in the RBF network response.
2+2
ans = ii. Add uniform random noise in the range 0.1 0.1 to the training
4 targets. Repeat Exercise E17.10 ii. with no regularization and with
regularization parameter = 0.2 2 20 . Which case produces the
best results. Explain.
17-41
E17.12 Write a MATLAB program to implement the orthogonal least squares al-
gorithm. Repeat Exercise E17.10 using the orthogonal least squares algo-
2+2
rithm. Use the 10 random training point inputs as the potential centers,
ans = and use Eq. (17.9) to set the bias. Use only the first four selected centers.
4 Compare your final sum squared errors with the result from E17.10 part ii.
E17.13 Write a MATLAB program to implement the steepest descent algorithm

1
for the 1 S 1 RBF network. Train the network to approximate the func-
tion

g p = 1 + sin --- p for 2 p 2 .
8
2+2
ans = You should be able to use a slightly modified version of the program you
4 wrote for Exercise E11.25.
i. Select 10 data points at random from the interval 2 p 2 .
ii. Initialize all parameters (weights and biases in both layers) as
small random numbers, and then train the network to convergence.
(Experiment with the learning rate , to determine a stable value.)
Plot the network response for 2 p 2 , and show the training
points on the same plot. Compute the sum squared error over the
training set. Use 2, 4 and 8 centers. Try different sets of initial
weights.
iii. Repeat part ii., but use a different method for initializing the pa-
rameters. Start by setting the parameters as follows. First, select
basis function centers evenly spaced on the interval 2 p 2 .
Then, use Eq. (17.9) to set the bias. Finally, use linear least squares
to find the second layer weights and biases. Compute the squared
error for these initial weights and biases. Starting from these initial
conditions, train all parameters with steepest descent.
iv. Compare the final sum squared errors for all cases and explain your
results.
E17.14 Suppose that a radial basis function layer (Layer 1 of the RBF network)
were used in the second or third layer of a multilayer network. How could
you modify the backpropagation equation, Eq. (11.35), to accommodate this
change. (Recall that the weight update equations would be modified from
Eq. (11.23) and Eq. (11.24) to Eq. (17.84) and Eq. (17.85).)
E17.15 Consider again Exercise E16.10, in which you trained a feature map to
cluster the input space
0 p1 1 , 2 p2 3 .
17-42
Exercises
Assume that over this input space, we wish to use an RBF network to ap-
proximate the following function:
t = sin 2p 1 cos 2p 2 .
2+2
ans = i. Use MATLAB to randomly generate 200 input vectors in the region
4 shown above.
ii. Write a MATLAB M-file to implement a four-neuron by four-neuron
(two-dimensional) feature map. Calculate the net input by finding
the distance between the input and weight vectors directly, as is
done by the LVQ network, so the vectors do not need to be normal-
ized. Use the feature map to cluster the input vectors.
iii. Use the trained feature map weight matrix from part ii as the
weight matrix of the first layer of an RBF network. Use Eq. (17.79)
to determine the average distance between each cluster and its cen-
ter, and then use Eq. (17.80) to set the bias for each neuron in the
first layer of the RBF network.
iv. For each of the 200 input vectors in part i, compute the target re-
sponse for the function above. Then use the resulting input/target
pairs to determine the second-layer weights and bias for the RBF
network.
v. Repeat parts ii to iv, using a two by two feature map. Compare your
results.
17-43
Objectives
Objectives 18-1
Biological Motivation: Vision 18-3
Illusions 18-4
Vision Normalization 18-8
Basic Nonlinear Model 18-9
Two-Layer Competitive Network 18-12
Layer 1 18-13
Layer 2 18-17
Choice of Transfer Function 18-20
Learning Law 18-22
Relation to Kohonen Law 18-24
Epilogue 18-42
Exercises 18-45
Objectives
This chapter is a continuation of our discussion of associative and compet-
itive learning algorithms in Chapters 15 and 16. The Grossberg network
described in this chapter is a self-organizing continuous-time competitive
network. This will be the first time we have considered continuous-time re-
current networks, and we will introduce concepts here that will be further
explored in Chapters 20 and 21. This Grossberg network is also the foun-
dation for the adaptive resonance theory (ART) networks that we will
present in Chapter 19.
We will begin with a discussion of the biological motivation for the Gross-
berg network: the human visual system. Although we will not cover this
material in any depth, the Grossberg networks are so heavily influenced by
biology that it is difficult to discuss his networks without putting them in
their biological context. It is also important to note that biology provided
the original inspirations for the field of artificial neural networks, and we
should continue to look for inspiration there, as scientists continue to de-
velop new understanding of brain function.
18-1
Theory and Examples

During the late 1960s and the 1970s the number of researchers in the field
of neural networks dropped dramatically. There were, however, a number
of researchers who continued to work during this period, including Tuevo
Kohonen, James Anderson, Kunihiko Fukushima and Shun-ichi Amari,
among others. One of the most prolific was Stephen Grossberg.
Grossberg has been continuously active, and highly productive, in neural
network research since the early 1960s. His work is characterized by the
use of nonlinear mathematics to model specific functions of mind and
brain, and his volume of output has been consistent with the magnitude of
the task of understanding the brain. The topics of his papers have ranged
from such specific areas as how competitive networks can provide contrast
enhancement in vision, to such general subjects as a universal theory for
human memory.
In part because of the scale of his efforts, his work has a reputation for dif-
ficulty. Each new paper is built on a foundation of 30 years of previous re-
sults, and is therefore difficult to assess on its own merits. In addition, his
terminology is self-consistent, but not in standard use by other researchers.
His work is also characterized by a high level of mathematical and neuro-
physiological sophistication. He is inspired by the interdisciplinary re-
search into brain function by Helmholtz, Maxwell and Mach, and he brings
this viewpoint to his work. His research lies at the intersection of mathe-
matics, psychology and neurophysiology. A lack of background in these ar-
eas can make his work difficult to approach on a first reading.
This chapter will take a rudimentary look at one of the seminal networks
developed by Grossberg. In order to obtain the maximum understanding of
his ideas, we will begin with a brief introduction to the biological motiva-
tion for the network: the visual system. Then we will present the mathe-
matical building block for many of Grossbergs networks: the shunting
model. After understanding the function of this simple model, we will dem-
onstrate how it can be used to build a neural network for adaptive pattern
recognition. This network will then form the basis for the adaptive reso-
nance theory networks that are discussed in Chapter 19. By building up
gradually to the more complex networks, we hope to make them more eas-
ily understandable.
There is an important lesson we should take from the work described in
this chapter. Although the original inspiration for the field of artificial neu-
ral networks came from biology, at times we forget to look back to biology
for new ideas. It will be the blending of biology, mathematics, psychology
and other disciplines that will provide the maximum growth in our under-
standing of neural networks.
18-2
Biological Motivation: Vision

The neural network described in this chapter was inspired by the develop-
mental physiology of the human visual system. In this section we want to
provide a brief introduction to vision, so that the function of the network
will be more understandable.
In Figure 18.1 we have a schematic representation of the first stages of the
visual system. Light passes through the cornea (the transparent front part
of the eye) and the lens, which bends the light to focus objects on the retina
(the interior layer of the external wall of the eye). It is after the light falls
on the retina that the immense job of translating this information into an
understandable image begins. As we will see later in this chapter, much of
what we see is not actually present in the image projected onto the retina.
Bipolar Cell
Ganglion Cell
Amacrine
Cell Optic
Nerve
Cone
Light Lens
Rod
Horizontal
Cell
Retina
Optic Nerve Fiber
Figure 18.1 Eyeball and Retina

Retina The retina is actually a part of the brain. It becomes separated from the
brain during fetal development, but remains connected to it through the
optic nerve. The retina consists of three layers of nerve cells. The outer lay-
er consists of the photoreceptors (rods and cones), which convert light into
Rods electrical signals. The rods allow us to see in dim light, whereas the cones
Cones allow us to see fine detail and color. For reasons not completely understood,
light must pass through the other two layers of the retina in order to stim-
ulate the rods and cones. As we will see later, this obstruction must be com-
pensated for in neural processing, in order to reconstruct recognizable
images.
The middle layer of the retina consists of three types of cells: bipolar cells,
Bipolar Cells horizontal cells and amacrine cells. Bipolar cells receive input from the re-
ceptors and feed into the third layer of the retina, containing the ganglion
18-3
Horizontal Cells cells. Horizontal cells link the receptors and the bipolar cells, and amacrine
Amacrine Cells cells link bipolar cells with the ganglion cells.
Ganglion Cells The final layer of the retina contains the ganglion cells. The axons of the
ganglion cells pass across the surface of the retina and collect in a bundle
to form the optic nerve. It is interesting to note that each eye contains
roughly 125 million receptors, but only 1 million ganglion cells. Clearly
there is significant processing done in the retina to perform data reduction.
The axons of the ganglion cells, bundled into the optic nerve, connect to an
area of the brain called the lateral geniculate nucleus, as illustrated in Fig-
ure 18.2. From this point the fibers fan out into the primary visual cortex,
located at the back of the brain. The axons of the ganglion cells make syn-
apses with lateral geniculate cells, and the axons of the lateral geniculate
Visual Cortex cells make synapses with cells in the visual cortex. The visual cortex is the
region of the brain devoted to visual function and consists of many layers
of cells.
Lateral
Primary Geniculate
Visual Nucleus
Cortex
Retina
Figure 18.2 Visual Pathway

The connections along the visual pathway are far from random. The map-
ping from each layer to the next is highly organized. The axons from the
ganglion cells in a certain part of the retina go to cells in a particular part
of the lateral geniculate, which in turn go to a particular part of the visual
cortex. (This topographic mapping was one of the inspirations for the self-
organizing feature map described in Chapter 14.) In addition, as we can see
in Figure 18.2, each hemisphere of the brain receives input from both eyes,
since half of the optic nerve fibers cross and the other half stay uncrossed.
It turns out that the left half of each visual field ends up in the right half
of the brain, and the right half of each visual field ends up in the left half
of the brain.
Illusions
We now have some idea of the general structure of the visual pathway, but
how does it function? What is the purpose of the three layers of the retina?
What operations are performed by the lateral geniculate? Some hints to the
18-4
answers to these questions can be obtained by investigating visual illu-

sions.
Why are there so many visual illusions? Mechanisms that overcome imper-
fections of the retinal uptake process imply illusions. Grossberg and others
have used the vast store of known illusions to probe adaptive perceptual
mechanisms [GrMi89]. If we can develop mathematical models that pro-
duce the same illusions the biological system does, then we may have a
mechanism that describes how this part of the brain works. To help us un-
derstand why illusions exist, we will first consider some of the imperfec-
tions of the retinal uptake process.
Figure 18.3 is the view of the retina that an ophthalmologist has when
Optic Disk looking into the eye through the cornea. The large pale circle is the optic
disk, where the optic nerve leaves the retina on its way to the lateral gen-
iculate. This is also where arteries enter and veins leave the retina. The op-
tic disk causes a blind spot in our vision, as we will discuss in a moment.
Fovea The darker disk to the right of the optic disk is the fovea, which constitutes
the center of our field of vision. This is a region of the retina, about half a
millimeter in diameter, that contains only cones. Although cones are dis-
tributed throughout the retina, they are most densely packed in the fovea.
In addition, in this area of the retina the other layers are displaced to the
side, so that the cones lie at the front. The densely packed photoreceptors,
and the lack of obstruction, give us our best fine-detail vision at the fovea,
which allows us to precisely focus the lens.
Optic Disk (Blind Spot)
Vein
Fovea
Figure 18.3 Back of the Eye (from [John01])

From Figure 18.3 we can see that there are a number of imperfections in
retinal uptake. First, there are no rods and cones in the optic disk, which
leaves a blind spot in our field of vision. We are not normally aware of the
blind spot because of processing done in the visual pathway, but we can
identify it with a simple test. Look at the blue circle on the left side of Fig-
ure 18.4, while covering your left eye. As you move your head closer to the
page, then farther away, you will notice a point (about nine inches away)
18-5
at which the circle on the right will disappear from your field of vision. (You
are still looking at the circle on the left.) If you havent tried this before, it
can be a little disconcerting. The interesting thing is that we dont see our
blind spot as a black hole. Somehow our brains fill in the missing region.
Figure 18.4 Test for the Blind Spot

Other imperfections in the retinal uptake are the arteries and veins that
cross in front of the photoreceptors at the back of the retina. These obstruct
the rods and cones from receiving all of the light in the visual field. In ad-
dition, because the photoreceptors are at the back of the retina, light must
pass through the other two layers to reach them.
Figure 18.5 illustrates the effect of these imperfections. Here we see an
edge displayed on the retina. The drawing on the right illustrates the im-
age initially perceived by the photoreceptors. The regions covered by the
blind spot and the veins are not observed by the rods and cones. (The rea-
son we do not see the arteries, veins, etc., is that the vision pathway does
not respond to stabilized images. The eyeballs are constantly jerking, in
what are called saccadic movements, so that even fixed objects in our field
of vision are moving relative to the eye. The veins are fixed relative to the
eyeball, so they fade from our vision.)
Edge
Vein
Blind Stabilized
Spot Images Fade
Retina
Figure 18.5 Perception of an Edge on the Retina (after [Gros90])

Because we do not see edges as displayed on the right side of Figure 18.5,
the neural systems in our visual pathway must be performing some opera-
tion that compensates for the distortions and completes the image. Gross-
berg suggests [GrMi89] that there are two primary types of compensatory
Emergent Segmentation processing involved. The first, which he calls emergent segmentation, com-
Featural Filling-in pletes missing boundaries. The second, which he calls featural filling-in,
fills in the color and brightness inside the resulting boundaries. These two
processes are illustrated in Figure 18.6. In the top figure we see an edge as
it is originally perceived by the rods and cones, with missing sections. In
18-6
the lower figure we see the completed edge, after the emergent segmenta-
tion and featural filling-in.
Before Processing
After Processing
Emergent Segmentation Featural Filling-in

Figure 18.6 Compensatory Processing (after [Gros90])
If the processing along the visual pathway is recreating missing parts of
the images we see, there must be times when it makes mistakes, since it
cannot know exactly those parts of a scene from which it receives no light.
These mistakes are illustrated by visual illusions. Consider, for example,
the two figures in the left margin. In the top figure you should be able to
see a bright white triangle lying on top of several other black objects. In
fact, no such triangle exists in the figure. It is purely a creation of the emer-
gent segmentation and featural filling-in process of your visual system.
The same is true of the bright white circle which appears to lie on top of the
lines in the lower-left figure.
The featural filling-in process is also demonstrated in Figure 18.7. This il-
lusion is called neon color spreading [vanT75]. In the diagram on the right
you may be able to see light blue diamonds, or even wide light blue lines
criss-crossing the figure. In the diagram on the left you may be able to see
a light blue ring. The blue you see filling in the diamonds and the ring is
not a result of the color having been smeared during the printing process,
nor is it caused by the scattering of light. This effect does not appear on the
retina at all. It does not exist, except in your brain. (The perception of neon
color spreading can vary from individual to individual, and the strength of
the perception is dependent on the colors used. If you do not notice the ef-
fect in Figure 18.7, look at the cover of any issue of the journal Neural Net-
works, Pergamon Press.)
Later in this chapter we will discuss some neural network models that can
help to explain the processes that implement emergent segmentation, as
well as other visual phenomena.
18-7
Figure 18.7 Neon Color Spreading (Featural Filling In)
Vision Normalization
In addition to emergent segmentation and featural filling-in, there are two
other phenomena that give us an indication of what operations are being
Brightness Constancy performed in the early vision system: brightness constancy and brightness
Brightness Contrast contrast. The brightness constancy effect is evidenced by the test illustrat-
ed in Figure 18.8. In this test a subject is shown a small grey disk inside a
darker grey annulus, which is illuminated by white light of a certain inten-
sity. The subject is asked to indicate the brightness of the central disk by
looking at a series of grey disks, separately illuminated, and selecting the
disk with the same brightness. Next, the brightness of the light illuminat-
ing the grey disk and dark annulus is increased, and the subject is again
asked to select the disk with the same brightness. This process is repeated
for several different levels of illumination. It turns out that in each case the
subject will choose the same disk as matching the original central disk.
Even though the total light entering the subjects eye is 10 to 100 times
brighter, it is only the relative brightness that registers.
Variable Separate
Illumination Constant Illumination
Figure 18.8 Test of Brightness Constancy (after [Gros90])
18-8
Basic Nonlinear Model
Another phenomenon of the vision system, which is closely related to

brightness constancy, is brightness contrast. This effect is illustrated by
the two figures in the left margin. At the centers of the two figures we have
two small disks with equivalent grey scale. The small disk in the top figure
is surrounded by a darker annulus, while the small disk in the lower figure
is surrounded by a lighter annulus. Even though both disks have the same
grey scale, the one inside the darker annulus appears brighter. This is be-
cause our vision system is sensitive to relative intensities. It would seem
that the total activity across the image is held constant.
The properties of brightness constancy and brightness contrast are very
important to our vision system. Since we see things in so many different
lighting conditions, if we were not able to compensate for the absolute in-
tensity of a scene, we would never learn to recognize things. Grossberg calls
this process of normalization discounting the illuminant.
In the rest of this chapter we want to present a neural network architecture
that is consistent with the physical phenomena discussed in this section.

Before we introduce the Grossberg network, we will begin by looking at
some of the building blocks that make up the network. The first building
Leaky Integrator block is the leaky integrator, which is shown in Figure 18.9. The basic
equation for this system is
dn(t)
------------ = n(t) + p(t) , (18.1)
dt
Time Constant where is the system time constant.
Leaky Integrator
.
+ n n
p 1/
-
dn/dt = - n + p
Figure 18.9 Leaky Integrator

The response of the leaky integrator to an arbitrary input p t is
t 1 t t
n(t) = e n(0) + --- e p() d . (18.2)
0
18-9
For example, if the input p t is constant and the initial condition n 0 is

zero, Eq. (18.2) will produce
t
n(t) = p 1 e . (18.3)
A graph of this response, for p = 1 and = 1 , is given in Figure 18.10. The

response exponentially approaches a steady state value of 1.
1
0.75
n 0.5
0.25
0
0 1 2 3 4 5
Figure 18.10 Leaky Integrator Response

There are two important properties of the leaky integrator that we want to
note. First, because Eq. (18.1) is linear, if the input p is scaled, then the
response n t will be scaled by the same amount. For example, if the input
is doubled, then the response will also be doubled, but will maintain the
same shape. This is evident in Eq. (18.3). Second, the speed of response of
the leaky integrator is determined by the time constant . When decreas-
es, the response becomes faster; when increases, the response becomes
slower. (See Problem P18.1.)
To experiment with the leaky integrator, use the Neural Network Design
Demonstration Leaky Integrator (nnd15li).
The leaky integrator forms the nucleus of one of Grossbergs fundamental
Shunting Model neural models: the shunting model, which is shown in Figure 18.11. The
equation of operation of this network is
dn(t) + + - -
------------ = n(t) + b n(t) p n(t) + b p , (18.4)
dt
+
Excitatory where p is a nonnegative value representing the excitatory input to the
-
network (the input that causes the response to increase), and p is a non-
Inhibitory negative value representing the inhibitory input (the input that causes the
+ -
response to decrease). The biases b and b are nonnegative constants that
determine the upper and lower limits on the neuron response, as we will
explain next.
18-10
Input Basic Shunting Model
b+ +
p+
-
.
+ + n n
1/
- -
+
p-
+
b-
dn/dt = -n + (b+ - n) p+ - (n + b-) p-
Figure 18.11 Shunting Model

There are three terms on the right-hand side of Eq. (18.4). When the net
sign of these three terms is positive, n(t) will go up. When the net sign is
negative, n(t) will go down. To understand the performance of the network,
lets investigate the three terms individually.
The first term, n(t) , is a linear decay term, which is also found in the leaky
integrator. It is negative whenever n(t) is positive, and positive whenever
+ +
n(t) is negative. The second term, b n(t) p , provides nonlinear gain
+
control. This term will be positive while n(t) is less than b , but will become
+ +
zero when n(t) = b . This effectively sets an upper limit on n(t) of b . The
- -
third term, n(t) + b p , also provides nonlinear gain control. It sets a low-
-
er limit on n(t) of b .
Figure 18.12 illustrates the performance of the shunting network when
+ -
b = 1 , b = 0 and = 1 . In the left graph we see the network response
+ -
when the excitatory input p = 1 and the inhibitory input p = 0 . For the
+ -
right graph p = 5 and p = 0 . Notice that even though the excitatory in-
put is increased by a factor of 5, the steady state network response is in-
creased by less than a factor of 2. If we were to continue to increase the
b+ excitatory input, we would find that the steady state network response
+
would increase, but would always be less than b = 1 .
If we apply an inhibitory input to the shunting network, the steady state
0 n(t)
-
network response will decrease, but will remain greater than b . To sum-
+
marize the operation of the shunting model, if n(0) falls between b and
-b- -
b , then n(t) will remain between these limits, as shown in the figure in
the left margin.
18-11
1 1
0.75 + 0.75 +
p = 1 p = 5
nt 0.5 0.5
0.25 0.25
0 0
0 1 2 3 4 5 0 1 2 3 4 5
t t
Figure 18.12 Shunting Network Response

The shunting model will form the basis for the Grossberg competitive net-
work, which we will discuss in the next section. The nonlinear gain control
will be used to normalize input patterns and to maintain relative intensi-
ties over a wide range of total intensity.
To experiment with the shunting network, use the Neural Network Design
Demonstration Shunting Network (nnd15sn).
Two-Layer Competitive Network

We are now ready to present the Grossberg competitive network. This net-
work was inspired by the operation of the mammalian visual system, which
we discussed in the opening section of this chapter. (Grossberg was influ-
enced by the work of Chistoph von der Malsburg [vond73], which was influ-
enced in turn by the Nobel-prize-winning experimental work of David
Hubel and Torsten Wiesel [HuWi62].) A block diagram of the network is
There are three components to the Grossberg network: Layer 1, Layer 2
and the adaptive weights. Layer 1 is a rough model of the operation of the
retina, while Layer 2 represents the visual cortex. These models do not ful-
ly explain the complexity of the human visual system, but they do illustrate
Short-Term Memory a number of its characteristics. The network includes short-term memory
Long-Term Memory (STM) and long-term memory (LTM) mechanisms, and performs adapta-
tion, filtering, normalization and contrast enhancement. In the following
subsections we will discuss the operation of each of the components of the
network.
As we analyze the various elements of the Grossberg network, you will no-
tice the similarity to the Kohonen competitive network of the previous
chapter.
18-12
Layer 1 Layer 2
(Retina) (Visual Cortex)
Input
STM
LTM
(Adaptive Weights)
Normalization Contrast
Enhancement
Figure 18.13 Grossberg Competitive Network
Layer 1
Layer 1 of the Grossberg network receives external inputs and normalizes
the intensity of the input pattern. (Recall from Chapter 14 that the Ko-
honen network performs best when the input patterns are normalized. For
the Grossberg network the normalization is accomplished by the first layer
of the network.) A block diagram of this layer is given in Figure 18.14. Note
that it uses the shunting model, with the excitatory and inhibitory inputs
computed from the input vector p .
Input Layer 1
+b1
+
+W1
-
S1 x S1 .
p + + n1 n1 a1
1/
S1 x 1
- -
+
-W1
S1 S1
S1 x S1
+
-b1
dn1/dt = - n1 + (+b1 - n1) [+W1] p - (n1 + -b1) [-W1] p
Figure 18.14 Layer 1 of the Grossberg Network

The equation of operation of Layer 1 is
1
dn (t) 1 + 1 1 + 1 1 - 1 - 1
-------------- = n (t) + b n (t) W p n (t) + b W p . (18.5)
dt
18-13
As we mentioned earlier, the parameter determines the speed of re-

sponse. It is chosen so that the neuron responses will be much faster than
the changes in the adaptive weights, which we will discuss in a later sec-
tion.
+ 1
Eq. (18.5) is a shunting model with excitatory input W p , where
1 0 0
0
W = 0 1
+ 1
. (18.6)

0 0 1
Therefore the excitatory input to neuron i is the ith element of the input
vector.
- 1
The inhibitory input to Layer 1 is W p , where
0 1 1
1
W = 1 0
- 1
. (18.7)

1 1 0
Thus the inhibitory input to neuron i is the sum of all elements of the input
vector, except the ith element.
+ 1 - 1
The connection pattern defined by the matrices W and W is called an
On-Center/Off-Surround on-center/off-surround pattern. This is because the excitatory input for
neuron i (which turns the neuron on) comes from the element of the input
vector centered at the same location (element i ), while the inhibitory input
(which turns the neuron off) comes from surrounding locations. This type
of connection pattern produces a normalization of the input pattern, as we
will show in the following discussion.
- 1
For simplicity, we will set the inhibitory bias b to zero, which sets the
lower limit of the shunting model to zero, and we will set all elements of the
+ 1
excitatory bias b to the same value, i.e.,
+ 1 + 1 1
bi = b , i = 1 2 S , (18.8)
so that the upper limit for all neurons will be the same.
To investigate the normalization effect of Layer 1, consider the response of
neuron i :
18-14
1
dn i (t)
- = n i (t) + b n i (t) p i n i (t) p j .
1 + 1 1 1
------------- (18.9)
dt
ji
1
In the steady state ( dn i (t) dt = 0 ) we have
0 = n i + b n i p i n i p j .
1 + 1 1 1
(18.10)
ji
1
If we solve for the steady state neuron output n i we find
+ 1
1 b pi
n i = ----------------------
1 . (18.11)
S
1+ pj
j=1
We now define the relative intensity of input i to be

1
S
p
p i = ----i where P =
P pj . (18.12)
j=1
Then the steady state neuron activity can be written

+ 1
n i = ------------- p i .
1 b P
(18.13)
1 + P
1
Therefore n i will be proportional to the relative intensity p i , regardless of
the magnitude of the total input P . In addition, the total neuron activity is
bounded:
1 1
S S
+ 1 + 1
------------
b P
- p = ------------- b .
b P

1 + 1
nj = (18.14)
1 + P j 1 + P
j=1 j=1
+ 1
The input vector is normalized so that the total activity is less than b ,
while the relative intensities of the individual elements of the input vector
1
are maintained. Therefore, the outputs of Layer 1, n i , code the relative in-
put intensities, p i , rather than the instantaneous fluctuations in the total
input activity, P . This result is produced by the on-center/off-surround con-
nection pattern of the inputs and the nonlinear gain control of the shunting
model.
Note that Layer 1 of the Grossberg network explains the brightness con-
stancy and brightness contrast characteristics of the human visual system,
which we discussed on page 15-8. The network is sensitive to the relative
18-15
intensities of an image, rather than absolute intensities. In addition, exper-

imental evidence has shown that the on-center/off-surround connection
- - pattern is a characteristic feature of the receptive fields of retinal ganglion
cells [Hube88]. (The receptive field is an area of the retina in which the pho-
+ toreceptors feed into a given cell. The figure in the left margin illustrates
- - the on-center/off-surround receptive field of a typical retinal ganglion cell.
A + indicates an excitatory region, and a - indicates an inhibitory re-
gion. It is a two-dimensional pattern, as opposed to the one-dimensional
connections of Eq. (18.6) and Eq. (18.7).)
2 To illustrate the performance of Layer 1, consider the case of two neurons,
+2 + 1
with b = 1 , = 0.1 :
1
dn 1(t) 1 1 1
0.1 -------------- = n 1(t) + 1 n 1(t) p 1 n 1(t)p 2 , (18.15)
dt
1
dn 2(t) 1 1 1
0.1 -------------- = n 2(t) + 1 n 2(t) p 2 n 2(t)p 1 . (18.16)
dt
The response of this network, for two different input vectors, is shown in
Figure 18.15. For both input vectors, the second element is four times as
large as the first element, although the total intensity of the second input
vector is five times as large as that of the first input vector (50 vs. 10). From
Figure 18.15 we can see that the response of the network maintains the rel-
ative intensities of the inputs, while limiting the total response. The total
1 1
response ( n 1(t) + n 2(t) ) will always be less than 1.
1 1
1
1 n2
0.75
n2 0.75
0.5
p1 = 2 0.5 p 2 = 10
8 40
1 1
0.25 n1 0.25 n1
0 0
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
t t
Figure 18.15 Layer 1 Response

To experiment with Layer 1 of the Grossberg network, use the Neural Net-
work Design Demonstration Grossberg Layer 1 (nnd15gl1).
18-16
Layer 2
Layer 2 of the Grossberg network, which is a layer of continuous-time in-
stars, performs several functions. First, like Layer 1, it normalizes total ac-
tivity in the layer. Second, it contrast enhances its pattern, so that the
neuron that receives the largest input will dominate the response. (This is
closely related to the winner-take-all competition in the Hamming network
Short-Term Memory and the Kohonen network.) Finally, it operates as a short-term memory
(STM) by storing the contrast-enhanced pattern.
Figure 18.16 is a diagram of Layer 2. As with Layer 1, the shunting model
forms the basis for Layer 2. The main difference between Layer 2 and Lay-
er 1 is that Layer 2 uses feedback connections. The feedback enables the
network to store a pattern, even after the input has been removed. The
feedback also performs the competition that causes the contrast enhance-
ment of the pattern. We will demonstrate these properties in the following
discussion.
Layer 2
+ + On-Center
W2 +W2
+b2
S2 x S1 + S2 x S2
-
.
a1 + + n2 n2 a2
1/ f2
S1 - -
+
S2
+
-b2
-W2
Off-Surround
S2 x S2
dn2/dt = - n2 + (+b2 - n2) {[+W2] f 2(n2) + W2 a1}

- (n2 + -b2) [-W2] f 2(n2)
Figure 18.16 Layer 2 of the Grossberg Network
2
dn (t) 2 + 2 2 + 2 2 2 2 1
-------------- = n (t) + b n (t) W f (n (t)) + W a
dt
2 - 2 - 2 2 2
n (t) + b W f (n (t)) (18.17)
18-17
+ 2 2 2 2 1
This is a shunting model with excitatory input W f (n (t)) + W a ,
+ 2 + 1 2
where W = W provides on-center feedback connections, and W con-
sists of adaptive weights, analogous to the weights in the Kohonen net-
2
work. The rows of W , after training, will represent the prototype patterns.
- 2 2 2
The inhibitory input to the shunting model is W f (n (t)) , where
- 2 - 1
W = W provides off-surround feedback connections.
2 To illustrate the performance of Layer 2, consider a two-neuron layer with
+2
2 T
w
= 0.1 , b = 1 , b = 0 , W = 1 = 0.9 0.45 ,
+ 2 - 2 2
(18.18)
1 0 2 T 0.45 0.9
2w
and
2
2 10 n
f (n) = ------------------2- . (18.19)
1 + n
The equations of operation of the layer will be

2
dn 1(t) 2 2 2 2 2 T 1
0.1 -------------
- = n 1(t) + 1 n 1(t) f (n 1(t)) + 1w a (18.20)
dt
2 2 2
n 1(t)f (n 2(t))
2
dn 2(t) 2 2 2 2 2 T 1
0.1 -------------
- = n 2(t) + 1 n 2(t) f (n 2(t)) + 2w a (18.21)
dt
2 2 2
n 2(t)f (n 1(t)) .
Note the relationship between these equations and the Hamming and Ko-
honen networks. The inputs to Layer 2 are the inner products between the
2
prototype patterns (rows of the weight matrix W ) and the output of Layer
1 (normalized input pattern). The largest inner product will correspond to
the prototype pattern closest to the input pattern. Layer 2 then performs a
Contrast Enhance competition between the neurons, which tends to contrast enhance the out-
put pattern maintaining large outputs while attenuating small outputs.
This contrast enhancement is generally milder than the winner-take-all
competition of the Hamming and Kohonen networks. In the Hamming and
Kohonen networks, the competition drives all but one of the neuron out-
puts to zero. The exception is the one with the largest input. In the Gross-
berg network, the competition maintains large values and attenuates small
values, but does not necessarily drive all small values to zero. The amount
18-18
2
of contrast enhancement is determined by the transfer function f , as we
will see in the next section.
Figure 18.17 illustrates the response of Layer 2 when the input vector
1 T
a = 0.2 0.8 (the steady state result obtained from our Layer 1 example)
is applied for 0.25 seconds and then removed.
2 T 1
2w a 2
n 2(t)
0.75
2 T 1
1w a
0.5
0.25 2
n 1(t)
0
0 0.1 0.2 0.3 0.4 0.5
t
Figure 18.17 Layer 2 Response

There are two important characteristics of this response. First, even before
the input is removed, some contrast enhancement is performed. The inputs
to Layer 2 are
2 T 1
1w a = 0.9 0.45 0.2 = 0.54 , (18.22)
0.8
2 T 1
2w a = 0.45 0.9 0.2 = 0.81 . (18.23)
0.8
Therefore the second neuron has 1.5 times as much input as the first neu-
ron. However, after 0.25 seconds the output of the second neuron is 6.34
times the output of the first neuron. The contrast between high and low has
been increased dramatically.
The second characteristic of the response is that after the input has been
set to zero, the network further enhances the contrast and stores the pat-
tern. In Figure 18.17 we can see that after the input is removed (at 0.25
seconds) the output of the first neuron decays to zero, while the output of
the second neuron reaches a steady state value of 0.79 . This output is
maintained, even after the input is removed. (Grossberg calls this behavior
reverberation [Gross76].) It is the nonlinear feedback that enables the net-
18-19
work to store the pattern, and the on-center/off-surround connection pat-

+ 2 - 2
tern (determined by W and W ) that causes the contrast enhancement.
As an aside, note that we have used the on-center/off-surround structure in

both layers of the Grossberg network. There are other connection patterns
that could be used for different applications. Recall, for instance, the emer-
gent segmentation problem discussed earlier in this chapter. A structure
Oriented Receptive Field that has been proposed to implement this mechanism is the oriented recep-
tive field [GrMi89], which is shown in the left margin. For this structure,
the on (excitatory) connections come from one side of the field (indicated
by the blue region), and the off (inhibitory) connections come from the
other side of the field (indicated by the white region).
The operation of the oriented receptive field is illustrated in Figure 18.18.
When the field is aligned with an edge, the corresponding neuron is acti-
vated (large response). When the field is not aligned with an edge, then the
neuron is inactive (small response). This explains why we might perceive
an edge where none exists, as can be seen in the right-most receptive field
For a complete discussion of oriented receptive fields and how they can be
incorporated into a neural network architecture for preattentive vision, see
[GrMi89]. This paper also discusses a mechanism for featural filling-in.
Active
Active
Inactive
Figure 18.18 Operation of Oriented Receptive Field
Choice of Transfer Function

The behavior of Layer 2 of the Grossberg network depends very much on
2
n2i(0) the transfer function f (n) . For example, suppose that an input has been
applied for some length of time, so that the output has stabilized to the pat-
tern shown in the left margin. (Each point represents the output of an in-
i dividual neuron.) If the input is then removed, Figure 18.19 demonstrates
2
how the choice of f (n) will affect the steady state response of the network.
(See [Gross82].)
18-20
Stored Pattern
f 2(n) n2() Comments
Linear
Perfect storage
of any pattern,
but amplifies
noise.
Slower than
Linear Amplifies noise,
reduces contrast.
Faster than
Winner-take-all,
Linear
suppresses noise,
quantizes total
activity.
Sigmoid Supresses
noise, contrast
enhances, not
quantized.
2
Figure 18.19 Effect of Transfer Function f (n) (after [Gross82])
If the transfer function is linear, the pattern is perfectly stored. Unfortu-

nately, the noise in the pattern will be amplified and stored as easily as the
significant inputs. (See Problem P18.6.) If the transfer function is slower-
2 n
than-linear (e.g., f (n) = 1 e ), the steady state response is independent
of the initial conditions; all neurons that begin with nonzero values will
come to the same level in the steady state. All contrast is eliminated, and
noise is amplified.
2 2
Faster-than-linear transfer functions (e.g., f (n) = n ) produce a winner-
take-all competition. Only those neurons with the largest initial values are
stored; all others are driven to zero. This minimizes the effect of noise, but
quantizes the response into an all-or-nothing signal (as in the Hamming
and Kohonen networks).
A sigmoid function is faster-than-linear for small signals, approximately
linear for intermediate signals and slower-than-linear for large signals.
When a sigmoid transfer function is used in Layer 2, the pattern is contrast
enhanced; larger values are amplified, and smaller values are attenuated.
All initial neuron outputs that are less than a certain level (called the
quenching threshold by Grossberg [Gros76]) decay to zero. This merges the
noise suppression of the faster-than-linear transfer functions with the per-
fect storage produced by linear transfer functions.
18-21
To experiment with Layer 2 of the Grossberg network, use the Neural Net-
work Design Demonstration Grossberg Layer 2 (nnd15gl2).
Learning Law
The third component of the Grossberg network is the learning law for the
2
Long-Term Memory adaptive weights W . Grossberg calls these adaptive weights the long-term
2
memory (LTM). This is because the rows of W will represent patterns that
have been stored and that the network will be able to recognize. As in the
Kohonen and Hamming networks, the stored pattern that is closest to an
input pattern will produce the largest output in Layer 2. In the next sub-
section we will look more closely at the relationship between the Grossberg
network and the Kohonen network.
2
One learning law for W is given by
2
dw i j(t) 2 2 1
----------------- = w i j(t) + n i (t)n j (t) . (18.24)
dt
The first term in the bracket on the right-hand side of Eq. (18.24) is a pas-
sive decay term, which we have seen in the Layer 1 and Layer 2 equations,
while the second term implements a Hebbian-type learning. Together,
these terms implement the Hebb rule with decay, which was discussed in
Chapter 13.
Recall from Chapter 13 that it is often useful to turn off learning (and for-
2
getting) when n i (t) is not active. This can be accomplished by the following
learning law:
2
dw i j(t) 2 2 1
----------------- = n i (t) w i j(t) + n j (t) , (18.25)
dt
or, in vector form,

2
d iw (t) 2 2 1
--------------------- = n i (t) iw (t) + n (t) , (18.26)
dt
2 2
where iw (t) is a vector composed of the elements of the ith row of W (see
Eq. (4.4)).
The terms on the right-hand side of Eq. (18.25) are multiplied (gated) by
2 2
n i (t) , which allows learning (and forgetting) to occur only when n i (t) is not
zero. This is the continuous-time implementation of the instar learning
rule, which we introduced in Chapter 13 (Eq. (15.32)). In the following sub-
section we will demonstrate the equivalence of Eq. (18.25) and Eq. (15.32).
18-22
To illustrate the performance of the Grossberg learning law, consider a net-

2
+2 work with two neurons in each layer. The weight update equations would
be
2
dw 1 1(t) 2 2 1
------------------- = n 1(t) w 1 1(t) + n 1(t) , (18.27)
dt
2
dw 1 2(t) 2 2 1
------------------- = n 1(t) w 1 2(t) + n 2(t) , (18.28)
dt
2
dw 2 1(t) 2 2 1
------------------- = n 2(t) w 2 1(t) + n 1(t) , (18.29)
dt
2
dw 2 2(t) 2 2 1
------------------- = n 2(t) w 2 2(t) + n 2(t) , (18.30)
dt
where the learning rate coefficient has been set to 1 . To simplify our ex-
ample, we will assume that two different input patterns are alternately
presented to the network for periods of 0.2 seconds at a time. We will also
assume that Layer 1 and Layer 2 converge very quickly, in comparison
with the convergence of the weights, so that the neuron outputs are effec-
tively constant over the 0.2 seconds. The Layer 1 and Layer 2 outputs for
the two different input patterns will be
for pattern 1: n =
1 0.9 , n 2 = 1 , (18.31)
0.45 0
for pattern 2: n = 0.45 , n = 0 .

1 2
(18.32)
0.9 1
Pattern 1 is coded by the first neuron in Layer 2, and pattern 2 is coded by

the second neuron in Layer 2.
Figure 18.20 illustrates the response of the adaptive weights, beginning
with all weights set to zero. Note that the first row of the weight matrix
2 2 2
( w 1 1(t) and w 1 2(t) ) is only adjusted during those periods when n 1(t) is non-
1 1
zero, and that it converges to the corresponding n pattern ( n 1(t) = 0.9 and
1
n 2(t) = 0.45 ). (The elements in the first row of the weight matrix are indi-
cated by the blue lines in Figure 18.20.) Also, the second row of the weight
2 2 2
matrix ( w 2 1(t) and w 2 2(t) ) is only adjusted during those periods when n 2(t)
1 1
is nonzero, and it converges to the corresponding n pattern ( n 1(t) = 0.45
1
and n 2(t) = 0.9 ). (The elements in the second row of the weight matrix are
indicated by the black lines in Figure 18.20.)
18-23
2
0.75
w 1 1(t) 2
w 2 2(t)
2
0.5
w 1 2(t)
0.25 2
w 2 1(t)
0
0 0.5 1 1.5 2 2.5 3
Figure 18.20 Response of the Adaptive Weights

To experiment with the adaptive weights, use the Neural Network Design
Demonstration Adaptive Weights (nnd15aw).
Relation to Kohonen Law

In the previous section we indicated that the Grossberg learning law was a
continuous-time version of the instar learning law, which we discussed in
Chapter 13. Now we want to demonstrate this fact. We will also show that
the Grossberg network is, in its simplest form, a continuous-time version
of the Kohonen competitive network of Chapter 14.
To begin, lets repeat the Grossberg learning law of Eq. (18.25):
2
d iw (t) 2 2 1
--------------------- = n i (t) iw (t) + n (t) . (18.33)
dt
If we approximate the derivative by

2 2 2
d iw (t) iw (t + t) iw (t)
--------------------- -------------------------------------------- , (18.34)
dt t
then we can rewrite Eq. (18.33) as

2 2 2 2 1
iw (t + t) = iw (t) + t n i (t) iw (t) + n (t) . (18.35)
(Compare this equation with the instar rule that was presented in Chapter
13 in Eq. (15.33).) If we rearrange terms, this can be reduced to
2 2 2 2 1
iw (t + t) = 1 t n i (t) iw (t) + t n i (t) n (t) . (18.36)
To simplify the analysis further, assume that a faster-than-linear transfer

function is used in Layer 2, so that only one neuron in that layer will have
18-24
Relation to Kohonen Law
a nonzero output; call it neuron i . Then only row i of the weight matrix
will be updated:
2 2 1
i
w (t + t) = 1 ' iw (t) + ' n (t) , (18.37)
2
where ' = t n i(t) .
This is almost identical to the Kohonen rule for the competitive network
that we introduced in Chapter 14 in Eq. (16.13). The weight vector for the
1
winning neuron (with nonzero output) will be moved toward n , which is a
normalized version of the current input pattern.
There are three major differences between the Grossberg network that we
have presented in this chapter and the basic Kohonen competitive network.
First, the Grossberg network is a continuous-time network (satisfies a set
of nonlinear differential equations). Second, Layer 1 of the Grossberg net-
work automatically normalizes the input vectors. Third, Layer 2 of the
Grossberg network can perform a soft competition, rather than the win-
ner-take-all competition of the Kohonen network. This soft competition al-
lows more than one neuron in Layer 2 to learn. This causes the Grossberg
network to operate as a feature map.
18-25
Summary of Results

Leaky Integrator
dn(t)
------------ = n(t) + p(t)
dt
Leaky Integrator
.
+ n n
p 1/
-
dn/dt = - n + p
Shunting Model
dn(t) + + - -
------------ = n(t) + b n(t) p n(t) + b p
dt
b+ +
p+
- b+
.
+ + n n
1/
- - 0 n(t)
+ -b-
p-
+
b-
dn/dt = -n + (b+ - n) p+ - (n + b-) p-
18-26
Summary of Results

Layer 1 Layer 2
(Retina) (Visual Cortex)
Input
STM
LTM
(Adaptive Weights)
Normalization Contrast
Enhancement
Layer 1
Input Layer 1
+b1
+
+W1
-
S1 x S1 .
p + + n1 n1 a1
1/
S1 x 1
- -
+
S1
-W1
+
S1 18
S1 x S1
-b1
dn1/dt = - n1 + (+b1 - n1) [+W1] p - (n1 + -b1) [-W1] p
1
dn (t) 1 + 1 1 + 1 1 - 1 - 1
-------------- = n (t) + b n (t) W p n (t) + b W p
dt
1 0 0 0 1 1
0 1
W = 0 1 W = 1 0
+ 1 - 1

0 0 1 1 1 0
On-Center Off-Surround
18-27
Steady State Neuron Activity

1
S
+ 1
p
n i = ------------- p i , where p i = ----i and P =
b P
pj
1
1 + P P
j=1
Layer 2
Layer 2
+ + On-Center
W2 +W2
+b2
S2 x S1 + S2 x S2
-
.
a1 + + n2 n2 a2
1/ f2
S1 - -
+
S2
+
-b2
-W2
Off-Surround
S2 x S2
dn2/dt = - n2 + (+b2 - n2) {[+W2] f 2(n2) + W2 a1}

- (n2 + -b2) [-W2] f 2(n2)
2
dn (t) 2 + 2 2 + 2 2 2 2 1
-------------- = n (t) + b n (t) W f (n (t)) + W a
dt
2 - 2 - 2 2 2
n (t) + b W f (n (t))
18-28
Summary of Results
Choice of Transfer Function

Stored Pattern
f 2(n) n2() Comments
Linear
Perfect storage
of any pattern,
but amplifies
noise.
Slower than
Linear
Amplifies noise,
reduces contrast.
Faster than
Winner-take-all,
Linear
suppresses noise,
quantizes total
activity.
Sigmoid Supresses
noise, contrast
enhances, not
quantized.
Learning Law
2
d iw (t) 2 2 1
--------------------- = n i (t) iw (t) + n (t)
dt 18
(Continuous-Time Instar Learning)
18-29
Solved Problems
P18.1 Demonstrate the effect of the coefficient on the performance of
the leaky integrator, which is shown in Figure P18.1, with the in-
put p = 1 .
Leaky Integrator
.
+ n n
p 1/
-
dn/dt = - n + p
Figure P18.1 Leaky Integrator

The equation of operation for the leaky integrator is
dn(t)
------------ = n(t) + p(t) .
dt
The solution to this differential equation, for an arbitrary input p t , is
t 1 t t
n(t) = e n(0) + --- e p(t ) d .
0
If p t = 1 , the solution will be
t 1 t t
n(t) = e n(0) + --- e d .
0
We want to show how this response changes as a function of . The re-

sponse will be
t t t
n(t) = e n( 0 ) + 1 e = e n(0) 1 + 1 .
This response begins at n(0) , and then grows exponentially (or decays ex-
ponentially, depending on whether or not n(0) is greater than or less than
1 ), approaching the steady state response of n() = 1 . As is decreased,
t
the response becomes faster (since e decays more quickly), while the
steady state value remains constant. Figure P18.2 illustrates the responses
for = 1 0.5 0.25 0.125 , with n(0) = 0 . Notice that the steady state value
remains 1 for each case. Only the speed of response changes.
18-30
Solved Problems
1.25
1
= 0.125
0.75
0.5
= 1
0.25
0
0 0.5 1 1.5 2
t
Figure P18.2 Effect of on Leaky Integrator Response
P18.2 Again using the leaky integrator of Figure P18.1, set = 1 .
i. Find a difference equation approximation to the leaky inte-

grator differential equation by approximating the deriva-
tive using
dn(t) n(t + t) n(t)

------------ ----------------------------------- .
dt t
ii. Using t = 0.1 , compare the response of this difference

equation with the response of the differential equation for
p t = 1 and n(0) = 0 . Compare the two over the range
0t1.
iii. Using the difference equation model for the leaky integra-
tor, show that the response is a weighted average of previ-
ous inputs.
i. If we make the approximation to the derivative, we find
n(t + t) n(t)
----------------------------------- = n(t) + p(t)
t
or
n(t + t) = n(t) + t n(t) + p(t) = 1 t n(t) + t p(t) .
ii. If we let t = 0.1 we obtain the difference equation
n(t + 0.1) = 0.9n(t) + 0.1p(t) .
If we let p t = 1 and n(0) = 0 , then we can solve for n t :
18-31
n(0.1) = 0.9n(0) + 0.1p(0) = 0.1
n(0.2) = 0.9n(0.1) + 0.1p(0.1) = 0.9 0.1 + 0.1 1 = 0.19 ,
n(0.3) = 0.9n(0.2) + 0.1p(0.2) = 0.9 0.19 + 0.1 1 = 0.271 ,
n(0.4) = 0.9n(0.3) + 0.1p(0.3) = 0.9 0.271 + 0.1 1 = 0.3439 ,
n(0.5) = 0.9n(0.4) + 0.1p(0.4) = 0.9 0.3439 + 0.1 1 = 0.4095 ,
n(0.6) = 0.4686 , n(0.7) = 0.5217 , n(0.8) = 0.5695 ,
n(0.9) = 0.6126 , n(1.0) = 0.6513 .
From Problem P18.1, the solution to the differential equation is

t t t
n(t) = e n(0) + 1 e = 1 e .
Figure P18.3 illustrates the relationship between the difference equation

solution and the differential equation solution. The black line represents
the differential equation solution, and the blue circles represent the differ-
ence equation solution. The two solutions are very close, and can be made
arbitrarily close by decreasing the interval t .
1
0.75
0.5
0.25
0
0 0.25 0.5 0.75 1
Figure P18.3 Comparison of Difference and Differential Equations

iii. Consider again the difference equation model of the leaky integrator,
which we developed in part (ii):
n(t + 0.1) = 0.9n(t) + 0.1p(t) .
If we start from a zero initial condition we find
n(0.1) = 0.9n(0) + 0.1p(0) = 0.1p(0) ,
n(0.2) = 0.9n(0.1) + 0.1p(0.1) = 0.9 0.1p(0) + 0.1p(0.1) = 0.09p(0) + 0.1p(0.1)
18-32
Solved Problems
n 0.3 = 0.9n(0.2) + 0.1p(0.2) = 0.081p(0) + 0.09p(0.1) + 0.1p(0.2)

p(0.1) + + p( k 1 0.1) .
k1 k2
n k0.1 = 0.1 0.9 p(0) + 0.9
Therefore the response of the leaky integrator is a weighted average of pre-

vious inputs, p(0) p(0.1) p( k 1 0.1) . Note that the recent inputs con-
tribute more to the response than the early inputs.
P18.3 Find the response of the shunting network shown in Figure P18.4
+ - + -
for = 1 , b = 1 , b = 1 , p = 0 , p = 10 and n(0) = 0.5 .
b+ +
p+
-
.
+ + n n
1/
- -
+
p-
+
b-
dn/dt = -n + (b+ - n) p+ - (n + b-) p-
Figure P18.4 Shunting Network

The equation of operation of the shunting network is
dn(t) + + - -
------------ = n(t) + b n(t) p n(t) + b p .
dt
For the given parameter values this becomes
dn(t)
------------ = n(t) n(t) + 1 10 = 11n(t) 10 .
dt
The solution to this equation is

t 11 t
n(0) + e
11t
n(t) = e 10 d ,
0
18-33
or
0.5 + ------ 1 e .
11t 10 11t
n(t) = e
11
The response is plotted in Figure P18.5.

1
0.5
-0.5
-1
0 0.25 0.5 0.75 1
Figure P18.5 Shunting Network Response

There are two things to note about this response. First, as with all shunting
-
networks, the response will never drop below b , which in this case is 1 .
-
As the inhibitory input p is increased, the steady state response will de-
-
crease, but it can never be less than b . The second characteristic of the
response is that the speed of the response will increase as the input is in-
- -
creased. For instance, if the input were changed from p = 10 to p = 100 ,
the response would be
0.5 + --------- 1 e
101t 100 101t
n(t) = e .
101
101t 11t
Since e decays more rapidly than e , the response will be faster.
P18.4 Find the response of Layer 1 of the Grossberg network for the case
+ 1 - 1
of two neurons, with b = 1 , b = 0 , = 1 and input vector
T
p = c 2c . Assume that the initial conditions are set to zero. Dem-
onstrate the effect of c on the response.
The Layer 1 differential equations for this case are

1
dn 1(t) 1 1 1 1
-------------- = n 1(t) + 1 n 1(t) c n 1(t) 2c = 1 + 3c n 1(t) + c ,
dt
18-34
Solved Problems
1
dn 2(t) 1 1 1 1
-------------- = n 2(t) + 1 n 2(t) 2c n 2(t) c = 1 + 3c n 2(t) + 2c .
dt
The solutions to these equations would be
1 + 3c t 1 t 1 + 3c t
+ e
1
n 1(t) = e n 1(0) c d ,
0
1 + 3c t 1 t 1 + 3c t
+ e
1
n 2(t) = e n 2( 0) 2c d .
0
If the initial conditions are set to zero, these equations reduce to
n 1(t) = --------------- 1 e
1 c 1 + 3c t
,
1 + 3c
n 2(t) = --------------- 1 e
1 2c 1 + 3c t
.
1 + 3c
Note that the outputs of Layer 1 retain the same relative intensities as the
inputs; the output of neuron 2 is always twice the output of neuron 1. This
behavior is consistent with Eq. (18.13). In addition, the total output inten-
1 1 + 1
sity ( n 1(t) + n 2(t) ) is never larger than b = 1 , as predicted in Eq. (18.14).
As c is increased, it has two effects on the response. First, the steady state
values increase slightly. Second, the response becomes faster, since
1 + 3c t
e decays more rapidly as c increases.
P18.5 Consider Layer 2 of the Grossberg network. Assume that the input
to Layer 2 is applied for some length of time and then removed (set
to zero).
i. Find a differential equation that describes the variation in
the total output of Layer 2,
2
S
nk (t) ,
2 2
N (t) =
k=1
after the input to Layer 2 has been removed.

ii. Find a differential equation that describes the variation in
the relative outputs of Layer 2,
2
2 n i (t)
n i (t) -,
= -----------
2
N (t)
18-35
after the input to Layer 2 has been removed.

i. The operation of Layer 2 is described by Eq. (18.17):
2
dn (t) 2 + 2 2 + 2 2 2 2 1
-------------- = n (t) + b n (t) W f (n (t)) + W a
dt
2 - 2 - 2 2 2
n (t) + b W f (n (t)) .
2 1
If the input is removed, then W a is zero. For simplicity, we will set the
- 2
inhibitory bias b to zero, and we will set all elements of the excitatory bias
+ 2 + 2
b to b . The response of neuron i is then given by
2
dn i (t) 2
- = n i (t) + b n i (t) f (n i (t)) n i (t) f (n k (t)) .
2 + 2 2 2 2 2 2
-------------
dt
ki
This can be rearranged to produce
2 S2
dn i (t) 2
-------------- = n i (t) + b f (n i (t)) n i (t) f (n k (t)) .
2 + 2 2 2 2 2
dt k = 1

If we then make the definition

2
S
f (n k(t)) ,
2 2 2
F (t) =
k=1
we can simplify the equation to

2
dn i (t) 2 2 + 2 2 2
-------------
- = 1 + F (t) n i (t) + b f (n i (t)) .
dt
To get the total activity, sum this equation over i to produce

2
dN (t) 2 2 + 2 2
--------------- = 1 + F (t) N (t) + b F (t) .
dt
This equation describes the variation in the total activity in Layer 2 over
time.
ii. The derivative of the relative activity is
18-36
Solved Problems
2 2
d 2 d n i (t) 1 d 2 n i (t) d 2
n (t) = - n (t) ------------------
------------ = ----------- - N (t) .
dt i d t N 2(t) 2 dt i
N (t) N (t) d t
2 2
If we then substitute our previous equations for these derivatives, we find
d 2 1 2 2 + 2 2 2
n i (t) = -----------
2
- 1 + F (t) n i (t) + b f (n i (t))
dt N (t)
2
n i (t) 2 2 + 2 2
- 1 + F (t) N (t) + b F (t) .
-----------
2
N (t)
Two terms on the right-hand side will cancel to produce
2
d 2 1 + 2 2 2 n i (t) + 2 2
n i (t) = -----------
2
- b f (n i (t)) -----------
2
- b F (t) ,
dt N (t) N (t)
or
+ 2 2 2 2 2
d 2 b F (t) f (n i (t)) n i (t)
n i (t) = ------------------
2
- ------------------
2
- .
-----------
2
dt N (t) F (t) N (t)
We can put this in a more useful form if we expand the terms in the brack-
ets:
2 2
f (n i (t)) n i (t)
------------------
2
2
1
- = -----------------------
-----------
2 2 2
2 2 2 2 2
- f (n i (t))N (t) n i (t)F (t)
18
F (t) N (t) F (t)N (t)
2 2
S S
1
- g (n i (t))n i (t) n k (t) n i (t) g (n k (t))n k (t)
2 2 2 2 2 2 2 2
= -----------------------
2 2
F (t)N (t)
k=1 k=1
2
2 S
n i (t)
nk(t) g (ni (t)) g (n k(t))
2 2 2 2 2
= -----------------------
2 2
- ,
F (t)N (t) k=1
where
2 2
2 2 f (n i (t))
g (n i (t)) = ------------------
2
.
n i (t)
Combining this expression with our previous equation, we obtain
18-37
2
S
d 2 + 2 2 2
n k(t) g (ni (t)) g (n k(t))
2 2 2 2
n i (t) = b n i (t) .
dt
k=1
This form of the differential equation describing the evolution of the rela-
tive outputs is very useful in demonstrating the characteristics of Layer 2,
as we will see in the next solved problem.
P18.6 Suppose that the transfer function in Layer 2 of the Grossberg net-
work is linear.
i. Show that the relative outputs of Layer 2 will not change af-
ter the input has been removed.
ii. Under what conditions will the total output of Layer 2 decay
to zero after the input has been removed?
i. From Problem P18.5 we know that the relative outputs of Layer 2, after
the input has been removed, evolve according to
2
S
d 2 + 2 2 2
n k(t) g (ni (t)) g (n k(t))
2 2 2 2
n i (t) = b n i (t) .
dt
k=1
2
If the transfer function for Layer 2, f (n) , is linear, then
2
f ( n) = c n .
Therefore
2
2 f ( n) cn
g (n) = ----------- = ------- = c .
n n
If we substitute this expression into our differential equation, we find

2
S
d 2 + 2 2 2

dt
n i (t) = b n i (t) n k(t) c c = 0.
k=1
Therefore the relative outputs do not change.

ii. From Problem P18.5, the total output of Layer 2, after the input has
been removed, evolves according to
2
dN (t) 2 2 + 2 2
--------------- = 1 + F (t) N (t) + b F (t) .
dt
18-38
Solved Problems
2
If f (n) is linear, then
2 2 2
S S S
f (n k(t)) c n k(t) n k(t)

2 2 2 2 2 2
F (t) = = = c = c N (t) .
k=1 k=1 k=1
Therefore the differential equation can be written

2
dN (t) 2 2 + 2 2 + 2 2 2
--------------- = 1 + c N (t) N (t) + b c N (t) = 1 b c + c N (t) N (t) .
dt
To find the equilibrium solutions of this equation, we set the derivative to

zero:
+ 2 2 2
0 = 1 b c + c N (t) N (t) .
Therefore there are two equilibrium solutions:

+ 2
2 2 b c1
N (t) = 0 or N (t) = ------------------- .
c
We want to know the conditions under which the total output will converge
to each of these possible solutions. Consider two cases:
+ 2
1. 1 b c
For this case, the derivative of the total output,
dN (t)
2
+ 2 2 2
--------------- = 1 b c + c N (t) N (t) ,
dt
18
2
will always be negative for positive N (t) . (Recall that the outputs of
Layer 2 are never negative.) Therefore, the total output will decay
to zero.
2
lim N (t) = 0
t
+ 2
2. 1 b c
2 + 2
(a) If N (0) b c 1 c , then the derivative of the total output will
2 + 2
be negative until N (t) = b c 1 c , when the derivative will be
zero. Therefore,
+ 2
2 b c 1
lim N (t) = ------------------------ .
t c
18-39
2 + 2
(b) If N (0) b c 1 c , then the derivative of the total output will
2 + 2
be positive until N (t) = b c 1 c , when the derivative will be ze-
ro. Therefore,
+ 2
2 b c 1
lim N (t) = ------------------------ .
t c
Therefore, if the transfer function of Layer 2 is linear, the total output will
+ 2 + 2
decay to zero if 1 b c . If 1 b c , then the total output will converge to
+ 2
b c 1 c . In any case, the relative outputs will remain constant.
As an example of these results, consider the following Layer 2 equations:
2
dn 1(t) 2 2 2 2 2
-------------- = n 1(t) + 1.5 n 1(t) n 1(t) n 1(t) n 2(t) ,
dt
2
dn 2(t) 2 2 2 2 2
-------------- = n 2(t) + 1.5 n 2(t) n 2(t) n 2(t) n 1(t) .
dt
+ 2 + 2
For this case, = 1 , b = 1.5 and c = 1 , therefore 1 b c . The total out-
put will converge to
+ 2
2 b c 1 1.5 1
lim N (t) = ------------------------ = --------------------- = 0.5 .
t c 1
In Figure P18.6 we can see the response of Layer 2 for two different sets of
initial conditions:
n (0) = 0.75 and n (0) = 0.15 .

2 2
0.5 0.1
As expected, the total output converges to 0.5 for both initial conditions. In
addition, since the relative values of the initial conditions are the same for
the two cases, the outputs converge to the same values in both cases.
18-40
Solved Problems
1.5
2
N (t)
0.5
2
n 1(t)
2
n 2(t)
0
0 5 10
2
Figure P18.6 Response of Layer 2 for Linear f (n)
P18.7 Show that the continuous-time Hebb rule with decay, given by Eq.
(18.24), is equivalent to the discrete-time version given by Eq.
(15.18).
The continuous-time Hebb rule with decay is
2
dw i j(t) 2 2 1
----------------- = w i j(t) + n i (t)n j (t) .
dt

2 2 2
dw i j(t) w i j(t + t) w i j(t)
----------------- ---------------------------------------------- ,
dt t
the Hebb rule becomes

2 2 2 2 1
w i j(t + t) = w i j(t) + t w i j(t) + n i (t)n j (t) .
This can be rearranged to obtain

2 2 2 1
w i j(t + t) = 1 t w i j(t) + t n i (t)n j (t) .
In vector form this would be
2 2 2 1 T
W (t + t) = 1 t W (t) + t n (t) n (t) .
If we compare this with Eq. (15.18),

T
W q = 1 W q 1 + a q p q ,
we can see that they have the identical form.
18-41
Epilogue
The Grossberg network presented in this chapter was inspired by the visu-
al system of higher vertebrates. To motivate the network, we presented a
brief description of the primary visual pathway. We also discussed some vi-
sual illusions, which help us to understand the mechanisms underlying the
visual system.
The Grossberg network is a two-layer, continuous-time competitive net-
work, which is very similar in structure and operation to the Kohonen com-
petitive network presented in Chapter 14. The first layer of the Grossberg
network normalizes the input pattern. It demonstrates how the visual sys-
tem can use on-center/off-surround connection patterns and a shunting
model to implement an automatic gain control, which normalizes total ac-
tivity.
The second layer of the Grossberg network performs a competition, which
contrast enhances the output pattern and stores it in short-term memory.
It uses nonlinear feedback and the on-center/off-surround connection pat-
tern to produce the competition and the storage. The choice of the transfer
function and the feedback connection pattern determines the degree of
competition (e.g., winner-take-all, mild contrast enhancement, or no
change in the pattern).
The adaptive weights in the Grossberg network use an instar learning rule,
which stores prototype patterns in long-term memory. When a winner-
take-all competition is performed in the second layer, this learning rule is
equivalent to the Kohonen learning rule used in Chapter 14.
As with the Kohonen network, one key problem of the Grossberg network
is the stability of learning; as more inputs are applied to the network, the
weight matrix may never converge. This problem was discussed extensive-
ly in Chapter 14. In Chapter 16 we will present a class of networks that is
designed to overcome this difficulty: the Adaptive Resonance Theory (ART)
networks. The ART networks are direct descendents of the Grossberg net-
work presented in this chapter.
Another problem with the Grossberg network, which we have not discussed
in this chapter, is the stability of the differential equations that implement
the network. In Layer 2, for example, we have a set of differential equations
with nonlinear feedback. Can we make some general statement about the
stability of such systems? Chapter 17 will present a comprehensive discus-
sion of this problem.
18-42
Further Reading
Further Reading
[GrMi89] S. Grossberg, E. Mingolla and D. Todorovic, A neural net-
work architecture for preattentive vision, IEEE Transac-
tions on Biomedical Engineering, vol. 36, no. 1, pp. 6584,
1989.
The objective of this paper is to develop a neural network
for general purpose preattentive vision. The network con-
sists of two main subsystems: a boundary contour system
and a feature contour system.
feature detectors, Biological Cybernetics, vol. 23, pp. 121
134, 1976.
Grossberg describes a continuous-time competitive net-
work, inspired by the developmental physiology of the visu-
al cortex. The structure of this network forms the
foundation for other important networks.
mental concepts that are used in later Grossberg networks,
such as the adaptive resonance theory networks.
[Hube88] D.H. Hubel, Eye, Brain, and Vision, New York: Scientific
18
American Library, 1988.
David Hubel has been at the center of research in this area
for 30 years, and his book provides an excellent introduc-
tion to the human visual system. He explains the current
view of the visual system in a way that is easily accessible
to anyone with some scientific training.
[vanT75] H. F. J. M. van Tuijl, A new visual illusion: Neonlike color
spreading and complementary color induction between
subjective contours, Acta Psychologica, vol. 39, pp. 441
445, 1975.
This paper describes the original discovery of the illusion in
which crosses of certain colors, when placed inside Ehren-
stein figures, appear to spread into solid shapes.
18-43
[vond73] C. von der Malsburg, Self-organization of orientation sen-

sitive cells in the striate cortex, Kybernetic, vol. 14, pp. 85
100, 1973.
Malsbergs is one of the first papers to present a self-orga-
nizing feature map neural network. The network is a model
for the visual cortex of higher vertebrates. This paper influ-
enced the work of Kohonen and Grossberg on feature maps.
18-44
Exercises
Exercises
E18.1 Consider the leaky integrator shown in Figure E18.1.
i. Find the response n t if = 1 , n 0 = 1 and p t = 0.5 .
ii. Find the response n t if = 1 , n 0 = 1 and p t = 2 .
iii. Find the response n t if = 4 , n 0 = 1 and p t = 2 .
2+2
ans = iv. Check your answers to the previous parts by writing a MATLAB
4 M-file to simulate the leaky integrator. Use the ode45 routine. Plot
the response for each case.
Leaky Integrator
.
+ n n
p 1/
-
dn/dt = - n + p
Figure E18.1 Leaky Integrator
E18.2 Consider the shunting network shown in Figure E18.2.

i. Find and sketch the response of the shunting network if = 2 ,
+ - + -
b = 3 , b = 1 , p = 0 , p = 5 and n(0) = 1 .
ii. Find and sketch the response of the shunting network if = 2 ,
+ - + -
b = 3 , b = 1 , p = 0 , p = 50 and n(0) = 1 .
iii. Find and sketch the response of the shunting network if = 2 ,
+ - + -
b = 3 , b = 1 , p = 50 , p = 0 and n(0) = 1 .
iv. Find and sketch the response of the shunting network if = 5 ,
+ - + -
b = 2 , b = 6 , p = 5 , p = 0 and n(0) = 0 .
v. Find and sketch the response of the shunting network if = 5 ,
+ - + -
b = 2 , b = 6 , p = 0 , p = 5 and n(0) = 0 .
vi. Find and sketch the response of the shunting network if = 0.25 ,
+ - + -
b = 4 , b = 2 , p = 2 , p = 2 and n(0) = 0 .
vii. Find and sketch the response of the shunting network if = 0.25 ,
+ - + -
b = 4 , b = 2 , p = 2 , p = 4 and n(0) = 0 .
18-45
viii. Check your answers to the previous parts by writing a MATLAB

2+2 M-file to simulate the shunting network. Use the ode45 routine.
ans =
4
Plot the response for each case. Verify that your responses agree
with the known characteristics of the shunting model.
ix. Explain the differences in operation of the leaky integrator and the
shunting network.
b+ +
p+
-
.
+ + n n
1/
- -
+
p-
+
b-
dn/dt = -n + (b+ - n) p+ - (n + b-) p-
Figure E18.2 Shunting Network
E18.3 Suppose that Layer 1 of the Grossberg network has two neurons, with
+ 1 T
b = 0.5 , = 0.5 and input vector p = 2 1 . Assume that the initial
conditions are set to zero.
i. Find the steady state response of Layer 1, using Eq. (18.13).

ii. Find the solution to the differential equation for Layer 1. Verify that
the steady state response agrees with your answer to part (i).
2+2
ans = iii. Check your answer by writing a MATLAB M-file to simulate Layer
4 1 of the Grossberg network. Use the ode45 routine. Plot the re-
sponse.
T
E18.4 Repeat Exercise E18.3 for input vector p = 20 10 .
E18.5 Consider the first layer of the Grossberg network. The parameters are set
+ 1 - 1 T
to be b = 2 , b = 0 , = 2 . The input to the network is p = 4 1 . Find
the first layer outputs and sketch them versus time.
18-46
Exercises
E18.6 Find the differential equation that describes the variation in the total out-
put of Layer 1,
1
S
ni (t) .
1 1
N (t) =
i=1
(Use the technique presented in Problem P18.5.)
E18.7 Assume that Layer 2 of the Grossberg network has two neurons, with
2 + 2 - 2
f n = 2n , = 1 , b = 1 and b = 0 . The inputs have been applied for
some length of time, then removed.
2
i. What will be the steady state total output, lim N (t) ?
t
+ 2
ii. Repeat part (i) if b = 0.25 .
2+2
ans = iii. Check your answers to the previous parts by writing a MATLAB
4 M-file to simulate Layer 2 of the Grossberg network. Use the ode45
routine. Plot the responses for the following initial conditions:
n (0) = 2 and n (0) = 0.2 .

2 2
1 0.1
E18.8 Suppose that the transfer function for Layer 2 of the Grossberg network is
2 2 + 2
f n = c n , and = 1 , b = 1 .
i. Using the results of Problem P18.5, show that, after the inputs have
been removed, all of the relative outputs of Layer 2 will decay to ze-
ro, except the one with the largest initial condition (winner-take-all
competition).
2
ii. For what values of c will the total output N (t) have a nonzero sta-
ble point (steady state value)?
iii. If the condition of part (ii) is satisfied, what will be the steady state
2 2
value of N (t) ? Will this depend on the initial condition N (0) ?
2+2
ans = iv. Check your answers to the previous parts by writing a MATLAB
4 M-file and simulating the total response of Layer 2 for c = 4 and
2
N ( 0) = 3 .
E18.9 Simulate the response of the adaptive weights for the Grossberg network.
Assume that the coefficient is 1 . Assume that two different input pat-
2+2 terns are alternately presented to the network for periods of 0.2 seconds at
ans = a time. Also, assume that Layer 1 and Layer 2 converge very quickly, in
4
comparison with the convergence of the weights, so that the neuron out-
18-47
puts are effectively constant over the 0.2 seconds. The Layer 2 and Layer 1
outputs for the two different input patterns will be:
for pattern 1: n = 0.8 , n = 1 ,

1 2
0.2 0
for pattern 2: n = 0.5 , n = 0 .

1 2
0.5 1
E18.10 Repeat Exercise E18.9, but use the Hebb rule with decay, Eq. (18.24), in-
2+2
stead of the instar learning of Eq. (18.25). Explain the differences between
ans = the two responses.
4
18-48
Objectives
19 Adaptive Resonance Theory

Objectives 19-1
Overview of Adaptive Resonance 19-2
Layer 1 19-4
Steady State Analysis 19-6
Layer 2 19-10
Orienting Subsystem 19-13
Subset/Superset Dilemma 19-17
Learning Law 19-18
ART1 Algorithm Summary 19-21
Initialization 19-21
Algorithm 19-21
Other ART Architectures 19-23
Epilogue 19-44
Exercises 19-47
Objectives
In Chapter 16 and Chapter 18 we learned that one key problem of compet-
itive networks is the stability of learning. There is no guarantee that, as
more inputs are applied to the network, the weight matrix will eventually
converge. In this chapter we will present a modified type of competitive
learning, called adaptive resonance theory (ART), which is designed to
overcome the problem of learning stability.
19-1
Theory and Examples

A key problem of the Grossberg network presented in Chapter 18 and the
competitive networks of Chapter 16, is that they do not always form stable
clusters (or categories). Grossberg did show [Gros76] that if the number of
input patterns is not too large, or if the input patterns do not form too many
clusters relative to the number of neurons in Layer 2, then the learning
eventually stabilizes. However, he also showed that the standard competi-
tive networks do not have stable learning in response to arbitrary input
patterns. The learning instability occurs because of the networks adapt-
ability (or plasticity), which causes prior learning to be eroded by more re-
cent learning.
Stability/Plasticity Grossberg refers to this problem as the stability/plasticity dilemma. How
can a system be receptive to significant new patterns and yet remain stable
in response to irrelevant patterns? We know that biological systems are
very good at this. For example, you can easily recognize your mothers face,
even if you have not seen her for a long time and have met many new people
in the mean time.
Grossberg and Gail Carpenter developed a theory, called adaptive reso-
nance theory (ART), to address the stability/plasticity dilemma (see
[CaGr87a], [CaGr87b], [CaGr90], [CaGrRe91] and [CaGrMa92]). The ART
networks are based on the Grossberg network of Chapter 18. The key inno-
vation of ART is the use of expectations. As each input pattern is present-
ed to the network, it is compared with the prototype vector that it most
closely matches (the expectation). If the match between the prototype and
the input vector is not adequate, a new prototype is selected. In this way,
previously learned memories (prototypes) are not eroded by new learning.
It is beyond the scope of this text to discuss all of the variations of adaptive
resonance theory. Instead, we will present one of the ART networks in de-
tail ART1 (see [CaGr87a]). This particular network is designed for bina-
ry input vectors only. However, from this one architecture, the key features
of adaptive resonance theory can be understood.
Overview of Adaptive Resonance

The basic ART architecture is shown in Figure 19.1. It is a modification of
the Grossberg network of Chapter 18 (compare with Figure 18.13), which
is designed to stabilize the learning process. The innovations of the ART ar-
chitecture consist of three parts: Layer 2 (L2) to Layer 1 (L1) expectations,
the orienting subsystem and gain control. In this section we will describe
the general operation of the ART system; then, in later sections, we will dis-
cuss each subsystem in detail.
19-2
Overview of Adaptive Resonance
Layer 1 Layer 2
Gain Control
Input
Expectation
Reset
Orienting
Subsystem
Figure 19.1 Basic ART Architecture

Recall from Chapter 18 that the L1-L2 connections of the Grossberg net-
work are instars, which perform a clustering (or categorization) operation.
When an input pattern is presented to the network, it is multiplied (after
normalization) by the L1-L2 weight matrix. Then, a competition is per-
formed at Layer 2 to determine which row of the weight matrix is closest to
the input vector. That row is then moved toward the input vector. After
learning is complete, each row of the L1-L2 weight matrix is a prototype
pattern, which represents a cluster (or category) of input vectors.
In the ART networks, learning also occurs in a set of feedback connections
from Layer 2 to Layer 1. These connections are outstars (see Chapter 15),
which perform pattern recall. When a node in Layer 2 is activated, this re-
produces a prototype pattern (the expectation) at Layer 1. Layer 1 then
performs a comparison between the expectation and the input pattern.
When the expectation and the input pattern are not closely matched, the
orienting subsystem causes a reset in Layer 2. This reset disables the cur-
rent winning neuron, and the current expectation is removed. A new com-
petition is then performed in Layer 2, while the previous winning neuron
is disabled. The new winning neuron in Layer 2 projects a new expectation
to Layer 1, through the L2-L1 connections. This process continues until the
L2-L1 expectation provides a close enough match to the input pattern.
In the following sections we will investigate each of the subsystems of the
ART system, as they apply to one particular ART network ART1
([CaGr87a]). We will first describe the differential equations that describe
the subsystem operations. Then we will derive the steady state responses
of each subsystem. Finally, we will summarize the overall operation of the
ART1 system.
19-3
Layer 1
The main purpose of Layer 1 is to compare the input pattern with the ex-
pectation pattern from Layer 2. (Both patterns are binary in ART1.) If the
patterns are not closely matched, the orienting subsystem will cause a re-
set in Layer 2. If the patterns are close enough, Layer 1 combines the ex-
pectation and the input to form a new prototype pattern.
Layer 1 of the ART1 network, which is displayed in Figure 19.2, is very
similar to Layer 1 of the Grossberg network (see Figure 18.14). The differ-
ences occur at the excitatory and inhibitory inputs to the shunting model.
For the ART1 network, no normalization is performed at Layer 1; therefore
we dont have the on-center/off-surround connections from the input vector.
The excitatory input to Layer 1 of ART1 consists of a combination of the in-
put pattern and the L1-L2 expectation. The inhibitory input consists of the
gain control signal from Layer 2. In the following we will explain how these
inputs work together.
Input Layer 1
+ +
Expectation
W2:1 a2
+b1 S1 x S2
+
-
.
p + + n1 n1 a1
1/ f1
S1 x 1
- -
+
S1 S1
-b1 +
-W1 a2
Gain Control
S1 x S2
dn1/dt = - n1 + (+b1 - n1) { p + W2:1 a2} - (n1 + -b1) [-W1] a2
Figure 19.2 Layer 1 of the ART1 Network

1
dn (t) 1 + 1 1 2:1 2
-------------- = n (t) + b n (t) p + W a t (19.1)
dt
1 - 1 - 1 2
n (t) + b W a t
and the output of Layer 1 is computed
19-4
Layer 1
1 + 1
a = hardlim n , (19.2)
where
+ 1, n 0
hardlim n = . (19.3)
0, n 0
2:1 2
Eq. (19.1) is a shunting model with excitatory input p + W a t , which is
the sum of the input vector and the L2-L1 expectation. For example, as-
sume that the jth neuron in Layer 2 has won the competition, so that its
output is 1, and the other neurons have zero output. For this case we have
0
0
2:1 2 2:1
W a = w 2:1 2:1 2:1
1 w2 wj
2:1
w 2 = wj ,

S
(19.4)
1

2:1 2:1 2:1
where w j is the jth column of the matrix W . (The W matrix is
trained using an outstar rule, as we will show in a later section.) Now we
can see that
2:1 2 2:1
p+W a = p + wj . (19.5)
Therefore the excitatory input to Layer 1 is the sum of the input pattern
and the L2-L1 expectation. Each column of the L2-L1 matrix represents a
different expectation (prototype pattern). Layer 1 combines the input pat-
tern with the expectation using an AND operation, as we will see later.
- 1 2
The inhibitory input to Layer 1 is the gain control term W a t , where
1 1 1
1
W = 1 1
- 1
. (19.6)

1 1 1
Therefore, the inhibitory input to each neuron in Layer 1 is the sum of all
of the outputs of Layer 2. Since we will be using a winner-take-all compe-
tition in Layer 2, whenever Layer 2 is active there will be one, and only one,
2
nonzero element of a after the competition. Therefore the gain control in-
put to Layer 1 will be one when Layer 2 is active, and zero when Layer 2 is
inactive (all neurons having zero output). The purpose of this gain control
will become apparent as we analyze the steady state behavior of Layer 1.
19-5
Steady State Analysis

The response of neuron i in Layer 1 is described by
1 S
2
S
2
dn i 1 2:1 2
-------- = n i + b n i p i + w i j a j n i + b a j ,
1 + 1 1 - 1 2
(19.7)
dt
j=1
j=1
where 1 so that the short-term memory traces (the neuron outputs)

change much faster than the long-term memory traces (the weight matri-
ces).
We want to investigate the steady state response of this system for two dif-
2
ferent cases. In the first case Layer 2 is inactive, therefore a j = 0 for all j .
In the second case Layer 2 is active, and therefore one neuron has an out-
put of 1, and all other neurons output 0.
2
Consider first the case where Layer 2 is inactive. Since each a j = 0 , Eq.
(19.7) simplifies to
1
dn 1 + 1 1
-------i- = n i + b n i p i . (19.8)
dt
1
1 + 1 1 1 + 1
0 = n i + b n i p i = 1 + p i n i + b p i . (19.9)
1
+ 1
1 b p
ni = -------------i . (19.10)
1 + pi
1 1 + 1
Therefore, if p i = 0 then n i = 0 , and if p i = 1 then n i = b 2 0 . Since
+
we chose the transfer function for Layer 1 to be the hardlim function, then
we have
1
a = p. (19.11)
Therefore, when Layer 2 is inactive, the output of Layer 1 is the same as

the input pattern.
Now lets consider the second case, where Layer 2 is active. Assume that
2 2
neuron j is the winning neuron in Layer 2. Then a j = 1 and a k = 0 for
k j . For this case Eq. (19.7) simplifies to
19-6
Layer 1
1
dn 1 + 1 1 2:1 1 - 1
-------i- = n i + b n i p i + w i j n i + b . (19.12)
dt
1
1 + 1 1 2:1 1 - 1
0 = n i + b n i p i + w i j n i + b (19.13)
2:1 1 + 1 2:1 - 1
= 1 + p i + w i j + 1 n i + b p i + w i j b .
1
+ 1 2:1 - 1
1 b p i + w i j b
ni -.
= --------------------------------------------
2:1
(19.14)
2 + p i + w i j
Recall that Layer 1 should combine the input vector with the expectation
2:1
from Layer 2 (represented by w j ). Since we are dealing with binary pat-
terns (both the input and the expectation), we will use a logical AND oper-
1
ation to combine the two vectors. In other words, we want n i to be less than
2:1 1
zero when either p i or w i j is equal to zero, and we want n i to be greater
2:1
than zero when both p i and w i j are equal to one.
If we apply these conditions to Eq. (19.14), we obtain the following equa-

tions:
+ 1 - 1
b 2 b 0 , (19.15)
+ 1 - 1
b b 0, (19.16)
which we can combine to produce

+ 1 - 1 + 1
b 2 b b . (19.17)
+ 1 - 1
For example, we can use b = 1 and b = 1.5 to satisfy these conditions.
Therefore, if Eq. (19.17) is satisfied, and neuron j of Layer 2 is active, then
the output of Layer 1 will be
1 2:1
a = p wj , (19.18)
where represents the logical AND operation.

Notice that we needed the gain control in order to implement the AND op-
eration. Consider the numerator of Eq. (19.14):
+ 1 2:1 - 1
b p i + w i j b . (19.19)
19-7
- 1
The term b is multiplied by the gain control term, which in this case is 1.
If this term did not exist, then Eq. (19.19) would be greater than zero (and
1 2:1
therefore n i would be greater than zero) whenever either p i or w i j was
greater than zero. This would represent an OR operation, rather than an
AND operation. As we will see when we discuss the orienting subsystem, it
is critical that Layer 1 perform an AND operation.
When Layer 2 is inactive, the gain control term is zero. This is necessary
because in that case we want Layer 1 to respond to the input pattern alone,
since no expectation will be activated by Layer 2.
To summarize the steady state operation of Layer 1:
2
If Layer 2 is not active (i.e., each a j = 0 ),
1
a = p. (19.20)
2
If Layer 2 is active (i.e., one a j = 1 ),
1 2:1
a = p wj . (19.21)
2 To demonstrate the operation of Layer 1, assume the following network pa-
+2
rameters:
+ 1 - 1
= 0.1 , b = 1 and b = 1.5 . (19.22)
Assume also that we have two neurons in Layer 2, two elements in the in-
put vector and the following weight matrix and input:
= 1 1 and p = 0 .
2:1
W (19.23)
0 1 1
If we take the case where Layer 2 is active, and neuron 2 of Layer 2 wins
the competition, the equations of operation of Layer 1 are
1
dn 1 1 1 2:1 1
0.1 -------- = n 1 + 1 n 1 p 1 + w 1 2 n 1 + 1.5 (19.24)
dt
1 1 1 1
= n 1 + 1 n 1 0 + 1 n 1 + 1.5 = 3n 1 0.5
1
dn 1 1 2:1 1
0.1 -------2- = n 2 + 1 n 2 p 2 + w 2 2 n 2 + 1.5 (19.25)
dt
1 1 1 1
= n 2 + 1 n 2 1 + 1 n 2 + 1.5 = 4n 2 + 0.5 .
These can be simplified to obtain
19-8
Layer 1
1
dn 1 1
-------- = 30n 1 5 , (19.26)
dt
1
dn 2 1
-------- = 40n 2 + 5 . (19.27)
dt
In this simple case we can find closed-form solutions for these equations. If
we assume that both neurons start with zero initial conditions, the solu-
tions are
1 1 30t
n 1 t = --- 1 e , (19.28)
6
1 1 40t
n 2 t = --- 1 e . (19.29)
8
These are displayed in Figure 19.3.

0.2
1
0.1
n2 t
-0.1
1
n1 t
-0.2
0 0.05 0.1 0.15 0.2
Figure 19.3 Response of Layer 1

1 1
Note that n1 t
converges to a negative value, and n 2 t converges to a pos-
1 1
itive value. Therefore, a 1 t converges to 0, and a 2 t converges to 1 (recall
+
that the transfer function for Layer 1 is hardlim ). This agrees with our
steady state analysis (see Eq. (19.21)), since
= 0 1 = 0 = a .
2:1 1
p w2 (19.30)
1 1 1
To experiment with Layer 1 of the ART1 network, use the Neural Network
Design Demonstration ART1 Layer 1 (nnd16al1).
19-9
Layer 2
Layer 2 of the ART1 network is almost identical to Layer 2 of the Grossberg
network of Chapter 18. Its main purpose is to contrast enhance its output
pattern. For our implementation of the ART1 network, the contrast en-
hancement will be a winner-take-all competition, so only the neuron that
receives the largest input will have a nonzero output.
There is one major difference between the second layers of the Grossberg
a0 and the ART1 networks. Layer 2 of the ART1 network uses an integrator
. that can be reset. In this type of integrator, whose symbol is shown in the
n n 0
left margin, any positive outputs are reset to zero whenever the a signal
becomes positive. The outputs that are reset remain inhibited for a long pe-
riod of time, so that they cannot be driven above zero. (By a long period
of time we mean until an adequate match has occurred and the weights
have been updated.)
In the original ART1 paper, Carpenter and Grossberg suggested that the
reset mechanism could be implemented using a gated dipole field
[CaGr87]. They later suggested a more sophisticated biological model, us-
ing chemical neurotransmitters, in their ART3 architecture [CaGr90]. For
our purposes we will not be concerned with the specific biological imple-
mentation.
Figure 19.4 displays the complete Layer 2 of the ART1 network. Again, it
is almost identical to Layer 2 of the Grossberg network (see Figure 18.16),
0
with the primary exception of the resetable integrator. The reset signal, a ,
is the output of the orienting subsystem, which we will discuss in the next
section. It generates a reset whenever there is a mismatch at Layer 1 be-
tween the input signal and the L2-L1 expectation.
One other small difference between Layer 2 of the ART1 network and Lay-
er 2 of the Grossberg network is that two transfer functions are used in
2 2
ART1. The transfer function f n is used for the on-center/off-surround
feedback connections, while the output of Layer 2 is computed as
2 + 2
a = hardlim n . The reason for the second transfer function is that we
want the output of Layer 2 to be a binary signal.
19-10
Layer 2
Layer 2
+ +
On-Center +W2
W 1:2
S2 x S1
+b2 S2 x S2
+
-
. f2
a1 + + n2 n2
1/
- - a2
+
a0 -b2 +
S2
Reset -W2
Off-Surround
S2 x S2
dn2/dt = - n2 + (+b2 - n2) {[+W2] f 2(n2) + W1:2 a1}

- (n2 + -b2) [-W2] f 2(n2)
Figure 19.4 Layer 2 of the ART1 Network
2
dn (t) 2 + 2 2 + 2 2 2 1:2 1
-------------- = n (t) + b n (t) W f (n (t)) + W a (19.31)
dt
2 - 2 - 2 2 2
n (t) + b W f (n (t)) .
+ 2 2 2 1:2 1
This is a shunting model with excitatory input W f (n (t)) + W a ,
+ 2
where W provides on-center feedback connections (identical to Layers 1
1:2
and 2 of the Grossberg network of Chapter 18, Eq. (18.6)), and W con-
sists of adaptive weights, analogous to the weights in the Kohonen net-
work. They are trained according to an instar rule, as we will see in a later
1:2
section. The rows of W , after training, will represent the prototype pat-
terns.
- 2 2 2 - 2
The inhibitory input to the shunting model is W f (n (t)) , where W pro-
vides off-surround feedback connections (identical to Layers 1 and 2 of the
Grossberg network Eq. (18.7)).
19-11
To illustrate the performance of Layer 2, consider a two-neuron layer with

2
+2
1:2 T
w
= 0.1 , b = 1 , b = 1 , W = 1 = 0.5 0.5 , (19.32)
+ 2 - 2 1:2
1 1 1:2 T 1 0
2w
and
2 10 n 2 , n0
f ( n) = . (19.33)
0, n0
The equations of operation of the layer will be

2
dn 1(t) 2 2 2 2 1:2 T 1
0.1 -------------
- = n 1(t) + 1 n 1(t) f (n 1(t)) + 1w a (19.34)
dt
2 2 2
n 1(t) + 1 f (n 2(t))
2
dn 2(t) 2 2 2 2 1:2 T 1
0.1 -------------
- = n 2(t) + 1 n 2(t) f (n 2(t)) + 2w a (19.35)
dt
2 2 2
n 2(t) + 1 f (n 1(t)) .
This is identical in form to the Grossberg Layer 2 example in Chapter 18

- 2 2
(see Eq. (18.20) and Eq. (18.21)), except that b = 1 . This will allow n 1(t)
2
and n 2(t) to range between -1 and +1.
The inputs to Layer 2 are the inner products of the prototype patterns
1:2
(rows of the weight matrix W ) with the output of Layer 1. (The rows of
this weight matrix are normalized, as will be explained in a later section.)
The largest inner product will correspond to the prototype pattern that is
closest to the output of Layer 1. Layer 2 then performs a competition be-
2
tween the neurons. The transfer function f (n) is chosen to be a faster-than-
linear transfer function (see Chapter 18, page 18-20, for a discussion of the
2
effect of f (n) ). This choice will force the neuron with largest input to have
a positive n , and the other neuron to have a negative n (with appropriate
choice of network parameters). After the competition, one neuron output
will be 1, and the other neuron output will be zero, since we are using the
+
hardlim transfer function to compute the layer output.
Figure 19.5 illustrates the response of Layer 2 when the input vector is
1 T 1:2 1
a = 1 0 . The second row of W has a larger inner product with a
19-12
Orienting Subsystem
than the second row, therefore neuron 2 wins the competition. At steady
2 2
state, n 2(t) has a positive value, and n 1(t) has a negative value. The steady
state Layer 2 output will then be
a = 0 .
2
(19.36)
1
1
1:2 T 1
2w a 2
n2 t
0.5
1:2 T 1
1w a
0
-0.5 2
n1 t
-1
0 0.05 0.1 0.15 0.2

t
Figure 19.5 Response of Layer 2

We can summarize the steady state operation of Layer 2 as follows:
1:2 T 1 1:2 T 1
2 1, if iw a = max jw a
ai = . (19.37)
0, otherwise

To experiment with Layer 2 of the ART1 network, use the Neural Network
Design Demonstration ART1 Layer 2 (nnd16al2).
Orienting Subsystem
One of the key elements of the ART architecture is the Orienting Sub-
system. Its purpose is to determine if there is a sufficient match between
the L2-L1 expectation and the input pattern. When there is not enough of
a match, the Orienting Subsystem should send a reset signal to Layer 2.
The reset signal will cause a long-lasting inhibition of the previous winning
neuron, and thus allow another neuron to win the competition.
Figure 19.6 displays the Orienting Subsystem.
19-13
Orienting Subsystem
+b0
+
p +W0
-
1 x S1 .
+ + n0 n0 a0
1/ f0
- - Reset
+
a1 -W0 1
1 x S1 -b0 +
dn0/dt = -n0 + (+b0 - n0) [+W0] p - (n0 + -b0) [-W0] a 1
Figure 19.6 Orienting Subsystem of the ART1 Network

The equation of operation of the Orienting Subsystem is
0
dn (t) 0 + 0 0 + 0 0 - 0 - 0 1
-------------- = n (t) + b n (t) W p n (t) + b W a . (19.38)
dt
+ 0
This is a shunting model, with excitatory input W p , where
+ 0
W = . (19.39)
Therefore, the excitatory input can be written

1
S
W p = p = pj = p ,
+ 0 2
(19.40)
j=1
where the last equality holds because p is a binary vector.

- 0 1
The inhibitory input to the Orienting Subsystem is W a , where
- 0
W = . (19.41)
Therefore, the inhibitory input can be written

1
S
1 2
W a = a = aj t = a
- 0 1 1 1
. (19.42)
j=1
19-14
Orienting Subsystem
Whenever the excitatory input is larger than the inhibitory input, the Ori-
enting Subsystem will be driven on. Consider the following steady state op-
eration:
0 + 0 0 2 0 - 0 1 2
0 = n + b n p n + b a (19.43)
2 1 2 0 + 0 2 - 0 1 2
= 1 + p + a n + b p b a .
0
If we solve for n , we find
+ 0 2 - 0 1 2
0 b p b a
n = ------------------------------------------------------------ . (19.44)
2 1 2
1 + p + a
+ 0 - 0 0 2 1 2
Let b = b = 1 , then n 0 if p a 0 , or in other words:
1 2
0 a
n 0 if -----------2- --- = . (19.45)
p
This is the condition that will cause a reset of Layer 2, since

0 + 0
Vigilance a = hardlim n . The term is called the vigilance parameter, and must
fall in the range 0 1 . If the vigilance is close to 1, a reset will occur un-
1 1
less a is close to p . If the vigilance is close to 0, a need not be close to p
to prevent a reset. The vigilance parameter determines the coarseness of
the categorization (or clustering) created by the prototype vectors.
1 2:1
Recall from Eq. (19.21) that a = p w j whenever Layer 2 is active.
2 1 2
Therefore, p will always be greater than or equal to a . They will be
2:1
equal when the expectation w j has a 1 wherever the input p has a 1.
Therefore, the orienting subsystem will cause a reset when there is enough
2:1
of a mismatch between p and w j . The amount of mismatch required for
a reset is determined by the vigilance parameter .
2 To demonstrate the operation of the Orienting Subsystem, suppose that
+2
= 0.1 , = 3 , = 4 ( = 0.75 ),
p = 1 and a = 1 .
1
(19.46)
1 0
The equation of operation becomes

0
dn (t) 0 0
0.1 -------------- = n (t) + 1 n (t) 3 p 1 + p 2 (19.47)
dt
0 1 1
n (t) + 1 4 a 1 + a 2
19-15
or
0
dn (t) 0
-------------- = 110n (t) + 20 . (19.48)
dt
The response is plotted in Figure 19.7. In this case a reset signal will be
0
sent to Layer 2, since n t is positive. In this example, because the vigi-
lance parameter is set to = 0.75 , and p has only two elements, we will
1
have a reset whenever p and a are not identical. (If the vigilance param-
1
eter were set to = 0.25 , we would not have had a reset for the p and a
1 2 2
of Eq. (19.46), since a p = 1 2 .)
0.2
0
n t
0.1
-0.1
-0.2
0 0.05 0.1 0.15 0.2
Figure 19.7 Response of the Orienting Subsystem

The steady state operation of the Orienting Subsystem can be summarized
as follows:
1 2 2
0 1, if a p
a = . (19.49)
0, otherwise

To experiment with the Orienting Subsystem, use the Neural Network De-
sign Demonstration Orienting Subsystem (nnd16os).
Learning Law: L1-L2

The ART1 network has two separate learning laws: one for the L1-L2 con-
nections, and another for the L2-L1 connections. The L1-L2 connections
use a type of instar learning to learn to recognize a set of prototype pat-
terns. The L2-L1 connections use outstar learning in order to reproduce (or
recall) a set of prototype patterns. In this section we will describe the L1-
19-16
Learning Law: L1-L2
L2 instar learning law, and in the following section we will present the L2-
L1 outstar learning law.
We should note that the L1-L2 connections and the L2-L1 connections are
updated at the same time. Whenever the input pattern and the expectation
have an adequate match, as determined by the Orienting Subsystem, both
1:2 2:1
W and W are adapted. This process of matching, and subsequent ad-
Resonance aptation, is referred to as resonance, hence the name adaptive resonance
theory.
Subset/Superset Dilemma
The learning in the L1-L2 connections of the ART1 network is very close to
the learning in the Grossberg network of Chapter 18, with one major dif-
ference. In the Grossberg network, the input patterns are normalized in
Layer 1, and therefore all of the prototype patterns will have the same
length. In the ART1 network no normalization takes place in Layer 1.
Therefore a problem can occur when one prototype pattern is a subset of
another. For example, suppose that the L1-L2 connection matrix is
= 1 1 0 ,
1:2
W (19.50)
1 1 1
so that the prototype patterns are
1 1
1:2 1:2
1w = 1 and 2w = 1 . (19.51)
0 1
1:2 1:2 1:2 1:2
We say that 1w is a subset of 2w , since 2w has a 1 wherever 1w has
a 1.
If the output of Layer 1 is 19
1
1
a = 1 , (19.52)
0
then the input to Layer 2 will be
1
W a = 11 0 1 = 2 .
1:2 1
(19.53)
11 1 2
0
19-17
1
Both prototype vectors have the same inner product with a , even though
1
the first prototype is identical to a and the second prototype is not. This is
called the subset/superset dilemma.
One solution to the subset/superset dilemma is to normalize the prototype
patterns. That is, when a prototype pattern has a large number of nonzero
entries, the magnitude of each entry should be reduced. For example, using
our preceding problem, we could modify the L1-L2 matrix as follows:
1 1
--- --- 0
= 2 2
1:2
W . (19.54)
1 1 1
--- --- ---
3 3 3
The input to Layer 2 will then be
1 1
--- --- 0 1 1
a = 2 2
1:2 1
W 1 = 2 . (19.55)
1 1 1 ---
--- --- --- 0 3
3 3 3
Now we have the desired result: the first prototype has the largest inner
1
product with a . The first neuron in Layer 2 will be activated.
In the Grossberg network of Chapter 18 we obtained normalized prototype
patterns by normalizing the input patterns in Layer 1. In the ART1 net-
work we will normalize the prototype patterns by using an on-center/off-
surround competition in the L1-L2 learning law.
Learning Law
1:2
The learning law for W is
1:2
d iw t 2 + 1:2 + 1
------------------------- = a i t b iw t W a t (19.56)
dt
1:2 - - 1
iw t + b W a t ,
where
1 0 1 0 0 0 1 1
1
b = 1 , b = 0 , W = 0 1 0 and -
W = 1 0
+ - +
. (19.57)

1 0 0 0 1 1 1 0
19-18
Learning Law: L1-L2
This is a modified form of instar learning. When neuron i of Layer 2 is ac-

1:2 1:2 1
tive, the ith row of W , iw , is moved in the direction of a . The differ-
ence between Eq. (19.56) and the standard instar learning is that the
1:2 1:2
elements of iw compete, and therefore iw is normalized. In the bracket
on the right side of Eq. (19.56) we can see that it has the form of a shunting
1
model, with on-center/off-surround input connections from a . The excita-
+ -
tory bias is b = 1 (a vector of 1s), and the inhibitory bias is b = 0 , which
1:2
ensures that the elements of iw remain between 0 and 1. (Recall our dis-
cussion of the shunting model in Chapter 18.)
To verify that Eq. (19.56) causes normalization of the prototype patterns,

lets investigate the steady state operation. For this analysis we will as-
sume that the outputs of Layer 1 and Layer 2 remain constant until the
Fast Learning weights reach steady state. This is called fast learning.
1:2
The equation for element w i j is
1:2
dw i j t
= a i t 1 w i j t a j t w i j t a k t .
2 1:2 1 1:2 1
------------------- (19.58)
dt
kj
2
If we assume that neuron i is active in Layer 2 ( a i t = 1 ) and set the de-
rivative to zero in Eq. (19.58), we see that
ak
1:2 1 1:2 1
0 = 1 w i j a j w i j . (19.59)
kj
1:2
To find the steady state value of w i j , we will consider two cases. First, as-
1
sume that a j = 1 . Then we have
1:2 1:2 1 2 1 2 1:2

0 = 1 w i j w i j a 1 = + a 1 w i j + , (19.60)
or
1:2
-.
w i j = ----------------------------- (19.61)
1 2
+ a 1
1
S
1 2
ak
1 1
(Note that = a , since a is a binary vector.)
k=1
1
On the other hand, if a j = 0 , then Eq. (19.59) reduces to
1:2 1 2
0 = w i j a , (19.62)
or
19-19
1:2
w i j = 0 . (19.63)
To summarize Eq. (19.61) and Eq. (19.63):

1
1:2 a
iw -,
= ----------------------------- (19.64)
1 2
+ a 1
where 1 to ensure that the denominator will never equal zero.

Therefore the prototype patterns will be normalized, and this will solve the
subset/superset dilemma. (By normalized here we do not mean that all
prototype vectors will have unit length in Euclidean distance, but simply
1:2
that the rows of W that have more nonzero entries will have elements
with smaller magnitudes. In this case, vectors with more nonzero entries
may actually have a smaller length than vectors with fewer nonzero en-
tries.)
Learning Law: L2-L1

2:1
The L2-L1 connections, W , in the ART1 architecture are trained using
an outstar learning rule. The purpose of the L2-L1 connections is to recall
an appropriate prototype pattern (the expectation), so that it can be com-
pared and combined, in Layer 1, with the input pattern. When the expec-
tation and the input pattern do not match, a reset is sent to Layer 2, so that
a new prototype pattern can be chosen (as we have discussed in previous
sections).
2:1
The learning law for W is a typical outstar equation:
2:1
d wj t 2 2:1 1
- = aj t wj t + a t .
------------------------ (19.65)
dt
Therefore, if neuron j in Layer 2 is active (has won the competition), then

2:1 1
column j of W is moved toward the a pattern. To illustrate this, lets
investigate the steady state operation of Eq. (19.65).
For this analysis we will assume the fast learning scenario, where the out-
puts of Layer 1 and Layer 2 remain constant until the weights reach steady
2
state. Assume that neuron j in Layer 2 is active, so that a j = 1 . Setting
the derivative in Eq. (19.65) to zero, we find
2:1 1 2:1 1
0 = w j + a , or w j = a . (19.66)
2:1 1
Therefore column j of W converges to the output of Layer 1, a . Recall
1
from Eq. (19.20) and Eq. (19.21) that a is a combination of the input pat-
tern and the appropriate prototype pattern. Therefore the prototype pat-
19-20
ART1 Algorithm Summary
tern is modified to incorporate the current input pattern (if there is a close
enough match).
1:2 2:1
Keep in mind that W and W are updated at the same time. When neu-
ron j of Layer 2 is active and there is a sufficient match between the expec-
tation and the input pattern (which indicates a resonance condition), then
1:2 2:1
row j of W and column j of W are adapted. In fast learning, column
2:1 1 1:2
j of W is set to a , while row j of W is set to a normalized version of
1
a .
ART1 Algorithm Summary

Now that we have investigated each of the subsystems of the ART1 archi-
tecture, we can gain some insight into its overall operation if we summarize
the key steady state equations and organize them into an algorithm.
Initialization
The ART1 algorithm begins with an initialization of the weight matrices
1:2 2:1 2:1
W and W . The initial W matrix is set to all 1s. Thus, the first time
a new neuron in Layer 2 wins a competition, resonance will occur, since
1 2:1 1 2 2
a = p w j = p and therefore a p = 1 . This means that any
2:1
untrained column in W is effectively a blank slate and will cause a
match with any input pattern.
1:2
Since the rows of the W matrix should be normalized versions of the col-
2:1 1:2
umns of W , every element of the initial W matrix is set to
1
+ S 1 .
Algorithm
After initialization, the ART1 algorithm proceeds as follows:
1. First, we present an input pattern to the network. Since Layer 2 is not
2
active on initialization (i.e., each a j = 0 ), the output of Layer 1 is (Eq.
(19.20))
1
a = p. (19.67)
2. Next, we compute the input to Layer 2,

1:2 1
W a , (19.68)
and activate the neuron in Layer 2 with the largest input (Eq. (19.37)):
19-21
1:2 T 1 1:2 T 1
2 1, if iw a = max kw a
ai = . (19.69)
0, otherwise

In case of a tie, the neuron with the smallest index is declared the win-
ner.
3. We then compute the L2-L1 expectation (where we assume neuron j of
Layer 2 is activated):
2:1 2 2:1
W a = wj . (19.70)
4. Now that Layer 2 is active, we adjust the Layer 1 output to include the
L2-L1 expectation (Eq. (19.21)):
1 2:1
a = p wj . (19.71)
5. Next, the Orienting Subsystem determines the degree of match be-

tween the expectation and the input pattern (Eq. (19.49)):
1 2 2
0 1, if a p
a = . (19.72)
0, otherwise

0 2
6. If a = 1 , then we set a j = 0 , inhibit it until an adequate match occurs
0
(resonance), and return to step 1. If a = 0 , we continue with step 7.
1:2
7. Resonance has occurred. Therefore we update row j of W (Eq.
(19.61)):
1
1:2 a
jw -.
= ----------------------------- (19.73)
1 2
+ a 1
2:1
8. We now update column j of W (Eq. (19.66)):
2:1 1
wj = a . (19.74)
9. We remove the input pattern, restore all inhibited neurons in Layer 2,
and return to step 1 with a new input pattern.
The input patterns continue to be applied to the network until the weights
stabilize (do not change). Carpenter and Grossberg have shown [CaGr87a]
that the ART1 algorithm will always form stable clusters for any set of in-
put patterns.
See Problems P19.5, P19.6 and P19.7 for detailed examples of the ART1 al-
gorithm.
19-22
Other ART Architectures
To experiment with the ART1 algorithm, use the Neural Network Design
Demonstration ART1 (nnd16a1).
Other ART Architectures

The ART1 network is just one example of adaptive resonance theory. Car-
penter and Grossberg, and others in their research group, have developed
many variations on this theme.
One disadvantage of the ART1 network is that it can only be used for bina-
ry input patterns. Carpenter and Grossberg developed a variation of ART1,
called ART2, to handle either analog or binary patterns [CaGr87b]. The ba-
sic structure of ART2 is very similar to ART1, with the exception of Layer
1. In ART2 several sublayers take the place of Layer 1. These sublayers are
needed because analog vectors, unlike binary vectors, can be arbitrarily
close together. The sublayers perform a combination of normalization and
noise suppression, in addition to the comparison of the input vector and the
expectation that is needed by the orienting subsystem.
Carpenter and Grossberg later developed the ART3 network [CaGr90],
which introduced a more sophisticated biological model for the reset mech-
anism required for ART. Up to the present time, this network has not been
widely applied.
In 1991 Carpenter, Grossberg and Reynolds introduced the ARTMAP net-
work [CaGrRe91]. In contrast with all of the previous ART networks, it is
a supervised network. The ARTMAP architecture consists of two ART mod-
ules that are connected by an inter-ART associative memory. One ART
module receives the input vector, while the other ART module receives the
desired output vector. The network learns to predict the correct output vec-
tor whenever the input vector is presented.
More recently, Carpenter, Grossberg, Markuzon, Reynolds and Rosen have
modified the ARTMAP architecture to incorporate fuzzy logic. The result is
referred to as Fuzzy ARTMAP [CaGrMa92]. It seems to improve perfor-
mance, especially with noisy input patterns. 19
All of these ART architectures incorporate the key modules discussed in
this chapter, including:
L1-L2 instars for pattern recognition.
L2-L1 outstars for pattern recall.
Layer 2 for contrast enhancement (competition).
Layer 1 for comparison of input and expectation.
Orienting Subsystem for resetting when a pattern mismatch occurs.
19-23
Summary of Results
Basic ART Architecture
Layer 1 Layer 2
Gain Control
Input
Expectation
Reset
Orienting
Subsystem
ART1 Network (Binary Patterns)

ART1 Layer 1
Input Layer 1
+ +
Expectation
W2:1 a2
+b1 S1 x S2
+
-
.
p + + n1 n1 a1
1/ f1
S1 x 1
- -
+
S1 S1
-b1 +
-W1 a2
Gain Control
S1 x S2
dn1/dt = - n1 + (+b1 - n1) { p + W2:1 a2} - (n1 + -b1) [-W1] a2
19-24
Summary of Results
Layer 1 Equation
1
dn (t) 1 + 1 1 2:1 2 1 - 1 - 1 2
-------------- = n (t) + b n (t) p + W a t n (t) + b W a t
dt
Steady State Operation

2 1
If Layer 2 is not active (i.e., each a j = 0 ), a = p .
2 1 2:1
If Layer 2 is active (i.e., one a j = 1 ), a = p w j .
ART1 Layer 2
Layer 2
+ +
On-Center +W2
W 1:2
S2 x S1
+b2 S2 x S2
+
-
. f2
a1 + + n2 n2
1/
- - a2
+
a0 -b2 +
S2
Reset -W2
Off-Surround
S2 x S2
dn2/dt = - n2 + (+b2 - n2) {[+W2] f 2(n2) + W1:2 a1}

- (n2 + -b2) [-W2] f 2(n2)
Layer 2 Equation
19
2
dn (t) 2 + 2 2 + 2 2 2 1:2 1 2 - 2 - 2 2 2
-------------- = n (t) + b n (t) W f (n (t)) + W a n (t) + b W f (n (t))
dt
1:2 T 1 1:2 T 1
ai =
0, otherwise

19-25
Orienting Subsystem
Orienting Subsystem
+b0
+
p +W0
-
1 x S1 .
+ + n0 n0 a0
1/ f0
- - Reset
+
a1 -W0 1
1 x S1 -b0 +
dn0/dt = -n0 + (+b0 - n0) [+W0] p - (n0 + -b0) [-W0] a 1
Orienting Subsystem Equation

0
dn (t) 0 + 0 0 + 0 0 - 0 - 0 1
-------------- = n (t) + b n (t) W p n (t) + b W a
dt
+ 0 - 0 + 0 - 0
where W = , W = , b = b = 1
1 2 2
0 1, if a p
a =
0, otherwise

L1-L2 Learning Law

1:2
d iw t 2 + 1:2 + 1 1:2 - - 1
------------------------- = a i t b iw t W a t iw t + b W a t
dt
1 0 1 0 0 0 1 1
0 1
b = 1 b = 0 W = 0 1
+ - +
W = 1 0
-

1 0 0 0 1 1 1 0
19-26
Summary of Results
Steady State Operation (Fast Learning)

1
1:2 a
iw - (Neuron i in Layer 2 Active)
= -----------------------------
1 2
+ a 1
L2-L1 Learning Law

2:1
d wj t 2 2:1 1
- = aj t wj t + a t
------------------------
dt
Steady State Operation (Fast Learning)

2:1 1
wj = a (Neuron j in Layer 2 Active)
ART1 Algorithm (Fast Learning) Summary

Initialization
2:1
The initial W matrix is set to all 1s.
1:2 1
Every element of the initial W matrix is set to + S 1 .
Algorithm
1. First, we present an input pattern to the network. Since Layer 2 is not
2
active on initialization (i.e., each a j = 0 ), the output of Layer 1 is
1
a = p.
2. Next, we compute the input to Layer 2,

1:2 1
W a ,
and activate the neuron in Layer 2 with the largest input: 19

1:2 T 1 1:2 T 1
2 1, if iw a = max kw a
ai = .
0, otherwise

In case of a tie, the neuron with the smallest index is declared the win-
ner.
3. We then compute the L2-L1 expectation (where we assume neuron j of
Layer 2 is activated):
2:1 2 2:1
W a = wj .
19-27
4. Now that Layer 2 is active, we adjust the Layer 1 output to include the
L2-L1 expectation:
1 2:1
a = p wj .

tween the expectation and the input pattern:
1 2 2
0 1, if a p
a = .
0, otherwise

0 2
6. If a = 1 , then we set a j = 0 , inhibit it until an adequate match occurs
0
(resonance), and return to step 1. If a = 0 , we continue with step 7.
1:2
7. Resonance has occurred, therefore we update row j of W :
1
1:2 a
jw -.
= -----------------------------
1 2
+ a 1
2:1
8. We now update column j of W :
2:1 1
wj = a .
9. We remove the input pattern, restore all inhibited neurons in Layer 2,

and return to step 1 with a new input pattern.
19-28
Solved Problems
Solved Problems
P19.1 Consider Layer 1 of the ART1 network with the following parame-
ters:
+ 1 - 1
= 0.01 b = 2 b = 3.
Assume two neurons in Layer 2, two elements in the input vector

and the following weight matrix and input:
= 0 1 p = 1 .
2:1
W
1 1 1
Also, assume that neuron 1 of Layer 2 is active.

1
i. Find and plot the response n .
ii. Check to see that the answer to part (i) satisfies the steady
state response predicted by Eq. (19.21).
i. Since Layer 2 is active, and neuron 1 of Layer 2 wins the competition,
the equations of operation of Layer 1 are
1
dn 1 1 2:1 1
0.01 -------1- = n 1 + 2 n 1 p 1 + w 1 1 n 1 + 3
dt
1 1 1 1
= n 1 + 2 n 1 1 + 0 n 1 + 3 = 3n 1 1
1
dn 1 1 2:1 1
0.01 -------2- = n 2 + 2 n 2 p 2 + w 2 1 n 2 + 3
dt
1 1 1 1
= n 2 + 2 n 2 1 + 1 n 2 + 3 = 4n 2 + 1 .
These can be simplified to obtain

1
dn 1 1
-------- = 300n 1 100 ,
dt
1
dn 2 1
-------- = 400n 2 + 100 .
dt
If we assume that both neurons start with zero initial condition, the solu-
tions are
19-29
1 1 300t
n 1 t = --- 1 e ,
3
1 1 400t
n 2 t = --- 1 e .
4
These are displayed in Figure P19.1.

0.4
1
0.2
n2 t
-0.2
1
n1 t
-0.4
0 0.005 0.01 0.015 0.02
t
Figure P19.1 Response of Layer 1

1 1
ii. Note that n1 t
converges to a negative value, and n 2 t converges to a
1 1
positive value. Therefore, a 1 t converges to 0, and a 2 t converges to 1 (re-
+
call that the transfer function for Layer 1 is hardlim ). This agrees with
our steady state analysis (see Eq. (19.21)), since
= 1 0 = 0 = a .
2:1 1
p w1 (19.75)
1 1 1
P19.2 Consider Layer 2 of the ART1 network with the following parame-
ters:
1:2 T
1w
b = 2 b = 2 = 0.5 0.5
+ 2 - 2 1:2
= 0.1 W =
2 2 1:2 T 1 0
2w
and
2 10 n 2 n0
f ( n) = .
0 n0
19-30
Solved Problems
Assume that the output of Layer 1 is
a = 1 .
1
This is equivalent to the Layer 2 example in the text (page 19-12),

with the exception of the bias values.
i. Write the equations of operation of Layer 2 and simulate and
plot the response. Explain the effect of increasing the bias
values.
ii. Verify that the steady state operation of Layer 2 is correct.
i. The equations of operation of the layer will be
2
dn 1(t) 2 2 2 2 1:2 T 1
0.1 -------------- = n 1(t) + 2 n 1(t) f (n 1(t)) + 1w a
dt
2 2 2
n 1(t) + 2 f (n 2(t)) ,
2
dn 2(t) 2 2 2 2 1:2 T 1
0.1 -------------
- = n 2(t) + 2 n 2(t) f (n 2(t)) + 2w a
dt
2 2 2
n 2(t) + 2 f (n 1(t)) .
Figure P19.2 illustrates the response of Layer 2 when the input vector is
1 T 1:2 1
a = 1 0 . The second row of W has a larger inner product with a
than the first row, therefore neuron 2 wins the competition.
2
2
1:2 T 1 n2 t
2w a
1
1:2 T 1
1w a
-1
2
n1 t
-2
0 0.05 0.1 0.15 0.2
t
Figure P19.2 Response of Layer 2
19-31
If we compare Figure P19.2 with Figure 19.5, we can see that the bias value
has three effects. First, the speed of response is increased; the neuron out-
puts move more quickly to their steady state values. Second, the range of
the response is increased from [ 1, 1] to [ 2, 2] . (Recall from Chapter 18
+
that for the shunting model the upper limit will be the excitatory bias b .
-
The lower limit will be the inhibitory bias b .) Third, the neuron responses
move closer to the upper and lower limits.
2 2
ii. At steady state, n 1(t) has a positive value, and n 2(t) has a negative val-
ue. The steady state Layer 2 output will then be
a = 1 .
2
This agrees with the desired steady state response characteristics of Layer
2:
1:2 T 1 1:2 T 1
ai = .
0, otherwise

P19.3 Consider the Orienting Subsystem of the ART1 network with the
following parameters:
+ 0 - 0
= 0.1 = 0.5 = 2 ( = 0.25 ) b = b = 0.5 .
The inputs to the Orienting Subsystem are
1 1
1
p = 1 a = 0 .
1 1
0
i. Find and plot the response of the Orienting Subsystem n t .
ii. Verify that the steady state conditions are satisfied.
The equation of operation of the Orienting Subsystem is
0
dn (t) 0 0
0.1 -------------- = n (t) + 0.5 n (t) 0.5 p 1 + p 2 + p 3
dt
0 1 1 1
n (t) + 0.5 2 a 1 + a 2 + a 3
or
19-32
Solved Problems
0
dn (t) 0
-------------- = 65n (t) 12.5 .
dt
The response is then
0 65t
n t = 0.1923 1 e
0
This response is plotted in Figure P19.3. In this case, since n t is nega-
0 + 0
tive, a = hardlim n = 0 , and therefore a reset signal will not be sent to
Layer 2.
0.3
0.15
0
-0.15 n t
-0.3
0 0.05 0.1 0.15 0.2
Figure P19.3 Response of the Orienting Subsystem

ii. The steady state operation of the Orienting Subsystem can be summa-
rized as follows:
1 2 2
0 1, if a p
a = .
0, otherwise

For this problem

2 2
1 1
1 2 2 2
a p = 0 1 = --- = 0.25 .
3
1 1
0
Therefore a = 0 , which agrees with the results of part (i).
19-33
P19.4 Show that the learning equation for the L2-L1 connections is
equivalent to the outstar equation described in Chapter 15.
The L2-L1 learning law (Eq. (19.65)) is
2:1
d wj t 2 2:1 1
- = aj t wj t + a t .
------------------------
dt

2:1 2:1 2:1
d w j (t) w j (t + t) w j (t)
----------------------- ------------------------------------------------- ,
dt t
then we can rewrite Eq. (19.65) as

2:1 2:1 2 2:1 1
w j (t + t) = w j (t) + t a j (t) w j (t) + a (t) .
This is equivalent to the outstar rule of Chapter 15 (Eq. (15.51)). Here the
2
input to the L2-L1 connections is a j (t) , and the output of the L2-L1 connec-
1
tions is a .
P19.5 Train an ART1 network using the following input vectors:
0 1 1
p1 = 1 , p2 = 0 , p3 = 1 .
0 0 0
2
Use the parameters = 2 , and = 0.4 , and choose S = 3 (3 cate-
gories).
Our initial weights will be
1 1 1 0.5 0.5 0.5

2:1 1:2
W = 1 1 1 , W = 0.5 0.5 0.5 .
1 1 1 0.5 0.5 0.5
We now begin the algorithm.

1. Compute the Layer 1 response:
0
1
a = p1 = 1 .
0
19-34
Solved Problems
2. Next, compute the input to Layer 2:
0.5 0.5 0.5 0 0.5

1:2 1
W a = 0.5 0.5 0.5 1 = 0.5 .
0.5 0.5 0.5 0 0.5
Since all neurons have the same input, pick the first neuron as the win-
ner. (In case of a tie, pick the neuron with the smallest index.)
1
2
a = 0
0
3. Now compute the L2-L1 expectation:
11 1 1 1
2:1 2 2:1
W a = 1 1 1 0 = 1 = w1 .
11 1 0 1
4. Adjust the Layer 1 output to include the L2-L1 expectation:
0 1 0
1 2:1
a = p1 w1 = 1 1 = 1 .
0 1 0

1 2 2 1 0
a p1 = --- = 0.4 , therefore a = 0 (no reset).
0
1
19
6. Since a = 0 , continue with step 7.
1:2
7. Resonance has occurred, therefore update row 1 of W :
1 0 0 1 0
1:2 2a 1 1:2
1w = -----------------------------
- = a = 1 , W = 0.5 0.5 0.5 .
1 2
2+ a 1
0 0.5 0.5 0.5
2:1
8. Update column 1 of W :
19-35
0 0 1 1
2:1 1 2:1
w1 = a = 1 ,W = 1 1 1 .
0 0 1 1
9. Remove p 1 , and return to step 1 with input pattern p 2 .

1. Compute the new Layer 1 response (Layer 2 inactive):
1
1
a = p2 = 0 .
0
0 1 0 1 0
1:2 1
W a = 0.5 0.5 0.5 0 = 0.5 .
0.5 0.5 0.5 0 0.5
Since neurons 2 and 3 have the same input, pick the second neuron as
the winner:
0
2
a = 1 . (19.76)
0
01 1 0 1
2:1 2 2:1
W a = 1 1 1 1 = w2 = 1 .
01 1 0 1
1 1 1
1 2:1
a = p2 w2 = 0 1 = 0 .
0 1 0

1 2 2 1 0
1
19-36
Solved Problems
0
1:2
1 1 0 1 0
1:2 2a 1 1:2
2w = -----------------------------
- = a = 0 , W = 1 0 0 .
1 2
2+ a 1
0 0.5 0.5 0.5
2:1
1 0 1 1
2:1 1 2:1
w2 = a = 0 ,W = 1 0 1 .
0 0 0 1
9. Remove p 2 , and return to step 1 with input pattern p 3 .

1. Compute the Layer 1 response with the new input vector:
1
1
a = p3 = 1 .
0
0 1 0 1 1
1:2 1
W a = 1 0 0 1 = 1 .
0.5 0.5 0.5 0 1
ner:
19
1
2
a = 0 .
0
01 1 1 0
2:1 2 2:1
W a = 1 0 1 0 = w1 = 1 .
00 1 0 0
19-37
1 0 0
1 2:1
a = p3 w1 = 1 1 = 1 .
0 0 0

1 2 2 1 0
2
0
1:2
1 0 0 1 0
1:2 2a 1 1:2
1w = -----------------------------
- = a = 1 , W = 1 0 0 .
1 2
2+ a 1
0 0.5 0.5 0.5
2:1
0 0 1 1
2:1 1 2:1
w2 = a = 1 ,W = 1 0 1 .
0 0 0 1
This completes the training, since if you apply any of the three patterns
again they will not change the weights. These patterns have been success-
fully clustered. This type of result (stable learning) is guaranteed for the
ART1 algorithm, since it has been proven to always produce stable clus-
ters.
P19.6 Repeat Problem P19.5, but change the vigilance parameter to

= 0.6 .
The training will proceed exactly as in Problem P19.5, until pattern p 3 is

presented, so lets pick up the algorithm at that point.
1. Compute the Layer 1 response:
1
1
a = p3 = 1 .
0
19-38
Solved Problems
0 1 0 1 1
1:2 1
W a = 1 0 0 1 = 1 .
0.5 0.5 0.5 0 1
ner:
1
2
a = 0 .
0
01 1 1 0
2:1 2 2:1
W a = 1 0 1 0 = w1 = 1 .
00 1 0 0
1 0 0
1 2:1
a = p3 w1 = 1 1 = 1 .
0 0 0

1 2 2 1 0
a p3 = --- = 0.6 , therefore a = 1 (reset).
2
0 2
6. Since a = 1 , set a 1 = 0 , inhibit it until an adequate match occurs
(resonance), and return to step 1.
19
1. Recompute the Layer 1 response (Layer 2 inactive):
1
1
a = p3 = 1 .
0
19-39
0 1 0 1 1
1:2 1
W a = 1 0 0 1 = 1 .
0.5 0.5 0.5 0 1
Since neuron 1 is inhibited, neuron 2 is the winner:
0
2
a = 1 .
0
01 1 0 1
2:1 2 2:1
W a = 1 0 1 1 = w2 = 0 .
00 1 0 0
1 1 1
1 2:1
a = p3 w2 = 1 0 = 0 .
0 0 0

1 2 2 1 0
a p3 = --- = 0.6 , therefore a = 1 (reset).
2
0 2
6. Since a = 1 , set a 2 = 0 , inhibit it until an adequate match occurs
(resonance), and return to step 1.
1. Recompute the Layer 1 response:
1
1
a = p3 = 1 .
0
0 1 0 1 1
1:2 1
W a = 1 0 0 1 = 1 .
0.5 0.5 0.5 0 1
19-40
Solved Problems
Since neurons 1 and 2 are inhibited, neuron 3 is the winner:
0
2
a = 0 .
1
01 1 0 1
2:1 2 2:1
W a = 1 0 1 0 = w3 = 1 .
00 1 1 1
1 1 1
1 2:1
a = p3 w3 = 1 1 = 1 .
0 1 0

1 2 2 2 0
2
0
1:2
2
--- 0 1 0
1 3
2a 2 1
3w
1:2
2+ a 1
1 2
- = --- a = 2 , W = 1 0 0 .
= -----------------------------
3 ---
3
1:2
2 2
--- --- 0
19
3 3
0
2:1
1 0 1 1
2:1 1 2:1
w3 = a = 1 ,W = 1 0 1 .
0 0 0 0
This completes the training, since if you apply any of the three patterns
again they will not change the weights. (Verify this for yourself by applying
19-41
each input pattern to the network.) These patterns have been successfully
clustered.
Note that in Problem P19.5, where the vigilance was = 0.4 , the patterns
were clustered into two categories. In this problem, with vigilance = 0.6 ,
the patterns were clustered into three categories. The closer the vigilance
is to 1, the more categories will be used. This is because an input pattern
must be closer to a prototype in order to be incorporated into that proto-
type. When the vigilance is close to zero, many different input patterns can
be incorporated into one prototype. The vigilance parameter adjusts the
coarseness of the categorization.
P19.7 Train an ART1 network using the following input vectors (see
[CaGr87a]):
p1 p2 p3 p4
Present the vectors in the order p 1 p 2 p 3 p 1 p 4 (i.e., p 1 is pre-

sented twice in each epoch). Use the parameters = 2 and = 0.6 ,
2
and choose S = 3 (3 categories). Train the network until the
weights have converged.
2:1
We begin by initializing the weight matrices. The initial W matrix is an
1 2 1:2
S S = 25 3 matrix of 1s. The initial W matrix is normalized, there-
2 1
fore it is an S S = 3 25 matrix, with each element equal to
2
- = ----------------------------- = 0.0769 .
---------------------------
+ S 1
1 2 + 25 1
To create the input vectors we will scan each pattern row-by-row, where
each blue square will be represented by a 1 and each white square will be
represented by a 0. Since the input patterns are 5 5 grids, this will create
25-dimensional input vectors.
We now begin the training. Since it is not practical to display all of the cal-
culations when the vectors are so large, we have summarized the results of
the algorithm in Figure P19.4. Each row represents one iteration of the
ART1 algorithm (presentation of one input vector). The left-most pattern
in each row is the input vector. The remainder of the patterns represent the
2:1
three columns of the W matrix. At each iteration, a star indicates the
2:1
resonance point the column of W that matched with the input pattern.
Whenever a reset occurred, it is represented by a check mark. When more
than one reset occurred in a given iteration, the number beside the check
mark indicates the order in which the reset occurred.
19-42
Solved Problems
p w12:1 w22:1 w32:1 p w12:1 w22:1 w32:1
1 6
1 2
2 7
3 8
1
4 9
5 10
(a) First Epoch (b) Second Epoch
Figure P19.4 ART1 Iterations for Problem

A total of 10 iterations of the algorithm were performed (two epochs of the
sequence p 1 p 2 p 3 p 1 p 4 ). The weights are now stable. (You may want
to check this by presenting each input pattern.)
There are several interesting points to notice in this example. First, notice
2:1
that at iteration 4 both p 1 and p 3 are coded by w 2 . However, on iteration
2:1 2:1
5, when p 4 is presented, w 2 is modified to include p 4 . This new w 2 no
longer provides an adequate match with p 1 and p 3 , as we can see at itera-
tions 6 and 8. This requires them to take over neuron 3 , which was unused
during the first epoch.
The results of the algorithm could be modified by changing the vigilance
parameter. How small would you have to make the vigilance, so that only
two neurons in Layer 2 would be required to code all 4 input vectors? How
large would the vigilance have to be before a fourth Layer 2 neuron was
needed?
19
19-43
Epilogue
Competitive learning, and many other types of neural network training al-
gorithms, suffer from a problem called the stability/plasticity dilemma. If a
learning algorithm is sensitive to new inputs (plastic), then it runs the risk
of forgetting prior learning (unstable). The ART networks were designed to
achieve learning stability while maintaining sensitivity to novel inputs.
In this chapter, the ART1 network was used to illustrate the key concepts
of adaptive resonance theory. The ART1 network is based on the Grossberg
competitive network of Chapter 18, with a few modifications. The key inno-
vation of ART is the use of expectations. As each input pattern is present-
ed to the network, it is compared with the prototype vector that it most
closely matches (the expectation). If the match between the prototype and
the input vector is not adequate, a new prototype is selected. In this way,
previously learned memories (prototypes) are not eroded by new learning.
One important point to keep in mind when analyzing ART networks, is
that they were designed to be biologically plausible mechanisms for learn-
ing. They have as much to do with understanding how the brain works as
they do with inspiring practical pattern recognition systems. For this rea-
son, the learning mechanisms are required to use only local information at
each neuron. This is not true of all of the learning rules discussed in this
text.
Although the ART networks solve the problem of learning instability, in
which the network weights never stabilize, there is another stability prob-
lem that we have not yet discussed. This is the stability of the differential
equations that implement the short-term memory equations of the net-
work. In Layer 2, for example, we have a set of differential equations with
nonlinear feedback. Can we make some general statement about the sta-
bility of such systems? Chapter 20 will present a comprehensive discussion
of this problem.
19-44
Further Reading
Further Reading
[CaGr87a] G. A. Carpenter and S. Grossberg, A massively parallel ar-
chitecture for a self-organizing neural pattern recognition
machine, Computer Vision, Graphics, and Image Process-
ing, vol. 37, pp. 54115, 1987.
In this original presentation of the ART1 architecture, Car-
penter and Grossberg demonstrate that the architecture
self-organizes and self-stabilizes in response to an arbi-
trary number of binary input patterns. The key feature of
ART is a top-down matching mechanism.
[CaGr87b] G. A. Carpenter and S. Grossberg, ART2: Self-organiza-
tion of stable category recognition codes for analog input
patterns, Applied Optics, vol. 26, no. 23, pp. 49194930,
1987.
This article describes an extension of the ART1 architec-
ture that is designed to handle analog input patterns.
[CaGr90] G. A. Carpenter and S. Grossberg, ART3: Hierarchical
search using chemical transmitters in self-organizing pat-
tern recognition architectures, Neural Networks, vol. 3,
no. 23, pp. 129152, 1990.
This article demonstrates how the Orienting Subsystem of
the ART networks could be implemented in biological neu-
rons through the use of chemical transmitters.
[CaGrMa92] G. A. Carpenter, S. Grossberg, N. Markuzon, J. Reynolds
and D. Rosen, Fuzzy ARTMAP: An adaptive resonance ar-
chitecture for incremental learning of analog maps, Pro-
ceedings of the International Joint Conference on Neural
Networks, Baltimore, MD, vol. 3, no. 5, pp. 309314, 1992. 19
The authors present a modification of the ARTMAP archi-
tecture to include fuzzy logic that enables better perfor-
mance in a noisy environment.
[CaGrRe91] G. A. Carpenter, S. Grossberg and J. Reynolds, ARTMAP:
Supervised real-time learning and classification of nonsta-
tionary data by a self-organizing neural network, Neural
Networks, vol. 4, no. 5, pp. 169181, 1991.
This article presents an adaptive resonance theory net-
work for supervised learning. The network consists of two
interconnected ART modules. One module receives the in-
put vector, and the other module receives the desired out-
put vector.
19-45

134, 1976.
Grossberg describes a continuous-time competitive net-
work, inspired by the developmental physiology of the visu-
al cortex. The structure of this network forms the
foundation for other important networks.
mental concepts used in later Grossberg networks, such as
the adaptive resonance theory networks.
19-46
Exercises
Exercises
E19.1 Consider Layer 1 of the ART1 network with = 0.02 . Assume two neurons
in Layer 2, two elements in the input vector and the following weight ma-
trix and input:
= 0 1 p = 0 .
2:1
W
1 1 1
Also assume that neuron 2 of Layer 2 is active.

1 + 1 - 1
i. Find and plot the response n if b = 2 and b = 3 .
1 + 1 - 1
ii. Find and plot the response n if b = 4 and b = 5 .
1 + 1 - 1
iii. Find and plot the response n if b = 4 and b = 4 .
iv. Check to see that the answers to parts (i)(iii) satisfy the steady
state response predicted by Eq. (19.21). Explain any inconsisten-
cies.
2+2
ans = v. Check your answers to parts (i)(iii) by writing a MATLAB M-file to
4 simulate Layer 1 of the ART1 network. Use the ode45 routine. Plot
the response for each case.
E19.2 Consider Layer 2 of the ART1 network with the following parameters:
1:2 T 2 2
1:2 1w --- ---
= 0.1 W = = 3 3
1:2 T
2w 1 0
and
2 10 n 2 , n0
f (n) = .
0, n0
Assume that the output of Layer 1 is
a = 1 .
1
1
2+2
ans = i. Write the equations of operation of Layer 2, and simulate and plot
4 the response if the following bias vectors are used:
19-47
b = 2 b = 2 .
+ 2 - 2
2 2
ii. Repeat part (i) for the following bias vectors:
b = 3 b = 3 .
+ 2 - 2
3 3
iii. Repeat part (i) for the following bias vectors:
b = 3 b = 0 .
+ 2 - 2
3 0
iv. Do the results of all of the previous parts satisfy the desired steady
state response described in Eq. (19.37)? If not, explain why.
E19.3 Consider the Orienting Subsystem of the ART1 network with the following
parameters:
+ 0 - 0
= 0.1 b = b = 2.
The inputs to the Orienting Subsystem are
1 0
1
p = 1 a = 0 .
1 1
0
i. Find and plot the response of the Orienting Subsystem n t , for
= 0.5 = 4 ( = 0.125 ).
0
ii. Find and plot the response of the Orienting Subsystem n t , for
= 0.5 = 2 ( = 0.25 ).
iii. Verify that the steady state conditions are satisfied in parts (i) and
(ii).
2+2
ans = iv. Check your answers to parts (i) and (ii) by writing a MATLAB M-file
4 to simulate the Orienting Subsystem.
19-48
Exercises
E19.4 To derive the steady state conditions for the L1-L2 and L2-L1 learning
rules, we have made the assumption that the input pattern and the neuron
outputs remain constant until the weight matrices converge. This is called
fast learning. Show that this fast learning assumption is equivalent to
setting the learning rate to 1 in the instar and outstar learning rules pre-
sented in Chapter 15 and the Kohonen competitive learning rule in Chap-
ter 16.
E19.5 Train an ART1 network using the following input vectors:
0 1 1 1
p1 = 1 , p2 = 0 , p3 = 1 , p3 = 1 .
0 0 0 1
1 1 0 1
2
Use the parameter = 2 , and choose S = 3 (3 categories).
i. Train the network to convergence using = 0.3 .
ii. Repeat part (i) using = 0.6 .
iii. Repeat part (ii) using = 0.9 .
E19.6 The ART1 algorithm can be modified to add a new neuron in Layer 2 when-
ever there is no adequate match between the existing prototypes and the
1:2
input pattern. This involves creating a new row of the W matrix and a
2:1
new column of the W matrix. Describe how this would be done.
E19.7 Write a Matlab M-file to implement the ART1 algorithm (with the modifi-
2+2
cation described in Exercise E19.6). Use this M-file to train an ART1 net-
ans = work using the following input vectors (see Problem P19.7):
4
p1 p2 p3 p4
Present the vectors in the order p 1 p 2 p 3 p 1 p 4 (i.e., p 1 is presented

twice in each epoch). Use the parameters = 2 and = 0.9 , and choose
2
S = 3 (3 categories). Train the network until the weights have converged.
Compare your results with Problem P19.7.
19-49
E19.8 Recall the digit recognition problem described in Chapter 7 (page 7-10).
Train an ART1 network using the digits 0 9 , as displayed below:
2+2
ans =
4
p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
2
Use the parameter = 2 , and choose S = 5 (5 categories). Use the Mat-
lab M-file from Exercise E19.7.
i. Train the network to convergence using = 0.3 .
ii. Train the network to convergence using = 0.6 .
iii. Train the network to convergence using = 0.9 .
iv. Discuss the results of parts (i)(iii). Explain the effect of the vigi-
lance parameter.
19-50
Objectives
20 Stability
Objectives 20-1
Stability Concepts 20-3
Definitions 20-4
Lyapunov Stability Theorem 20-5
Pendulum Example 20-6
LaSalles Invariance Theorem 20-12
Definitions 20-12
Theorem 20-13
Example 20-14
Comments 20-18
Epilogue 20-28
Exercises 20-30
Objectives
The problem of convergence in a recurrent network was first raised in our
discussion of the Hopfield network, in Chapter 3. It was noted there that
the output of a recurrent network could converge to a stable point, oscillate,
or perhaps even diverge. The stability of the steepest descent process and
of the LMS algorithm were discussed in Chapter 9 and Chapter 10, respec-
tively. The stability of Grossbergs continuous-time recurrent networks
was discussed in Chapter 18.
In this chapter we will define stability more carefully. Our objective is to
determine whether a particular set of nonlinear equations has points (or
trajectories) to which its output might converge. To help us study this topic
we will introduce Lyapunovs Stability Theorem and apply it to a simple,
but instructive, problem. Then, we will present a generalization of the
Lyapunov Theory: LaSalles Invariance Theorem. This will set the stage for
Chapter 21, where LaSalles theorem is used to prove the stability of
Hopfield networks.
20-1
20 Stability
Theory and Examples
Recurrent Networks
We first discussed recurrent neural networks, which have feedback connec-
tions from their outputs to their inputs, when we introduced the Hamming
and Hopfield networks in Chapter 3. The Grossberg networks of Chapter
18 and Chapter 19 also contain recurrent connections. Recurrent networks
are potentially more powerful than feedforward networks, since they are
able to recognize and recall temporal, as well as spatial, patterns. However,
the behavior of these recurrent networks is much more complex than that
of feedforward networks.
For feedforward networks, the output is constant (for a fixed input) and is
a function only of the network input. For recurrent networks, however, the
output of the network is a function of time. For a given input and a given
initial network output, the response of the network may converge to a sta-
ble output. However, it may also oscillate, explode to infinity, or follow a
chaotic pattern. In the remainder of this chapter we want to investigate
general nonlinear recurrent networks, in order to determine their long-
term behavior.
We will consider recurrent networks that can be described by nonlinear dif-
ferential equations of the form:
d
a t = g a t ,p t ,t . (20.1)
dt
Here p t is the input to the network, and a t is the output of the network.
(See Figure 20.1.)
Nonlinear Recurrent Network
p .
a a
g
a(0)
da(t)/dt = g (a(t), p(t), t)
Figure 20.1 Nonlinear, Continuous-Time, Recurrent Network
20-2
Stability Concepts
We want to know how these systems perform in the steady state. We will
be most interested in those cases where the network converges to a con-
stant output a stable equilibrium point. A nonlinear system can have
many stable points. For some neural networks these stable points repre-
sent stored prototype patterns. When possible, we would like to know
where the stable points are, and which initial conditions a 0 converge to
a given stable point (i.e., what is the basin of attraction for a given stable
point?).
Stability Concepts
To begin our discussion, lets introduce some basic stability concepts with
a simple, intuitive example. Consider the motion of a ball bearing, with dis-
sipative friction, in a gravity field. In the adjacent figure, we have a ball
bearing at the bottom of a trough (point a ). If we move the bearing to a
different position, it will oscillate back and forth in the trough, but, because
of friction, it will eventually settle back to the bottom of the trough. We will
call this position an asymptotically stable point, which we will define more
precisely in the next section.
Consider now the second figure in the left margin. Here we have a ball
bearing positioned at the center of a flat surface. If we place the bearing in
a different position, it will not move. The position at the center of the sur-
face is not asymptotically stable, since the bearing does not move back to
the center if it is moved away. However, it is stable in a certain sense, be-
cause at least the ball does not roll farther away from the center point. We
call this kind of point stable in the sense of Lyapunov, which we will define
in the next section.
Now consider the third figure in the left margin. The ball bearing is posi-
tioned at the top of a hill. This is an equilibrium point, since the ball will
remain at the top of the hill, if we position it carefully. However, if the bear-
ing is given the slightest disturbance, it will roll down the hill. This is an
unstable equilibrium point.
In the next chapter we will try to design Hopfield neural networks, in
which the stored prototype patterns will be asymptotically stable equilibri-
um points. We would also like the basins of attraction for these stable
points to be as large as possible.
For example, consider Figure 20.2. We would like to design neural net-
works with large basins of attraction such as those of Case A. One can cer-
tainly imagine that if a ball that rolls with high friction is placed (with zero
velocity) in any one of the basins of Case A, it will remain in that basin and
will eventually find its way to the bottom (stable point). However, Case B
is more complicated. If, for instance, one places a ball with friction at point
P, it is not clear which stable point will eventually capture the ball. The ball
may not come to rest at the stable point closest to P. It is also difficult to
tell how large the basin of attraction is for a specific stable point.
20-3
20 Stability
Case A Large Basin of Attraction
Case B
Complex Region of Attraction
P
Figure 20.2 Basins of Attraction

Now that we have presented some intuitive notions of stability, we will pur-
sue them with mathematical rigor in the remainder of this chapter.
Definitions
We will begin with specific mathematical definitions of the different types
of stability discussed in the previous section. In these definitions we will be
Equilibrium Point talking about the stability of an equilibrium point; a point a where the de-
rivative in Eq. (20.1) is zero. For simplicity, we will talk specifically about
the point a = 0 , which is referred to as the origin. This restriction does
not affect the generality of our discussion.
Stability Definition 1: Stability (in the sense of Lyapunov)
The origin is a stable equilibrium point if for any given value 0 there
exists a number 0 such that if a 0 , then the resulting motion
a t satisfies a t for t 0 .
This definition says that the system output is not going to move too far
away from a given stable point, so long as it is initially close to the stable
point. Lets say that you want the system output to remain within a dis-
tance of the origin. If the origin is stable, then you can always find a dis-
tance (which may be a function of ), such that if the system output is
within of the origin at time t = 0 , then it will always remain within of
the origin. The position of the ball (with zero velocity) in the figure to the
left is stable in the sense of Lyapunov, so long as the ball bearing has fric-
tion. If the ball bearing did not have friction, then any initial velocity would
produce a trajectory a t in which the position would go to infinity. (The
vector a t in this case would consist of the position and the velocity of the
ball.)
Next, lets consider the stronger concept of asymptotic stability.
20-4
Lyapunov Stability Theorem
Asymptotic Stability Definition 2: Asymptotic Stability

The origin is an asymptotically stable equilibrium point if there exists a
number 0 such that whenever a 0 the resulting motion satisfies
a t 0 as t .
This is a stronger definition of stability. It says that as long as the output

of the system is initially within some distance of the stable point, the out-
put will eventually converge to the stable point. The position of the ball
(with zero velocity) in the diagram in the left margin is asymptotically sta-
ble, so long as the ball bearing has friction. If there is no friction, the posi-
tion is only stable in the sense of Lyapunov.
We would like to build neural networks that have many specified asymp-
totically stable points, each of which represents a prototype pattern. This
is the design objective we will use for building Hopfield networks in Chap-
ter 21.
In addition to the stability definitions, there is another concept we will use
in analyzing stability. It is the concept of a definite function. The next two
definitions will clarify this concept.
Positive Definite Definition 3: Positive Definite
A scalar function V a is positive definite if V 0 = 0 and V a 0 for
a0.
Positive Semidefinite Definition 4: Positive Semidefinite

A scalar function V a is positive semidefinite if V a 0 for all a .
(These definitions can be modified appropriately to define the concepts neg-

ative definite and negative semidefinite.) Now that we have defined stabili-
ty, lets consider a method for testing stability.

One of the most important approaches for investigating the stability of
nonlinear systems is the theory introduced by Alexandr Mikhailovich
Lyapunov, a Russian mathematician. Although his major work was first
20
published in 1892, it received little attention outside Russia until much lat-
er. In this section we will discuss one of Lyapunovs most powerful tech-
niques for stability analysis the so-called direct method.
Consider the autonomous (unforced, no explicit time dependence) system:
da
= ga . (20.2)
dt
20-5
20 Stability
The Lyapunov stability theorem can then be stated as follows.

Theorem 1: Lyapunov Stability Theorem
If a positive definite function V a can be found such that dV a dt is neg-
ative semidefinite, then the origin ( a = 0 ) is stable for the system of Eq.
(20.2). If a positive definite function V a can be found such that dV a dt
is negative definite, then the origin ( a = 0 ) is asymptotically stable. In
each case, V is called a Lyapunov function of the system.
You can think of V a as a generalized energy function. The concept of the

theorem is that if the energy of a system is continually decreasing
( dV a dt negative definite), then it will eventually settle at some mini-
mum energy state. Lyapunovs insight was to generalize the concept of en-
ergy, so that the theorem could be applied to systems where the energy is
difficult to express or has no meaning.
We should note that the theorem only states that if a suitable Lyapunov
function V a can be found, the system is stable. It gives us no information
about the stability of the system in those situations where we are unable to
find such a function.
Pendulum Example
We can gain some insight into Lyapunovs stability theorem by applying it
to a simple mechanical system. This system is very simple, and its opera-
tion is easy to visualize, and yet it illustrates important concepts that we
will apply to neural network design in the next chapter. The example sys-
tem we will use is the pendulum shown in Figure 20.3.
l

m
mg
Figure 20.3 Pendulum
Using Newtons second law ( F = ma ), we can write the equation of opera-
tion of the pendulum as
2
d d
ml 2
() = c mg sin , (20.3)
dt dt
20-6
Pendulum Example
or
2
d d
ml 2
+c + mg sin = 0 , (20.4)
dt dt
where is the angle of the pendulum, m is the mass of the pendulum, l is

the length of the pendulum, c is the damping coefficient, and g is the grav-
itational constant.
The first term on the right side of Eq. (20.3) is the damping force, which is
proportional to the velocity of the pendulum. It is this term that represents
the energy dissipation in the system. The second term on the right side of
Eq. (20.3) is the gravitational force, which is proportional to the sine of the
angle of the pendulum. It is equal to zero when the pendulum is straight
down and has its maximum value when the pendulum is horizontal.
If the damping coefficient is not zero, the pendulum will eventually come to
rest hanging down in the vertical position. This solution might be viewed
as = 0 , but more generally it is = 2n , where n = 0 1 2 3 ... .
That is, given the appropriate initial conditions, the pendulum might sim-
ply settle to = 0 or it might rotate once to give a solution of = 2 , etc.
There are many possible equilibrium solutions. (The positions = n , for
odd values of n , are also equilibrium points, but they are not stable.)
To analyze the stability of this system, we will write the pendulum equa-
tion in state variable form, where it will appear as a pair of first-order dif-
ferential equations. Lets choose the following state variables:
d
a 1 = and a 2 = . (20.5)
dt
We can write equations for the pendulum in terms of these state variables
as follows:
da 1
= a2 , (20.6)
dt
da 2 g c
= --- sin a 1 ------ a 2 . (20.7)
dt l ml
Now we want to investigate the stability of the origin ( a = 0 ) for this pen-
dulum system. (The origin corresponds to a pendulum angle of zero and a
pendulum velocity of zero.) We first want to check that the origin is an equi-
librium point. We do this by substituting a = 0 into the state equations.
da 1
= a2 = 0 , (20.8)
dt
20-7
20 Stability
da 2 g c g c
= --- sin a 1 ------ a 2 = --- sin 0 ------ 0 = 0 (20.9)
dt l ml l ml
Since the derivatives are zero, the origin is an equilibrium point.

Next we need to find a Lyapunov function for the pendulum. For this ex-
ample we will use the energy of the system as the Lyapunov function V . To
obtain the total energy of the pendulum, we add the kinetic and potential
energies.
1 2 2
V a = --- ml a 2 + mgl 1 cos a 1 (20.10)
2
In order to test the stability of the system, we need to evaluate the deriva-
tive of V with respect to time.
d T V da 1 V da 2
V a = V a g a = + (20.11)
dt a1 d t a2 d t
The partial derivatives of V a can be obtained from Eq. (20.10), and the
derivatives of the two state variables are given by Eq. (20.6) and Eq. (20.7).
Thus we have
V a = mgl sin a 1 a 2 + ml a 2 --- sin a 1 ------ a 2 .

d 2 g c
(20.12)
dt l ml
The mgl sin a 1 a 2 terms cancel, which leaves only
d 2
V a = cl a 2 0 . (20.13)
dt
In order to prove that the origin ( a = 0 ) is asymptotically stable, we must

show that this derivative is negative definite. The derivative is zero at the
origin, but it also is zero for any value of a 1 , as long as a 2 = 0 . Thus,
dV a dt is negative semidefinite, rather than negative definite. From
Lyapunovs theorem, then, we know that the origin is a stable point. How-
ever, we cannot say, from the theorem and this Lyapunov function, that the
origin is asymptotically stable.
In this case we know that as long as the pendulum has friction, it will even-
tually settle in a vertical position, and, therefore, that the origin is asymp-
totically stable. However, Lyapunovs theorem, using our Lyapunov
function, can only tell us that the origin is stable. To prove that the origin
is asymptotically stable, we will need a refinement of Lyapunovs theorem,
LaSalles Invariance Theorem. We will discuss LaSalles theorem in the
next section.
20-8
Pendulum Example
First, lets investigate the pendulum further, by taking a specific numerical

2
+2 example. Let g = 9.8 m = 1 l = 9.8 c = 1.96 . Now we can rewrite the
state equations for the pendulum as
da 1
= a2 , (20.14)
dt
da 2
= sin a 1 0.2a 2 . (20.15)
dt
Expressions for V and its derivative follow:
2 1 2
V = 9.8 --- a 2 + 1 cos a 1 , (20.16)
2
dV 2
= 19.208 a 2 . (20.17)
dt
Note that dV dt is zero for any value of a 1 as long as a 2 = 0 .

Figure 20.4 displays the 3-D and contour plots of the energy surface, V , as
the angle varies between -10 and +10 radians and the angular velocity var-
ies between -2 and 2 radians per second. Note that in this range there are
three possible minimum points of the energy surface, at 0 and 2 .
400
1
300
200
a2 0
100
-1
0
2
1 10
0 5 -2
a2 0
-1 a1
-5
-2 -10 -10 -5 0 5 10
a1
Figure 20.4 Pendulum Energy Surface

(We will find in Chapter 21 that the minimum points of the Lyapunov func-
tion can correspond to prototype patterns in an autoassociative neural net-
work. The pendulum system, like recurrent neural networks, has many
minimum points.)
Of course, the energy plots shown in Figure 20.4 do not tell us in what way,
or by what route, the pendulum finds a particular energy minimum. To
20-9
20 Stability
show this, we have plotted the energy contours, and one particular path for
the pendulum, in Figure 20.5. The response trajectory, shown by the blue
line, starts from an initial position, a 1 0 , of 1.3 radians (74) and an initial
velocity, a 2 0 , of 1.3 radians per second. The trajectory converges to the
equilibrium point a = 0 .
Contour Plot
a2 0
x2
-1
-2
-10 -5 0 5 10
ax11
Figure 20.5 Pendulum Response on State Variable Plane

A time response plot of the two state variables is shown in Figure 20.6. No-
tice that, because the initial velocity is positive, the pendulum continues to
move up initially. (Check to see if this agrees with Figure 20.5.) It reaches
a maximum angle of about 2 radians before falling back down. The oscilla-
tions continue to decay as both state variables converge to zero.
In this case, both state variables converge to zero. However, this is not the
only possible equilibrium point, as we will show later.
It is also interesting to plot the pendulum energy, V , as in Figure 20.7. Re-
call from Eq. (20.17) that the energy should never increase; this is consis-
tent with Figure 20.7. Eq. (20.17) also predicts that the derivative of the
energy curve should only be zero when the velocity, a 2 , is zero. This is also
verified if we compare Figure 20.7 with Figure 20.6. At those times where
the a 2 graph crosses the zero axis, the slope of the energy curve is zero.
20-10
Pendulum Example
2 a1
-1
a2
-2
0 10 20 30 40
Figure 20.6 State Variables a 1 (blue) and a 2 vs. Time
Notice that, although there are points where the derivative of the energy
curve is zero, the derivative does not remain zero until the energy is also
zero. This observation will lead to LaSalles Invariance Theorem, which we
will discuss in the next section. The key idea of that theorem is to identify
those points where the derivative of the Lyapunov function is zero, and
then to determine if the system will be trapped at those points. (Those plac-
es where a trajectory can be trapped are called invariant sets.) If the only
point that can trap the trajectory, and that has zero derivative, is the ori-
gin, then the origin is asymptotically stable.
160
120
80
Vt
40
0 10 20
t
30 40
20
Figure 20.7 Pendulum Lyapunov Function (Energy) vs. Time
The particular pendulum behavior shown in the graphs in this section de-
pends on the initial conditions of the two state variables. The choice of a dif-
ferent set of initial conditions may give results entirely different from those
shown in these plots. We will expand on this in the next section.
20-11
20 Stability
To experiment with the pendulum, use the Neural Network Design Demon-
stration Dynamic System (nnd17ds).
LaSalles Invariance Theorem

The pendulum example demonstrated a problem with Lyapunovs theorem.
We found a Lyapunov function whose derivative was only negative
semidefinite (not negative definite), and yet we know that the origin is as-
ymptotically stable for the pendulum system. In this section we will intro-
duce a theorem that clarifies this uncertainty in Lyapunovs theorem. It
does so by defining those regions of the state space where the derivative of
the Lyapunov function is zero, and then identifying those parts of that re-
gion that can trap the trajectory.
Before we discuss LaSalles Invariance Theorem, we first need to introduce
the following definitions.
Definitions
Lyapunov Function Definition 5: Lyapunov Function
n
Let V be a continuously differentiable function from to . If G is any
n
subset of , we say that V is a Lyapunov function on G for the system
da dt = g a if
dV a T
--------------- = V a g a (20.18)
dt
does not change sign on G .
This is a generalization of our previous definition of the Lyapunov function,

which we used in Theorem 1. Here we do not require that the function be
positive definite. In fact, there is no direct requirement on the function it-
self (except that it be continuously differentiable). The only requirement is
on the derivative of V . The derivative cannot change sign anywhere on the
set G . Note that the derivative will not change sign if it is negative
semidefinite or if it is positive semidefinite.
We should note here that we have not yet explained how to choose the set
G . We will use the following definitions and theorems to help us select the
best G for a given system.
Set Z Definition 6: Set Z
Z = a: dV a dt = 0 a in the closure of G . (20.19)
20-12
Here the closure of G includes the interior and the boundary of G . This
is a key set. It contains all of those points where the derivative of the
Lyapunov function is zero. Later we will want to determine where in this
set the system trajectory can be trapped.
Invariant Set Definition 7: Invariant Set

n
A set of points in is invariant with respect to da dt = g a if every so-
lution of da dt = g a starting in that set remains in the set for all time.
If the system gets into an invariant set, then it cant get out.
Set L Definition 8: Set L

L is defined as the largest invariant set in Z .
This set includes all possible points at which the solution might converge.
The Lyapunov function does not change in L (because its derivative is ze-
ro), and the trajectory will be trapped in L (because it is an invariant set).
Now, if this set has only one stable point, then that point is asymptotically
stable. This is, in essence, what LaSalles theorem will say.
Theorem
LaSalles Invariance Theorem extends the Lyapunov Stability Theorem.
We will use it to design Hopfield networks in the next chapter. The theorem
proceeds as follows [Lasa67].
Theorem 2: LaSalles Invariance Theorem

If V is a Lyapunov function on G for da dt = g a , then each solution a t
that remains in G for all t 0 approaches L = L as t . ( G is a
basin of attraction for L , which has all of the stable points.) If all trajecto-
ries are bounded, then a t L as t .
If a trajectory stays in G , then it will either converge to L , or it will go to

infinity. If all trajectories are bounded, then all trajectories will converge
to L . 20
There is a corollary to LaSalles theorem that we will use extensively. It in-
volves choosing the set G in a special way.
20-13
20 Stability
Corollary 1: LaSalles Corollary

Let G be a component (one connected subset) of
= a:V a . (20.20)
Assume that G is bounded, dV a dt 0 on the set G , and let the set

L = closure L G be a subset of G . Then L is an attractor, and G is in
its region of attraction.
LaSalles theorem, and its corollary, are very powerful. Not only can they
tell us which points are stable ( L ), but they can also provide us with a par-
tial region of attraction ( G ). (Note that L is constructed differently in the
corollary than in the theorem.)
To clarify LaSalles Invariance Theorem, lets return to the pendulum ex-
ample we discussed earlier.
Example
2 Lets apply Corollary 1 to the pendulum example. The first step in using the
+2
corollary will be to choose the set . This set will then be used to select
the set G (a component of ).
For this example we will use the value = 100 , therefore 100 will be the
set of points where the energy is less than or equal to 100 .
100 = a:V a 100 (20.21)
This set is displayed in blue in Figure 20.8.
a2 0
-1
-2
-10 -5 0 5 10
a1
Figure 20.8 Illustration of the Set 100
The next step in our analysis is to choose a component (connected subset)

of 100 for the set G . Since we have been investigating the stability of the
20-14
origin, lets choose the component of 100 that contains a = 0 . The result-
ing set is shown in Figure 20.9.
a2 0
-1
-2
-10 -5 0 5 10
a1
Figure 20.9 Illustration of the Set G
Now that we have chosen G , we need to check that the derivative of the
Lyapunov function is less than or equal to zero on G . From Eq. (20.17) we
know that dV a dt is negative semidefinite. Therefore it will certainly be
less than or equal to zero on G .
We are now ready to determine the attractor set L . We begin with the set
L , which is the largest invariant set in Z .
Z = a: dV a dt = 0 a in the closure of G
= a: a 2 = 0, a in the closure of G .
(20.22)
This can also be written as
Z = a:a 2 = 0, 1.6 a 1 1.6 . (20.23)
We know from Eq. (20.17) that the derivative of V a is only zero when the
velocity is zero, which corresponds to the a 1 axis. Therefore Z consists of
the segment of the a 1 axis that falls within G . The set Z is displayed in
Figure 20.10.
The set L is the largest invariant set in Z . To find L we need to answer the
question: If we start the pendulum from an initial position between 1.6
and 1.6 radians, with zero initial velocity, will the velocity of the pendulum
remain zero? Clearly the only such initial condition would be 0 radians
(straight down). If we start the pendulum from any other position in Z , the
pendulum will start to fall, so the velocity will not remain zero and the tra-
jectory will move out of Z . Therefore, the set L consists only of the origin:
L = a: a = 0 . (20.24)
20-15
20 Stability
Z
a2 0
-1
-2
-10 -5 0 5 10
a1
Figure 20.10 Illustration of the Set Z
The set L is the closure of the intersection of L and G , which in this case
is simply L :
L = closure L G = L = a: a = 0 . (20.25)
Therefore, based on LaSalles corollary, L is an attractor (asymptotically

stable point) and G is in its region of attraction. This means that any tra-
jectory that starts in G will decay to the origin.
2
+2
Now suppose that we had taken a bigger region for , such as
300 = a: V a 300 . (20.26)
This set is shown in gray in Figure 20.11.
1
Z
a2 0
-1
-2
-10 -5 0 5 10
a1
Figure 20.11 Illustration of G = 300 (Gray) and Z
We let G = 300 , since 300 has only one component. The set Z is given by
20-16
Z = a: a 2 = 0 , (20.27)
which is shown by the blue bar on the horizontal axis of Figure 20.11. Thus,
it follows that
L = L = a: a 1 = n, a 2 = 0 . (20.28)
This is because there are now several different positions within the set Z
where we can place the pendulum, without causing the velocity to become
nonzero. The pendulum can be pointing directly up or directly down. This
corresponds to the positions n for any integer n . If we place the pendu-
lum in any of these positions, with zero velocity, then the pendulum will re-
main stationary. We can show this by setting the derivatives equal to zero
in Eq. (20.14) and Eq. (20.15).
da 1
= a2 = 0 , (20.29)
dt
da 2 = sin a 0.2a = sin a = 0 a = n (20.30)

dt 1 2 1 1
For this choice of G = 300 we can say very little about where the trajecto-
ry will converge. We tried to increase the size of our known region of attrac-
tion for the origin, but this G is a region of attraction for all of the
equilibrium points. We made G too large. The set L is illustrated by the
blue dots in Figure 20.12.
a2 0
-1
-2
-10 -5 0 5 10
a1
Figure 20.12 The Set L
We cannot tell which of the equilibrium points (blue dots) will attract the
trajectory. All we can say is that if we start somewhere in 300 , one of the
equilibrium points will attract the system solution, but we cannot say for
sure which one it will be. Consider, for instance, the trajectory shown in
20-17
20 Stability
Figure 20.13. This shows the pendulum response for an initial position of 2
radians and an initial velocity of 1.5 radians per second. This time the pen-
dulum had enough velocity to go over the top, and it converged to the equi-
librium point at 2 radians.
Now that we have discussed LaSalles Invariance Theorem, you might want
to experiment some more with the pendulum, in order to investigate the re-
gions of attraction for the various stable points. To experiment with the pen-
dulum, use the Neural Network Design Demonstration Dynamic System
(nnd17ds).
Contour Plot
a2 0
x2
-1
-2
-10 -5 0 5 10
ax11
Figure 20.13 Pendulum Trajectory for Different Starting Conditions
Comments
The keys to LaSalles theorem are the choices of the Lyapunov function V
and the set G . We want G to be as large as possible, because that will in-
dicate the region of attraction. However, we want to choose V so that the
set Z , which will contain the attractor set, is as small as possible.
For instance, we could try V = 0 . This is a Lyapunov function for the entire
n
space , since its derivative is zero (and therefore doesnt change sign) ev-
n
erywhere. However, it gives us no information since Z = .
Notice that if V 1 and V 2 are both Lyapunov functions on G , and dV 1 dt
and dV 2 dt have the same sign, then V = V 1 + V 2 is also a Lyapunov func-
tion, where Z = Z 1 Z 2 . If Z is smaller than both Z 1 and Z 2 , then Z is a
better Lyapunov function than either V 1 or V 2 . V is always at least as
good as either V 1 or V 2 , since Z can never be larger than the smaller of Z 1
and Z 2 . Therefore, if you have found two different Lyapunov functions and
their derivatives have the same sign, then add them together and you may
have a better function. The best Lyapunov function for a given system is
the one that has the smallest attractor set and the largest region of attrac-
tion.
20-18
Summary of Results
Summary of Results
Stability Concepts
Definitions
Definition 1: Stability (in the sense of Lyapunov)
The origin is a stable equilibrium point if for any given value 0 there
exists a number 0 such that if a 0 , then the resulting motion
a t satisfies a t for t 0 .
Definition 2: Asymptotic Stability

The origin is an asymptotically stable equilibrium point if there exists a
number 0 such that whenever a 0 the resulting motion satisfies
a 0 0 as t .
Definition 3: Positive Definite

A scalar function V a is positive definite if V 0 = 0 and V a 0 for
a0.
Definition 4: Positive Semidefinite

A scalar function V a is positive semidefinite if V a 0 for all a .

Consider the autonomous (unforced, no explicit time dependence) system
da
= ga .
dt
The Lyapunov stability theorem can then be stated as follows.

Theorem 1: Lyapunov Stability Theorem 20
If a positive definite function V a can be found such that dV a dt is neg-
ative semidefinite, then the origin ( a = 0 ) is stable for this system. If a pos-
itive definite function V a can be found such that dV a dt is negative
definite, then the origin ( a = 0 ) is asymptotically stable. In each case, V is
called a Lyapunov function of the system.
20-19
20 Stability

Definitions
Definition 5: Lyapunov Function
n
Let V be a continuously differentiable function from to . If G is any
n
subset of , we say that V is a Lyapunov function on G for the system
da dt = g a if
dV a T
--------------- = V a g a
dt
does not change sign on G .
Definition 6: Set Z
Z = a: dV a dt = 0 a in the closure of G . (20.31)
Definition 7: Invariant Set

n
A set of points G in is invariant with respect to da dt = g a if every
solution of da dt = g a starting in G remains in G for all time.
Definition 8: Set L
L is defined as the largest invariant set in Z .
Theorem
Theorem 2: LaSalles Invariance Theorem

If V is a Lyapunov function on G for da dt = g a , then each solution a t
that remains in G for all t 0 approaches L = L as t . ( G is a
basin of attraction for L , which has all of the stable points.) If all trajecto-
ries are bounded, then a t L as t .
Corollary 1: LaSalles Corollary

Let G be a component (one connected subset) of
= a:V a . (20.32)
Assume that G is bounded, dV a dt 0 on the set G , and let the set

L = closure L G be a subset of G . Then L is an attractor, and G is in
its region of attraction.
20-20
Solved Problems
Solved Problems
P20.1 Test the stability of the origin for the following system.
2
da 1 dt = a 1 + a 2
da 2 dt = a 2 a 1 + 1
The basic job here is to find a Lyapunov V a that is positive definite and
has a derivative that is negative semidefinite or, better yet, negative defi-
nite. (The latter is a stronger condition.)
2 2
Let us try V a = a 1 + a 2 . The derivative of V a is
dV a V da 1 V da 2
--------------- = V ------ =
T da
-------- + -------- ,
dt dt a 1 dt a 2 dt
or
dV a 2 2 2
--------------- = 2a 1 a 1 + a 2 + 2a 2 a 2 a 1 + 1 = 2 a 1 2 a 2 .
dt
The derivative dV a dt is negative definite. Therefore, the origin is as-

ymptotically stable.
P20.2 Test the stability of the origin for the following system.
5
da 1 dt = a 1
7
da 2 dt = 5 a 2
2 2
Let us try V a = a 1 + a 2 . Then we have
dV a 5 7 6 8
--------------- = 2a 1 a 1 + 2a 2 5 a 2 = 2 a 1 10 a 2 .
dt
Here again, dV a dt is negative definite, and therefore the origin is as-

ymptotically stable.
P20.3 Consider the mechanical system shown in Figure P20.1. This is a

spring-mass-damper system, with a nonlinear spring. We will de-
fine a 1 = x and a 2 = dx dt . Then the equations of motion are
da 1 dt = a 2 ,
20-21
20 Stability
3
da 2 dt = a 1 a 2 (nonlinear spring) .
Consider the candidate Lyapunov function
1 4 1 2
V a = --- a 1 + --- a 2 .
4 2
Use the corollary to LaSalles invariance theorem to provide as

much information as possible about the equilibrium points and ba-
sins of attraction.
Figure P20.1 Mechanical System

First calculate the derivative of V a as
dV a V da 1 V da 2 3 3 2
--------------- = -------- + -------- = a 1 a 2 + a 2 a 1 a 2 = a 2 .
dt a 1 dt a 2 dt
2
Thus, dV dt does not change sign on .
Now let us define
G = = a: V a
and consider the case for = 1 . A contour plot of V a is shown in Figure

P20.2. The set 1 is indicated in blue on the plot.
4
a2 0
-2
-4
-4 -2 0 2 4
a1
Figure P20.2 Contour Plot of V a and 1
20-22
Solved Problems
Now we need to determine the set Z .
Z = a: dV dt = 0, a in the closure of G = a: a 2 = 0 , a in the closure of G
or
Z = a: a 2 = 0, 2 a 1 2
Next we find the set L . Since a = 0 is the only invariant set,
L = a: a 1 = 0, a 2 = 0 .
Therefore, the origin,
a = 0 ,
0
is an attractor and 1 is in its region of attraction.

2
Further, we can increase to show that the entire is the basin of at-
traction for the origin.
Figure P20.3 shows the response of the spring-mass-damper from an initial
position of 2 and an initial velocity of 2 . Note that the trajectory is parallel
to the contour lines when the trajectory crosses the a 2 axis. This agrees
with our earlier result, which showed that the derivative of the Lyapunov
function was zero whenever a 2 = 0 . Fortunately, the a 2 axis is not an in-
variant set (except for the origin); therefore the trajectory is only attracted
to the origin.
Contour Plot
4
a2 0
x2
-2
-4
-4 -2 0 2 4
ax11
Figure P20.3 Spring-Mass-Damper Response
20-23
20 Stability
P20.4 Consider the following nonlinear system:

2 2
da 1 dt = a 1 a 1 + a 2 4 a 2
2 2
da 2 dt = a 1 + a 2 a 1 + a 2 4 .
This system has two invariant sets, the origin
a: a = 0 ,
and the circle

2 2
a: a 1 + a 2 = 4 .
Assuming the candidate Lyapunov function

2 2
V a = a1 + a2 ,
use LaSalles Invariance Theorem to find out as much as you can

about the region of attraction for the origin.
Our job, then, is to determine whether or not the given invariant sets rep-
resent a stable point or a stable trajectory. Lets first take a look at dV dt .
We recall that
dV a V da 1 V da 2
--------------- = -------- + -------- ,
dt a 1 dt a 2 dt
and substitute for the various terms to give
dV a 2 2 2 2
--------------- = 2a 1 a 1 a 1 + a 2 4 a 2 + 2a 2 a 1 + a 2 a 1 + a 2 4 .
dt
This can be simplified to
dV a 2 2 2 2
--------------- = 2 a 1 + a 2 a 1 + a 2 4 .
dt
2 2
Thus, dV dt is zero at a = 0 and on the circle a 1 + a 2 = 4 .
We now pick G , a region of attraction. Is there a change of sign of dV dt
2
over all ? Yes, there is. As we go from outside the circle of radius 2 to its
interior, the sign of dV dt changes from positive to negative. So dV dt is
2 2
negative semidefinite inside the circle a 1 + a 2 = 4 . Lets pick a G in-
side this circle, so that the circle will not be included. The following set will
do.
20-24
Solved Problems
G = 1 = a: V a 1
Now we consider 1 . There are just two places that dV dt = 0 , and the
only point inside 1 is a = 0 . Therefore,
Z = a: a 1 = 0, a 2 = 0
and
L = L = Z .
The origin is the attractor, and 1 is in its region of attraction.We can use
the same arguments to show that the region of attraction for the origin in-
2 2
cludes all points inside the circle a 1 + a 2 = 4 .
Figure P20.4 displays two trajectories for this system, one that begins in-
2 2
side the circle a 1 + a 2 = 4 , and one that begins outside the circle. Al-
though the circle is an invariant set, it is not an attractor. The only
attractor for this system is the origin.
Contour Plot
3
a2 0
x2
-1
-2
-3
-3 -2 -1 0 1 2 3
ax11
Figure P20.4 Sample Trajectories for Problem P20.4
P20.5 Consider the following nonlinear system.
da t dt = a t 1 a t 2
i. Find any equilibrium points for this system.

ii. Use the following candidate Lyapunov function to obtain
whatever information you can about the regions of attrac-
tion for the equilibrium points found in part (i). (Hint: Use
the corollary to LaSalles Invariance Theorem.)
20-25
20 Stability
2
Va = a 2
i. To find the equilibrium points, we set da t dt = 0 .
0 = a 1 a 2 a = 1, a = 2 are equilibrium points
ii. To use LaSalles corollary, we need to find dV dt .
dV V da 2
------- = = 2 a 2 a 1 a 2 = 2 a 1 a 2
dt a dt
Now we let
G = = a: V a .
For example, try = 0.5 . This gives

2
G = 0.5 = a: a 2 0.5 .
2
Note that a solution of a 2 0.5 yields
a 2 0.5 or 1.3 a 2.7 .
Thus, dV dt is negative definite on G .

Next we need to find the set Z , which contains those points within G where
dV dt is zero. There are two points where dV dt is zero, a = 1 and a = 2 .
Only one of these falls within G . Therefore
Z = a: a = 2 .
Now we need to find L , the largest invariant set in Z . There is only one
point in Z , and it is an equilibrium point. Thus
L = L = Z .
This means that G is in the region of attraction for 2.

We can use the same arguments with values of up to 1.0. So we can say
that the region for attraction for a = 2 must include at least
a: 1 a 3 .
What if we consider those regions where 1 ? Then Z includes both 1 and

2, and dV dt will change sign on G . Therefore we cannot say anything
about the region of attraction for a = 1 , using this Lyapunov function and
the corollary to LaSalles Invariance Theorem.
20-26
Solved Problems
Figure P20.5 displays some typical responses for this system. Here we can
see that the equilibrium point a = 1 is actually unstable. Any initial con-
dition above a = 1 converges to a = 2 . Anything less than a = 1 goes to
minus infinity.
4
0
0 1 2 3 4
t
Figure P20.5 Stable and Unstable Responses for Problem P20.5
20-27
20 Stability
Epilogue
In this chapter we have presented the concept of stability, as applied to dy-
namic systems. For nonlinear dynamic systems, like recurrent neural net-
works, we do not talk about the stability of the system. Rather, we discuss
the stability of certain system trajectories and, in particular, equilibrium
points.
There were two main stability theorems discussed in this chapter. The first
is the Lyapunov Stability Theorem, which introduces the concept of gener-
alized energy the Lyapunov function. The concept behind this theorem
is that if a systems energy is always decreasing, then it will eventually
stabilize at a point of minimum energy.
The second theorem presented was LaSalles Invariance Theorem, which is
an enhancement of the Lyapunov Stability Theorem. There are two key im-
provements made by LaSalle. The first is a clarification of the cases in
which the Lyapunov function does not decrease throughout the state space,
but stays constant in some regions. LaSalles theorem introduced the con-
cept of an invariant set to identify those regions that can trap the system
trajectory. The second improvement made by LaSalles theorem is that, in
addition to indicating the stability of equilibrium points, it also gave infor-
mation about the regions of attraction of each stable point.
The ideas presented in this chapter are important tools for the analysis of
recurrent neural networks, like the Grossberg networks of Chapters 18 and
19. (See [CoGr83] for an application of LaSalles Invariance Theorem to re-
current neural networks.) In Chapter 21 we will use LaSalles theorem to
explain the operation of the Hopfield network.
20-28
Further Reading
Further Reading
[CoGr83] M. A. Cohen and S. Grossberg, Absolute stability of global
pattern formation and parallel memory storage by compet-
itive neural networks, IEEE Transactions on Systems,
Man and Cybernetics, vol. 13, no. 5, pp. 815826, 1983.
Cohen and Grossberg apply LaSalles Invariance Theorem
to the analysis of the stability of competitive neural net-
works. The network description is very general, and the au-
thors show how their analysis can be applied to many
different types of recurrent neural networks.
[Lasa67] J. P. LaSalle, An invariance principle in the theory of sta-
bility, in Differential Equations and Dynamic Systems, J.
K. Hale and J. P. LaSalle, eds., New York: Academic Press,
pp. 277286, 1967.
This article provides a unified presentation of Lyapunovs
stability theory, including several extensions. It introduces
LaSalles Invariance Theorem and various corollaries.
[SlLi91] J.-J. E. Slotine and W. Li, Applied Nonlinear Control, En-
glewood Cliffs, NJ: Prentice-Hall, 1991.
This text is an introduction to nonlinear control systems. A
significant portion of the book is devoted to the analysis of
nonlinear dynamic systems. A number of stability theo-
rems are presented and demonstrated.
20
20-29
20 Stability
Exercises
E20.1 Use Lyapunovs Stability Theorem to test the stability of the origin for the
following systems.
3
i. da 1 dt = a 1 + a 2
da 2 dt = a 1 a 2
2
ii. da 1 dt = a 1 + a 2
da 2 dt = a 2 a 1 + 1
E20.2 Consider the following nonlinear system:

2 2
da 1 dt = a 2 2a 1 a 1 + a 2 ,
2 2
da 2 dt = a 1 2a 2 a 1 + a 2 .
i. Use Lyapunovs Stability Theorem and the candidate Lyapunov

function shown below to investigate the stability of the origin.
2 2
V a = a1 + a2
2+2
ans = ii. Check your stability result from part (i) by writing a MATLAB M-
4 file to simulate the response of this system for several different ini-
tial conditions. Use the ode45 routine. Plot the responses.
da dt = a a + 1 ,
i. Find any equilibrium points.

ii. The following Lyapunov function is proposed. Show that this is a
valid Lyapunov function for use in Lasalles invariance theorem.
3 2
V a = 2a + 3a .
iii. Use the corollary to Lasalles theorem and the proposed Lyapunov
function to provide as much information as you can about the stable
equilibrium points and their basins of attraction. Identify the sets
Z, G and L. Use graphs wherever possible.
20-30
Exercises
E20.4 Repeat E20.3 for the following systems and Lyapunov functions. (In some
cases, it may be useful to sketch the Lyapunov functions.)
2
i. da dt = a 2 a + 1 , V a = a + 1
3 2
ii. da dt = a a + 1 , V a = 2a + 3a
3 2
iii. da dt = a a + 2 , V a = a 3 a
3 2
iv. da dt = a a 1 , V a = 2a 3a
v. da dt = cos a , find a V a
vi. da dt = sin a , find a V a
da 1 dt = a 2 ,
2
da 2 dt = a 2 1 a 2 a 1 .
We want to use the corollary to Lasalles invariance theorem to locate at-

tractors and find out as much as we can about the basins of attraction, us-
ing the following Lyapunov function.
2 2
V a = a1 + a2 .

ii. Find dV a dt .
iii. Choose a set G.
iv. Find the corresponding set Z.
v. Find the set L.
vi. What have you learned about the attractors of this system and the
basins of attraction? Can you learn more by modifying the set G?
Explain.
2+2
ans = vii. Check your results by writing a MATLAB M-file to simulate the re-
4 sponse of this system for several different initial conditions. Use the
ode45 routine. Plot the responses.
da 1 dt = a 2 ,
20-31
20 Stability
3
da 2 dt = a 1 a 2 .

ii. Find as much information about the stability of the equilibrium
points as possible, using the corollary to LaSalles theorem and the
candidate Lyapunov function
2 2
V a = a1 + a2 .
2+2
ans = iii. Check your results from parts (i) and (ii) by writing a MATLAB M-
4 file to simulate the response of this system for several different ini-
tial conditions. Use the ode45 routine. Plot the responses.

2
da dt = 1 a 1 + a = 1 a .

ii. Find a suitable Lyapunov function. (Hint: Start with a form for
dV dt and work backward to find V .)
iii. Sketch the Lyapunov function.
iv. Use the corollary to LaSalles theorem and the Lyapunov function of
part (ii) to find as much information as possible about regions of at-
traction. Use graphs wherever possible.
(Hint: The graph shown in Figure E20.1 may be helpful.)
f (a) = 1 - (a / ) 2
Figure E20.1 Helpful Function for Exercise E20.7
20-32
Exercises

4 2
da 1 dt = a 2 a 1 a 1 + 2 a 2 10 ,
3 5 4 2
da 2 dt = a 1 3 a 2 a 1 + 2 a 2 10 .
2+2
i. Find any invariant sets. (You may want to simulate this system us-
ans = ing MATLAB in order to help identify the invariant sets.)
4
ii. Using the candidate Lyapunov function shown below and the corol-
lary to LaSalles theorem, investigate the stability of the invariant
sets you found in part (i).
4 2 2
V a = a 1 + 2 a 2 10
E20.9 Consider the following system:
da
= 1 0 a + 1
dt 0 2 2

ii. Find a Lyapunov function and identify attractors and basins of at-
traction. Use the corollary to Lasalles theorem and carefully iden-
tify and graph the sets , G, Z and L.
E20.10 For the nonlinear system

2 2
da 1 dt = a 2 + a 1 1 a 1 a 2 ,
2 2
da 2 dt = a 1 + a 2 1 a 1 a 2 ,
we know that the following sets are invariant:
a a = 0 ,
2
a a1 + a2 = 1 .
2 20
The following Lyapunov function is proposed:
2 2 2
V a = a1 + a2 1 .
20-33
20 Stability
Use the corollary to Lasalles theorem to find out as much as possible about
the basins of attraction for the two invariant sets given above, and graph
the sets , G, Z and L.
E20.11 Consider the system
da 1 dt = cos a 2 a 1 ,
da 2 dt = a 1 .

ii. The following Lyapunov function is proposed. Show that this is a
valid Lyapunov function for use in Lasalles invariance theorem.
2
a1
V a = -----------
- + sin a 2
2
iii. Use Lasalles theorem to find out as much information as you can
about the stable equilibrium points and their basins of attraction.
(Make a rough sketch of the contour plot for V a to assist you.)
20-34
Objectives
21 Hopfield Network
Objectives 21-1
Hopfield Model 21-3
Lyapunov Function 21-5
Invariant Sets 21-7
Example 21-8
Hopfield Attractors 21-11
Effect of Gain 21-13
Hopfield Design 21-16
Content-Addressable Memory 21-16
Hebb Rule 21-19
Lyapunov Surface 21-23
Epilogue 21-37
Exercises 21-41
Objectives
This chapter will discuss the Hopfield recurrent neural network a net-
work that was highly influential in bringing about the resurgence of neural
network research in the early 1980s. We will begin with a description of the
network, and then we will show how Lyapunov stability theory can be used
to analyze the network operation. Finally, we will demonstrate how the
network can be designed to behave as an associative memory.
This chapter brings together many topics discussed in previous chapters:
the discrete-time Hopfield network (Chapter 3), eigenvalues and eigenvec-
tors (Chapter 6); associative memory and the Hebb rule (Chapter 7); Hes-
sian matrices, conditions for optimality, quadratic functions and surface
and contour plots (Chapter 8); steepest descent and phase plane trajecto-
ries (Chapter 9); continuous-time recurrent networks (Chapter 18); and
Lyapunovs Stability Theorem and LaSalles Invariance Theorem (Chapter
20). This chapter is, in some ways, a culmination of all our previous efforts.
21-1
21 Hopfield Network
Theory and Examples

Much of the resurgence of interest in neural networks during the early
1980s can be attributed to the work of John Hopfield. As a well-known Cal.
Tech. physicist, Hopfields visibility and scientific credentials lent renewed
credibility to the neural network field, which had been tarnished by the
hype of the mid-1960s. Early in his career he studied the interaction be-
tween light and solids. Later he focused on the mechanism of electron
transfer between biological molecules. One can imagine that his academic
study in physics and mathematics, combined with his later experiences in
biology, prepared him uniquely for the conception and presentation of his
neural network contribution.
Hopfield wrote two highly influential papers in 1982 [Hopf82] and 1984
[Hopf84]. Many of the ideas in these papers were based on the previous
work of other researchers, such as the neuron model of McCulloch and Pitts
[McPi43], the additive model of Grossberg [Gros67], the linear associator of
Anderson [Ande72] and Kohonen [Koho72] and the Brain-State-in-a-Box
network of Anderson, Silverstein, Ritz and Jones [AnSi77]. However,
Hopfields papers are very readable, and they bring together a number of
important ideas and present them with a clear mathematical analysis (in-
cluding the application of Lyapunov stability theory).
There are several other reasons why Hopfields papers have had such an
impact. First, he identified a close analogy between his neural network and
the Ising model of magnetic materials, which is used in statistical physics.
This brought a significant amount of existing theory to bear on the analysis
of neural networks, and it encouraged many physicists, as well as other sci-
entists and engineers, to turn their attention to neural network research.
Hopfield also had close contacts with VLSI chip designers, because of his
long association with AT&T Bell Laboratories. As early as 1987, Bell Labs
had successfully developed neural network chips based on the Hopfield net-
work. One of the main promises of neural networks is their suitability for
parallel implementation in VLSI and optical devices. The fact that Hopfield
addressed the implementation issues of his networks distinguished him
from most previous neural network researchers.
Hopfield emphasized practicality, both in the implementation of his net-
works and in the types of problems they solved. Some of the applications
that he described in his early papers include content-addressable memory
(which we will discuss later in this chapter), analog-to-digital conversion
[TaHo86], and optimization [HoTa85] (as in the traveling salesman prob-
lem).
In the next section we will present the Hopfield model. We will use the con-
tinuous-time model from the 1984 paper [Hopf84]. Then we will apply
Lyapunov stability theory and LaSalles Invariance Theorem to the analy-
21-2
Hopfield Model
sis of the Hopfield model. In the final section we will demonstrate how the
Hebb rule can be used to design Hopfield networks as content-addressable
memories.
Hopfield Model
In keeping with his practical viewpoint, Hopfield presented his model as an
Hopfield Model electrical circuit. The basic Hopfield model (see [Hopf84]) is shown in Fig-
ure 21.1.
I1 I2 IS
R1,S
Amplifier RS,2
Inverting R2,1
Output
n1 n2 nS
C C C
Resistor
a1 a2 aS
Figure 21.1 Hopfield Model

Each neuron is represented by an operational amplifier and its associated
resistor/capacitor network. There are two sets of inputs to the neurons. The
first set, represented by the currents I 1 I 2 ... , are constant external inputs.
The other set consists of feedback connections from other op-amps. For in-
stance, the second output, a 2 , is fed to resistor R S 2 , which is connected, in
turn, to the input of amplifier S . Resistors are, of course, only positive, but
a negative input to a neuron can be obtained by selecting the inverted out-
put of a particular amplifier. (In Figure 21.1, the inverting output of the
first amplifier is connected to the input of the second amplifier through re-
sistor R 2 1 .)
The equation of operation for the Hopfield model, derived using Kirchhoffs
current law, is
S
dn i(t) n i(t)
C ------------- =
dt Ti j aj(t) ---------
Ri
- + Ii , (21.1)
j=1
21-3
21 Hopfield Network
where n i is the input voltage to the ith amplifier, a i is the output voltage
of the ith amplifier, C is the amplifier input capacitance and I i is a fixed
input current to the ith amplifier. Also,
S
1 1 1 1
-------
1
T i j = -------- , ----- = --- + -, n = f a i (or a i = f n i ), (21.2)
R i j R i R i j i
j=1
where f n is the amplifier characteristic. Here and in what follows we will

assume that the circuit is symmetric, so that T i j = T j i .
The amplifier transfer function, a i = f n i , is ordinarily a sigmoid function.
Both this sigmoid function and its inverse are assumed to be increasing
functions. We will provide a specific example of a suitable transfer function
later in this chapter.
If we multiply both sides of Eq. (21.1) by R i , we obtain
S
dn i(t)
R i C ------------
dt
- = Ri Ti j a j(t) n i(t) + Ri Ii . (21.3)
j=1
This can be transformed into our standard neural network notation if we

define
= R i C , w i j = R i T i j and b i = R i I i . (21.4)
Now Eq. (21.3) can be rewritten as

S
dn i(t)
------------
dt
- = n i(t) + wi j a j(t) + bi . (21.5)
j=1
In vector form we have
dn(t)
------------ = n(t) + Wa(t) + b . (21.6)
dt
and
a(t) = f n(t) . (21.7)
The resulting Hopfield network is displayed in Figure 21.2.

Thus, Hopfields original network of S operational amplifier circuits can be
represented conveniently in our standard network notation. Note that the
input vector p determines the initial network output. This form of the
Hopfield network is used for associative memory networks, as will be dis-
cussed at the end of this chapter.
21-4
Lyapunov Function
Input Recurrent Layer
p
W - .
Sx1 + n n a
SxS 1/ f
Sx1 Sx1
1 b
Sx1
S -1
S
f
-1
n(0) = f (p), (a(0) = p) dn/dt = - n + W f(n) + b
Lyapunov Function
The application of Lyapunov stability theory to the analysis of recurrent
networks was one of the key contributions of Hopfield. (Cohen and Gross-
berg also used Lyapunov theory for the analysis of competitive networks at
about the same time [CoGr83].) In this section we will demonstrate how
LaSalles Invariance Theorem, which was presented in Chapter 20, can be
used with the Hopfield network. The first step in using LaSalles theorem
is to choose a Lyapunov function. Hopfield suggested the following func-
tion:
ai
S
1 T 1
f u du b a .
T
V a = --- a Wa + (21.8)
2
i=1 0

Hopfields choice of this particular Lyapunov candidate is one of his key

contributions. Notice that the first and third terms make up a quadratic
function. In a later section of this chapter we will use our previous results
on quadratic functions to help develop some insight into this Lyapunov
function.
To use LaSalles theorem, we will need to evaluate the derivative of V a .
For clarity, we will consider each of the three terms of V a separately. Us-
ing Eq. (8.37), the derivative of the first term is
21-5
21 Hopfield Network
d 1 T 1 T T da T da T da
--- a Wa = --- a Wa ------ = Wa ------ = a W ------ . (21.9)
dt 2 2 dt dt dt
The second term in V a consists of a sum of integrals. If we consider one

of these integrals, we find
ai ai
d 1 d 1 da i da da

dt 0
f u du =
d ai 0
f u du
1
= f ai i = ni i .
dt dt
(21.10)
d t

The total derivative of the second term in V a is then
S ai
d 1 T da
-----
dt f u du = n -----
dt
-. (21.11)
i=1 0

Using Eq. (8.36), we can find the derivative of the third term in V a .
d T T T da T da
b a = b a ------ = b ------ (21.12)
dt dt dt
Therefore, the total derivative of V a is
d T da T da T da T T T da
V a = a W ------ + n ------ b ------ = a W + n b ------ . (21.13)
dt dt dt dt dt
From Eq. (21.6) we know that

T
T T T dn(t)
a W + n b = ------------ . (21.14)
dt
This allows us to rewrite Eq. (21.13) as

S
d dn(t) T da dn i da i .
V a = ------------ ------ = (21.15)
dt dt dt dt dt
i=1
1
Since n i = f a i , we can expand the derivative of n i as follows:
dn i d 1 d 1 da
= f ai = f ai i . (21.16)
dt dt d ai dt
Now Eq. (21.15) can be rewritten
21-6
Lyapunov Function
S S
dn da da 2
V a = i i = f a i i .
d d 1
(21.17)
dt dt dt d ai dt
i=1 i=1
1
If we assume that f a i is an increasing function, as it would be for an op-
erational amplifier, then
d 1
f ai 0 . (21.18)
d ai
From Eq. (21.17), this implies that
d
V a 0 . (21.19)
dt
1
Thus, if f a i is an increasing function, dV a dt is a negative semidefi-
nite function. Therefore, V a is a valid Lyapunov function.
Invariant Sets
Now we want to apply LaSalles Invariance Theorem to determine equilib-
rium points for the Hopfield network. The first step is to find the set Z (Eq.
(20.19)).
Z = a: dV a dt = 0 a in the closure of G (21.20)
This set includes all points at which the derivative of the Lyapunov func-
S
tion is zero. For now, lets assume that G is all of .
We can see from Eq. (21.17) that such derivatives will be zero if the deriv-
atives of all of the neuron outputs are zero.
da
------ = 0 (21.21)
dt
However, when the derivatives of the outputs are zero, the circuit is at
equilibrium. Thus, those points where the system energy is not changing
are also points where the circuit is at equilibrium.
This means that the set L , the largest invariant set in Z , is exactly equal
to Z .
L = Z (21.22)
Thus, all points in Z are potential attractors.

Some of these features will be illustrated in the following example.
21-7
21 Hopfield Network
Example
2 Consider the following example from Hopfields original paper [Hopf84].
+2
We will examine a system having an amplifier characteristic
2 1 n
a = f n = --- tan --------- . (21.23)
2
We can also write this expression as

n = ------ tan --- a .
2
(21.24)
2
Assume two amplifiers, with the output of each connected to the input of
the other through a unit resistor, so that
R 1 2 = R 2 1 = 1 and T 1 2 = T 2 1 = 1 . (21.25)
Thus we have a weight matrix
W = 0 1 . (21.26)
1 0
If the amplifier input capacitance is also set to 1, we have
= Ri C = 1 . (21.27)
Let us also take = 1.4 and I 1 = I 2 = 0 . Therefore
b = 0 . (21.28)
0
Recall from Eq. (21.8) that the Lyapunov function is
S ai
1 T 1
f u du b a .
T
V a = --- a Wa + (21.29)
2
i=1 0

The first term of the Lyapunov function, for this example, is
1 T 1 a
--- a Wa = --- a 1 a 2 0 1 1 = a 1 a 2 . (21.30)
2 2 1 0 a2
21-8
Lyapunov Function
The third term is zero, because the biases are zero. The ith part of the sec-
ond term is
ai ai
ai
- tan --- u du = ------ log cos --- u ---
2 2 2
f u du = -----
1
2 2
. (21.31)
0 0 0
This expression can be simplified to
ai

= -------2- log cos --- a i .
4
f 1 u du 2
(21.32)
0
Finally, substituting all three terms into Eq. (21.29), we have our
Lyapunov function:

V a = a 1 a 2 ------------2- log cos --- a 1 + log cos --- a 2 .
4
(21.33)
2 2
1.4
Now lets write out the network equation (Eq. (21.6)). With = 1 and
b = 0 , it is
dn
------ = n + Wf n = n + Wa . (21.34)
dt
If we substitute the weight matrix of Eq. (21.26), this expression can be

written as the following pair of equations:
dn 1 dt = a 2 n 1 , (21.35)
dn 2 dt = a 1 n 2 . (21.36)
The neuron outputs are
a 1 = --- tan ----------- n 1 ,

2 1 1.4
(21.37)
2
a 2 = --- tan ----------- n 2 .

2 1 1.4
(21.38)
2
Now that we have found expressions for the system Lyapunov function and
the network equation of operation, lets investigate the network behavior.
The Lyapunov function contour and a sample trajectory are shown in Fig-
ure 21.3.
21-9
21 Hopfield Network
0.5
a2 0
-0.5 1
0.5 1
0.5
a2 0 0
-1
-0.5
-0.5 a1
-1 -0.5 0 0.5 1 -1 -1
a1
Figure 21.3 Hopfield Example Lyapunov Function and Trajectory

The contour lines in this figure represent constant values of the Lyapunov
function. The system has two attractors, one in the lower left and one in the
upper right of Figure 21.3. Starting from the upper left, the system con-
verges, as shown by the blue line, to the stable point at the lower left.
Figure 21.4 displays the time response of the two neuron outputs.
1
0.5
0
a2
-0.5
a1
-1
0 2 4 6 8 10
Figure 21.4 Hopfield Example Time Response

Figure 21.5 displays the time response of the Lyapunov function. As ex-
pected, it decreases continuously as the equilibrium point is approached.
The system also has an equilibrium point at the origin. If the network is
initialized anywhere on a diagonal line drawn from the upper-left corner to
the lower-right corner, the solution converges to the origin. Any initial con-
ditions that do not fall on this line, however, will converge to one of the so-
lutions in the lower-left or upper-right corner. The solution at the origin is
a saddle point of the Lyapunov function, not a local minimum. We will dis-
21-10
Lyapunov Function
cuss this problem in a later section. Figure 21.6 displays a trajectory that
converges to the saddle point.
2
1.5
Va 1
0.5
0 2 4 6 8 10
Figure 21.5 Lyapunov Function Response

1
0.5
a2 0
-0.5
-1
-1 -0.5 0 0.5 1
a1
Figure 21.6 Hopfield Convergence to a Saddle Point

To experiment with the Hopfield network, use the Neural Network Design
Demonstration Hopfield Network (nnd18hn).
This example has provided some insight into the Hopfield attractors. In the
next section we will analyze them more carefully.
Hopfield Attractors
In the example network in the previous section we found that the Hopfield
network attractors were stationary points of the Lyapunov function. Now
we want to show that this is true in the general case. Recall from Eq.
(21.21) that the potential attractors of the Hopfield network satisfy
21-11
21 Hopfield Network
da
------ = 0 . (21.39)
dt
How are these points related to the minima of the Lyapunov function V a ?
In Chapter 8 (Eq. (8.27)) we showed that the minima of a function must be
stationary points (i.e., gradient equal to zero). The stationary points of
V a will satisfy
T
V = V V ... V = 0, (21.40)
a1 a2 aS
where
S ai
1 T 1
f u du b a .
T
V a = --- a Wa + (21.41)
2
i=1 0

If we follow steps similar to those we used to derive Eq. (21.13), we can find
the following expression for the gradient:
dn(t)
V a = W a + n b = ------------ . (21.42)
dt
The ith element of the gradient is therefore
dn d 1 d 1 da
V a = i = ( f a i ) = f a i i . (21.43)
ai dt dt d ai dt
1
Notice, incidentally, that if f a is linear, Eq. (21.43) implies that
da
------ = V a . (21.44)
dt
Therefore, the response of the Hopfield network is steepest descent. Thus,

1
if you are in a region where f a is approximately linear, the network so-
lution approximates steepest descent.
We have assumed that the transfer function and its inverse are monotonic
increasing functions. Therefore,
d 1
f ai 0 . (21.45)
d ai
From Eq. (21.43), this implies that those points for which
21-12
Effect of Gain
da(t)
------------ = 0 , (21.46)
dt
will also be points where
V a = 0 . (21.47)
Therefore, the attractors, which are members of the set L and satisfy Eq.
(21.39), will also be stationary points of the Lyapunov function V a .
Effect of Gain
The Hopfield Lyapunov function can be simplified if we consider those cas-
es where the amplifier gain is large. Recall that the nonlinear amplifier
characteristic for our previous example was
2 1 n
a = f n = --- tan --------- . (21.48)
2
This function is displayed in Figure 21.7 for four different gain values.
1
= 14
= 1.4
0.5
= 0.14
a 0
-0.5
-1
-5 -2.5 0 2.5 5
n
Figure 21.7 Inverse Tangent Amplifier Characteristic

The gain determines the steepness of the curve at n = 0 . As increases,
the slope of the curve at the origin increases. As goes to infinity, f n ap-
proaches a signum (step) function.
Now recall from Eq. (21.8) that the general Lyapunov function is
S ai
1 T 1
f u du b a .
T
V a = --- a Wa + (21.49)
2
i=1 0

21-13
21 Hopfield Network
For our previous example,
u
f u = ------ tan ------ .
1 2
(21.50)
2
Therefore, the second term in the Lyapunov function takes the form
ai
a a
= ------ --- log cos --------i = -------2- log cos --------i .
2 2 4
f 1 u du 2 2
(21.51)
0
A graph of this function is shown in Figure 21.8 for three different values
of the gain. Note that as increases the function flattens and is close to 0
most of the time. Thus, as the gain goes to infinity, the integral in the
second term of the Lyapunov function will be close to zero in the range
High-Gain 1 a i 1 . This allows us to eliminate that term, and the high-gain
Lyapunov Function Lyapunov function then reduces to
1 T T
V a = --- a Wa b a . (21.52)
2
1.5
= 0.14
1
= 1.4
0.5
0
= 14
-1 -0.5 0 0.5 1
a
Figure 21.8 Second Term in the Lyapunov Function

By comparing Eq. (21.52) with Eq. (8.35), we can see that the high-gain
Lyapunov function is, in fact, a quadratic function:
1 T T 1 T T
V a = --- a Wa b a = --- a Aa + d a + c , (21.53)
2 2
where
2V a = A = W , d = b and c = 0 . (21.54)
21-14
Effect of Gain
This is an important development, for now we can apply our results from
Chapter 8 on quadratic functions to the understanding of the operation of
Hopfield networks.
Recall that the shape of the surface of a quadratic function is determined
by the eigenvalues and eigenvectors of its Hessian matrix. The Hessian
matrix for our example Lyapunov function is
2V a = W = 0 1 . (21.55)
1 0
The eigenvalues of this Hessian matrix are computed as follows:
2V a I = 1 2
= 1 = + 1 1 . (21.56)
1
Thus, the eigenvalues are 1 = 1 and 2 = 1 . It follows that the eigen-

vectors are
z 1 = 1 and z 2 = 1 . (21.57)
1 1
What does the surface of the high-gain Lyapunov function look like? We
know, since the Hessian matrix has one positive and one negative eigenval-
ue, that we have a saddle point condition. The surface will have a negative
curvature along the first eigenvector and a positive curvature along the
second eigenvector. The surface is shown in Figure 21.9.
The function does not have a minimum. However, the network is con-
strained to the hypercube a: 1 a i 1 by the amplifier transfer func-
tion. Therefore, there will be constrained minima at the two corners of the
hypercube
a = 1 and a = 1 . (21.58)
1 1
When the gain is very small, there is a single minimum at the origin (see
Exercise E21.1). As the gain is increased, two minima move out from the
origin toward the two corners given by Eq. (21.58). Figure 21.3 displays an
intermediate case, where the gain is = 1.4 . The minima in that figure oc-
cur at
a = 0.57 and a = 0.57 . (21.59)

0.57 0.57
21-15
21 Hopfield Network
In the general case, where there are more than two neurons in the network,
the high-gain minima will fall in certain corners of the hypercube
a: 1 a i 1 . We will discuss the general case in more detail in later sec-
tions, after we describe the Hopfield design process.
0.5
0.5
0
0
-0.5
-0.5 -1
1
0.5 1
0 0.5
0
-1 -0.5
-0.5
-1 -0.5 0 0.5 1
-1 -1
Figure 21.9 Example High Gain Lyapunov Function
Hopfield Design
The Hopfield network does not have a learning law associated with it. It is
not trained, nor does it learn on its own. Instead, a design procedure based
on the Lyapunov function is used to determine the weight matrix.
Consider again the high-gain Lyapunov function
1 T T
V a = --- a Wa b a . (21.60)
2
The Hopfield design technique is to choose the weight matrix W and the
bias vector b so that V takes on the form of a function that you want to
minimize. Convert whatever problem you want to solve into a quadratic
minimization problem. Since the Hopfield network will minimize V , it will
also solve the original problem. The trick, of course, is in the conversion,
which is generally not straightforward.
Content-Addressable Memory
In this section we will describe how a Hopfield network can be designed to
work as an associative memory. The type of associative memory we will de-
Content-Addressable sign is called a content-addressable memory, because it retrieves stored
Memory memories on the basis of part of the contents. This is in contrast to stan-
dard computer memories, where items are retrieved based on their ad-
dresses. For example, if we have a content-addressable data base that
contains names, addresses and phone numbers of employees, we can re-
trieve a complete data entry simply by providing the employee name (or
21-16
Hopfield Design
perhaps a partial name). The content-addressable memory is effectively

the same as the autoassociative memory described in Chapter 7 (see page
7-10), except that in this chapter we will use the recurrent Hopfield net-
work instead of the linear associator.
Suppose that we want to store a set of prototype patterns in a Hopfield net-
work. When an input pattern is presented to the network, the network
should produce the stored pattern that most closely resembles the input
pattern. The initial network output is assigned to the input pattern. The
network output should then converge to the prototype pattern closest to the
input pattern. For this to happen, the prototype patterns must be minima
of the Lyapunov function.
Lets assume that the prototype patterns are
p 1 p 2 p Q . (21.61)
Each of these vectors consists of S elements, having the values 1 or 1 . As-

sume further that Q S , so that the state space is large, and that the pro-
totype patterns are well distributed in this space, and so will not be close
to each other.
In order for a Hopfield network to be able to recall the prototype patterns,
the patterns must be minima of the Lyapunov function. Since the high-gain
Lyapunov function is quadratic, we need the prototype patterns to be (con-
strained) minima of an appropriate quadratic function. We propose the fol-
lowing quadratic performance index:
Q
1 2
pq
T
J a = --- a . (21.62)
2
q=1
If the elements of the vectors a are restricted to be 1 , this function is min-

imized at the prototype patterns, as we will now show.
Assume that the prototype patterns are orthogonal. If we evaluate the per-
formance index at one of the prototype patterns, we find
Q
1 2 1 2 S
pq
T T
J p j = --- p j = --- p j p j = --- . (21.63)
2 2 2
q=1
The second equality follows from the orthogonality of the prototype pat-
terns. The last equality follows because all elements of p j are 1 .
Next, evaluate the performance index at a random input pattern a , which
is presumably not close to any prototype pattern. Each element in the sum
in Eq. (21.62) is an inner product between a prototype pattern and the in-
put. The inner product will increase as the input moves closer to a proto-
21-17
21 Hopfield Network
type pattern. However, if the input is not close to any prototype pattern,
then all terms of the sum in Eq. (21.62) will be small. Therefore, J a will
be largest (least negative) when a is not close to any prototype pattern, and
will be smallest (most negative) when a is equal to any one of the prototype
patterns.
We have now found a quadratic function that accurately indicates the per-
formance of the content-addressable memory. The next step is to choose the
weight matrix W and bias b so that the Hopfield Lyapunov function V will
be equivalent to the quadratic performance index J .
If we use the supervised Hebb rule to compute the weight matrix (with tar-
get patterns being the same as input patterns) as
Q
pq pq
T
W = , (21.64)
q=1
and set the bias to zero
b = 0, (21.65)
then the Lyapunov function is

Q Q
1 T 1 T 1
a
T T T
V a = --- a Wa = --- a p q p q a = --- pq pq a . (21.66)
2 2 2
q=1 q=1
This can be rewritten

Q
1 2
pq
T
V a = --- a = J a . (21.67)
2
q=1
Therefore, the Lyapunov function is indeed equal to the quadratic perfor-

mance index for the content-addressable memory problem. The Hopfield
network output will tend to converge to the stored prototype patterns
(among other possible equilibrium points, as we will discuss later).
As noted in Chapter 7, the supervised Hebb rule does not work well if there
is significant correlation between the prototype patterns. In that case the
pseudoinverse technique has been suggested. Another design technique,
which is beyond the scope of this text, is given in [LiMi89].
In the best situation, where the prototype patterns are orthogonal, every
prototype pattern will be an equilibrium point of the network. However,
there will be many other equilibrium points as well. The network may well
converge to a pattern that is not one of the prototype patterns. A general
rule is that, when using the Hebb rule, the number of stored patterns can
21-18
Hopfield Design
be no more than 15% of the number of neurons.The reference [LiMi89] dis-

cusses more complex design procedures, which minimize the number of
spurious equilibrium points.
In the next section we will analyze the location of the equilibrium points
more closely.
Hebb Rule
Lets take a closer look at the operation of the Hopfield network when the
Hebb rule is used to compute the weight matrix and the prototype patterns
are orthogonal. (The following analysis follows the discussion in the Chap-
ter 7, Problem P7.5.) The supervised Hebb rule is given by
Q
pq pq
T
W = . (21.68)
q=1
If we apply the prototype vector p j to the network, then

Q
pq pq
T T
Wp j = p j = p j p j p j = Sp j , (21.69)
q=1
where the second equality holds because the prototype patterns are orthog-
onal, and the third equality holds because each element of p j is either 1 or
1 . Eq. (21.69) is of the form
Wp j = p j . (21.70)
Therefore, each prototype vector is an eigenvector of the weight matrix and

they have a common eigenvalue of = S . The eigenspace X for the eigen-
value = S is therefore
X = span p 1 , p 2 , ... ,p Q . (21.71)
This space contains all vectors that can be written as linear combinations
of the prototype vectors. That is, any vector a that is a linear combination
of the prototype vectors is an eigenvector.
Wa = W 1 p 1 + 2 p 2 + + Q p Q
= 1 Wp 1 + 2 Wp 2 + + Q Wp Q
= 1 Sp 1 + 2 Sp 2 + + Q Sp Q
= S 1 p 1 + 2 p 2 + + Q p Q = Sa (21.72)
21-19
21 Hopfield Network
The eigenspace for the eigenvalue = S is Q-dimensional (assuming that

the prototype vectors are independent).
S
The entire space R can be divided into two disjoint sets [Brog85],
S
R = XX , (21.73)

where X is the orthogonal complement of X . (This is true for any set X ,

not just the one we are considering here.) Every vector in X is orthogonal

to every vector in X . This means that for any vector a X ,
T
p q a = 0, q = 1 2 Q . (21.74)

Therefore, if a X ,
Q Q
pq pq pq 0
T
Wa = a = = 0 = 0a. (21.75)
q=1 q=1

So X defines an eigenspace for the repeated eigenvalue = 0 .
To summarize, the weight matrix has two eigenvalues, S and 0 . The
eigenspace for the eigenvalue S is the space spanned by the prototype vec-
tors. The eigenspace for the eigenvalue 0 is the orthogonal complement of
the space spanned by the prototype vectors.
Since (from Eq. (21.54)) the Hessian matrix for the high-gain Lyapunov
function V is
2V = W , (21.76)
the eigenvalues for 2V will be S and 0 .

The high-gain Lyapunov function is a quadratic function. Therefore, the
eigenvalues of the Hessian matrix determine its shape. Because the first
eigenvalue is negative, V will have negative curvature in X . Because the

second eigenvalue is zero, V will have zero curvature in X .
What do these results say about the response of the Hopfield network? Be-
cause V has negative curvature in X , the trajectories of the Hopfield net-
work will tend to fall into the corners of the hypercube a: 1 a i 1 that
are contained in X .
Note that if we compute the weight matrix using the Hebb rule, there will
be at least two minima of the Lyapunov function for each prototype vector.
If p q is a prototype vector, then p q will also be in the space spanned by
the prototype vectors, X . Therefore, the negative of each prototype vector
will be one of the corners of the hypercube a: 1 a i 1 that are con-
21-20
Hopfield Design
tained in X . There will also be a number of other minima of the Lyapunov

function that do not correspond to prototype patterns.
The minima of V are in the corners of the hypercube a: 1 a i 1 that
are contained in X . These corners will include the prototype patterns, but
they will also include some linear combinations of the prototype patterns.
Spurious Patterns Those minima that are not prototype patterns are often referred to as spu-
rious patterns. The objective of Hopfield network design is to minimize the
number of spurious patterns and to make the basins of attraction for each
of the prototype patterns as large as possible. A design method that is guar-
anteed to minimize the number of spurious patterns is described in
[LiMi89].
2 To illustrate these principles, consider again the second-order example we
+2
have been discussing, where the connection matrix is
W = 0 1 . (21.77)
1 0
Suppose that this had been designed using the Hebb rule with one proto-
type pattern (obviously not an interesting practical case):
p1 = 1 . (21.78)
1
Then
W = p1 p1 = 1 1 1 = 1 1 .
T
(21.79)
1 1 1
Notice that
W' = W I = 0 1 (21.80)
1 0
corresponds to our original connection matrix. More about this in the next
section.
The high-gain Lyapunov function is
1 T 1 T
V a = --- a Wa = --- a 1 1 a .
2 2 1 1
(21.81) 21
The Hessian matrix for V a is
21-21
21 Hopfield Network
2V a = W = 1 1 . (21.82)
1 1
Its eigenvalues are
1 = S = 2 , and 2 = 0 , (21.83)
and the corresponding eigenvectors are
z 1 = 1 and z 2 = 1 . (21.84)
1 1
The first eigenvector, corresponding to the eigenvalue S , represents the

space spanned by the prototype vector:
X = a: a 1 = a 2 . (21.85)
The second eigenvector, corresponding to the eigenvalue 0 , represents the

orthogonal complement of the first eigenvector:

X = a: a 1 = a 2 . (21.86)
The Lyapunov function is displayed in Figure 21.10.

This surface has a straight ridge from the upper-left to the lower-right cor-

ner. This represents the zero curvature region of X . Initial conditions to
the left or to the right of the ridge will converge to the points
a = 1 or a = 1 , (21.87)
1 1
respectively. Initial conditions exactly on this ridge will stabilize where

they start. This situation is the same as that for our original example (see
Figure 21.9), except that in that case, initial points on the sloping ridge con-
verged to the origin, instead of remaining where they started (see Figure
21.6). Initial points to the right or to the left of the ridge, in both systems,
converge to the prototype design points. Thus, the convergence of our orig-
inal system, and the convergence of the system with zero diagonal ele-
ments, are identical in every important aspect. We will investigate this
further in the next section.
21-22
Hopfield Design
1
0
-0.5
0.5
-1
-1.5
0
-2
1
-0.5 0.5
-0.5 1
0.5
0
-1 -1 -0.5
-1 -0.5 0 0.5 1 -1
Figure 21.10 Example Lyapunov Function
Lyapunov Surface
In many discussions of the Hopfield network the diagonal elements of the
weight matrix are set to zero. In this section we will analyze the effect of
this operation on the Lyapunov surface. (See also Chapter 7, Exercise
E7.5.)
For the content-addressable memory network, all of the diagonal elements
of the weight matrix will be equal to Q (the number of prototype patterns),
since the elements of each p q are 1 . Therefore, we can zero the diagonal
by subtracting Q times the identity matrix:
W' = W QI . (21.88)
Lets investigate how this change affects the form of the Lyapunov func-
tion. If we multiply this new weight matrix times one of the prototype vec-
tors we find
W'p q = W QI p q = Sp q Qp q = S Q p q . (21.89)
Therefore, S Q is an eigenvalue of W' , and the corresponding eigens-

pace is X , the space spanned by the prototype vectors.
If we multiply the new weight matrix times a vector from the orthogonal

complement space, a X , we find
W'a = W QI a = 0 Qa = Qa . (21.90)
Therefore, Q is an eigenvalue of W' , and the corresponding eigenspace is

X .
To summarize, the eigenvectors of W' are the same as the eigenvectors of
W , but the eigenvalues are now S Q and Q , instead of S and 0 . There-
21-23
21 Hopfield Network
fore, the eigenvalues of the Hessian matrix of the modified Lyapunov func-
tion, 2V' a = W' , are S Q and Q .
This implies that the energy surface will have negative curvature in X and

positive curvature in X , in contrast with the original Lyapunov function,

which had negative curvature in X and zero curvature in X .
A comparison of Figure 21.9 and Figure 21.10 demonstrates the effect on
the Lyapunov function of setting the diagonal elements of the weight ma-
trix to zero. In terms of system performance, the change has little effect. If
the initial condition of the Hopfield network falls anywhere off of the line
a 1 = a 2 , then, in either case, the output of the network will converge to
one of the corners of the hypercube a: 1 a i 1 , which consists of the
T T
two points a = 1 1 and a = 1 1 .
If the initial condition falls exactly on the line a 1 = a 2 , and the weight
matrix W is used, then the network output will remain constant. If the ini-
tial condition falls exactly on the line a 1 = a 2 , and the weight matrix W'
is used, then the network output will converge to the saddle point at the
origin (as in Figure 21.6). Neither of these results is desirable, since the
network output does not converge to a minimum of the Lyapunov function.
Of course, the only case in which the network converges to a saddle point
is when the initial condition falls exactly on the line a 1 = a 2 , which would
be highly unlikely in practice.
21-24
Summary of Results
Summary of Results
Hopfield Model
dn(t)
------------ = n(t) + Wa(t) + b
dt
a(t) = f n(t)
Input Recurrent Layer
p
W - .
Sx1 + n n a
SxS 1/ f
Sx1 Sx1
1 b
Sx1
S S
f -1
n(0) = f -1(p), (a(0) = p) dn/dt = - n + W f(n) + b
Lyapunov Function
ai
S
1 T 1
f u du b a
T
V a = --- a Wa +
2
i=1 0

S
da 2
V a = f a i i
d d 1
dt d ai dt
i=1
d 1 d
If f a i 0 , then V a 0 .
d ai dt
Invariant Sets 21
The Invariant Set Consists of the Equilibrium Points.
L = Z = a:da dt = 0 a in the closure of G
21-25
21 Hopfield Network
Hopfield Attractors
The Equilibrium Points Are Stationary Points.
da(t)
If ------------ = 0 , then V a = 0 .
dt
dn(t)
V a = Wa + n b = ------------
dt
High-Gain Lyapunov Function

1 T T
V a = --- a Wa b a
2
2V a = W
Content-Addressable Memory
Q
pq pq
T
W = and b = 0
q=1
Energy Surface (Orthogonal Prototype Patterns)

Eigenvalues/Eigenvectors of 2V a = W Are
1 = S , with eigenspace X = span p 1 , p 2 , ... ,p Q .

2 = 0 , with eigenspace X .
T
X is defined such that for any vector a X , p q a = 0, q = 1 2 Q
Trajectories (Orthogonal Prototype Patterns)

Because the first eigenvalue is negative, V a will have negative curvature
in X . Because the second eigenvalue is zero, V a will have zero curvature

in X . Because V a has negative curvature in X , the trajectories of the
Hopfield network will tend to fall into the corners of the hypercube
a: 1 a i 1 that are contained in X .
21-26
Solved Problems
Solved Problems
P21.1 Assume the binary prototype vectors
1 1
p1 = 1 p2 = 1 .

1 1
1 1
i. Design a continuous-time Hopfield network (specify connec-

tion weights) to recognize these patterns, using the Hebb
rule.
ii. Find the Hessian matrix of the high-gain Lyapunov function
for this network. What are the eigenvalues and eigenvectors
of the Hessian matrix?
iii. Assuming large gain, what are the stable equilibrium points
for this Hopfield network?
i. First calculate the weight matrix from the reference vectors, using the
supervised Hebb rule.
1 1 1 1 1 1 1 1
W = p1 p1 + p2 p2 = 1 1 1 1 + 1 1 1 1 ,
T T
1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1
which simplifies to
2 0 0 2
W = 0 2 2 0 .
0 2 2 0
2 0 0 2
ii. The Hessian of the high-gain Lyapunov function, from Eq. (21.54), is
the negative of the weight matrix:
2 0 0 2
2V a = 0 2 2 0 .
0 2 2 0
2 0 0 2
21-27
21 Hopfield Network
T
The prototype patterns are orthogonal ( p 1 p 2 = 0 ). Thus, the eigenval-
ues are 1 = S = 4 and 2 = 0 . The eigenspace for 1 = 4 is
X = span p 1 , p 2 .
The eigenspace for 2 = 0 is the orthogonal complement of X :

1 1
1 , 1
X = span ,
1 1
1 1

where we have chosen two vectors that are orthogonal to both p 1 and p 2 .
iii. The stable points will be p 1 , p 2 p 1 p 2 since the negative of the pro-
totype patterns will also be equilibrium points. There may be other equilib-
rium points, if other corners of the hypercube lie in the span p 1 , p 2 .
4
There are a total of 2 = 16 corners of the hypercube. Four will fall in X
T
and four will fall in X . The other corners are partly in X and partly in X .
P21.2 Consider a high-gain Hopfield network with a weight matrix and

bias given by
W = 1 1 and b = 1 .
1 1 1
i. Sketch a contour plot of the high-gain Lyapunov function

for this network.
T
ii. If the network is given the initial condition 1 1 , where
will the network converge?
i. First consider the high-gain Lyapunov function
1 T T
V a = --- a Wa b a .
2
The Hessian matrix is
2V a = W = 1 1 .
1 1
Next, we need to compute the eigenvalues and eigenvectors:
21-28
Solved Problems
2V a I = 1 1 2
= 2 + 1 1 = 2 .
1 1
The eigenvalues are 1 = 0 and 2 = 2 .

Now we can find the eigenvectors. For 1 = 0 ,
2V a 1 I z 1 = 0 ,
and therefore
1 1 z = 0 or z = 1 .
1 1
11 1
Similarly, for 2 = 2 ,
2V a 2 I z 2 = 0
and therefore
1 1 z = 0 or z = 1 .
2 2
1 1 1
So the term
1 T
--- a Wa
2
has zero curvature in the direction z 1 and positive curvature in the direc-
tion z 2 .
Now we have to account for the linear term. First plot the contour without
the linear term, as in Figure P21.1.
The linear term will cause a negative slope in the direction of
b = 1 .
1
T
Therefore everything will curve down toward 1 1 , as is shown in Fig-
ure P21.2.
21-29
21 Hopfield Network
1
2
1.5
0.5
0.5
0
0
1
-0.5 0.5
-0.5 1
0.5
0
-1 -1 -0.5
-1 -0.5 0 0.5 1 -1
Figure P21.1 Contour Without Linear Term

T
ii. All trajectories will converge to 1 1 , regardless of the initial condi-
tions. As we can see in Figure P21.2, the energy function has only one min-
T
imum, which is located at 1 1 . (Keep in mind that the output of the
network is constrained to fall within the hypercube a: 1 a i 1 .)
0.5
1
0
0
-1
-2
-0.5 1
0.5 1
0 0.5
0
-0.5
-0.5
-1 -1 -1
-1 -0.5 0 0.5 1
Figure P21.2 Contour Including Linear Term
P21.3 Consider the following prototype vectors.
p1 = 1 p2 = 1
1 1
i. Design a Hopfield network to recognize these patterns.

ii. Find the Hessian matrix of the high-gain Lyapunov function
for this network. What are the eigenvalues and eigenvectors
of the Hessian matrix?
21-30
Solved Problems
iii. What are the stable points for this Hopfield network (as-
sume large gain)? What are the basins of attraction?
iv. How well does this network perform the pattern recognition
problem?
i. We will use the Hebb rule to find the weight matrix.
W = p1 p1 + p2 p2 = 1 1 + 1 1 = 2 0
T T
1 1 1 1 0 2
The bias is set to zero.
b = 0
0
ii. The Hessian matrix of the high-gain Lyapunov function is the negative
of the weight matrix.
2V a = W = 2 0
0 2
By inspection, we can see that there is a repeated eigenvalue.
1 = 2 = S = 2
The eigenvectors will then be
z 1 = 1 and z 2 = 0 ,
0 1
2
or any linear combination. (The entire is the eigenspace for the eigen-
value = 2 .)
iii. From Chapter 8 we know that when the eigenvalues of the Hessian are
equal, the contours will be circular. Because the eigenvalues are negative,
the function will have a single maximum at the origin. There will be four
minima at the four corners of the hypercube a: 1 a i 1 . There are also
four saddle points. The high-gain Lyapunov function is displayed in Figure
P21.3.
21-31
21 Hopfield Network
0.5
-0.5
-1
0
-1.5
-2
-0.5 1
0.5 1
0 0.5
0
-0.5
-0.5
-1 -1
-1 -0.5 0 0.5 1 -1
Figure P21.3 High-Gain Lyapunov Function for Problem P21.3

There are a total of nine stationary points. We could use the corollary to La-
Salles Invariance Theorem to show that the maximum point at the origin
has a basin of attraction that only includes the origin itself. Therefore it is
not a stable equilibrium point. The saddle points have regions of attraction
T
that are lines. (For example, the saddle point at 1 0 has a region of at-
traction along the negative a 1 axis.) The four corners of the hypercube are
the only attractors that have two-dimensional regions of attraction. The re-
gion of attraction for each corner is the corresponding quadrant of the hy-
percube. Figure P21.4 shows the low-gain Lyapunov function (with gain
= 1.4 ) and illustrates convergence to a saddle point and to a minimum.
0.5
-0.25
-0.5
0
-0.75
-1
-0.5 1
0.5 1
0 0.5
0
-0.5
-0.5
-1 -1
-1 -0.5 0 0.5 1 -1
Figure P21.4 Lyapunov Function for Problem P21.3

iv. The network does not do a good job on the pattern recognition problem.
Not only does it recognize the two prototype patterns, but it also recogniz-
es the other two corners of the hypercube as well. The network will con-
verge to whichever corner is closest to the input pattern, even though we
only wanted it to store the two prototype patterns. Since every possible
21-32
Solved Problems
two-bit pattern has been stored, the network is not very useful. This is not
unexpected, since the number of patterns that the Hebb rule is expected to
store is only 15% of the number of neurons. Since we only have two neu-
rons, we cant expect to successfully store many patterns. Try Exercise
E21.2 for a better network.
P21.4 A Hopfield network has the following high-gain Lyapunov func-

tion:
1 2 2
V a = --- 7 a 1 + 12a 1 a 2 2 a 2 .
2
i. Find the weight matrix.

ii. Find the gradient vector of the Lyapunov function.
iii. Find the Hessian matrix of the Lyapunov function.
iv. Sketch a contour plot of the Lyapunov function.
v. Sketch the path that a steepest descent algorithm would fol-
T
low for V a with an initial condition of 0.25 0.25 .
i. V a is a quadratic function, which can be rewritten as
1 1 T
V a = --- 7 a 1 + 12a 1 a 2 2 a 2 = --- a 7 6 a .
2 2
2 2 6 2
Therefore the weight matrix is
W = 7 6 .
6 2
ii. Since V a is a quadratic function, we can use Eq. (8.38) to find the gra-
dient.
V a = 7 6 a
6 2
iii. From Eq. (8.39), the Hessian is
2V a = 7 6 = 7 6 .
6 2 6 2
21-33
21 Hopfield Network
iv. To compute the eigenvalues,
2V a I = 7 6 2
= + 5 50 = + 10 5 .
6 2
The eigenvalues are 1 = 10 and 2 = 5 .

Now we can find the eigenvectors. For 1 = 10 ,
2V a 1 I z 1 = 0 ,
and therefore
3 6 z = 0 or z = 2 .
1 1
6 12 1
Similarly, for 2 = 5 ,
2V a 2 I z 2 = 0
and therefore
12 6 z = 0 or z = 1 .
2 2
6 3 2
Note that this is a saddle point case, since 1 0 2 . There will be nega-
tive curvature along z 1 and positive curvature along z 2 . The contour plot
of the high-gain Lyapunov function is shown in Figure P21.5.
1
0.5
0
-4
-8
-0.5 1
0.5 1
0 0.5
0
-0.5
-0.5
-1 -1
-1 -0.5 0 0.5 1 -1
Figure P21.5 High-Gain Lyapunov Func. & Steepest Descent Trajectory

v. The steepest descent path will follow the negative of the gradient and
will be perpendicular to the contour lines, as we saw in Chapter 9. When
21-34
Solved Problems
the trajectory hits the edge of the hypercube, it follows the edge down to the
minimum point. The resulting trajectory is shown in Figure P21.5.
The high-gain Lyapunov function is only an approximation, since it as-
sumes infinite gain. As a comparison, Figure P21.6 illustrates the
Lyapunov function, and the Hopfield trajectory, for a gain of 0.5 .
1
0.5 6
0 0
-2
-4
-6
-0.5 1
0.5 1
0 0.5
0
-0.5
-0.5
-1 -1
-1 -0.5 0 0.5 1 -1
Figure P21.6 Lyapunov Function & Hopfield Trajectory
P21.5 The Hopfield network has been used for applications other than
content-addressable memory. One of these other applications is
analog-to-digital (A/D) conversion [HoTa86]. The function of the
A/D converter is to take an analog signal y , and convert it into a se-
ries of bits (zeros and ones). For example, a two-bit A/D converter
would try to approximate the signal y as follows:
2
i 1
y ai 2 = a1 + a2 2 ,
i=1
where a 1 and a 2 are allowed values of 0 or 1 . (This A/D converter

would approximate analog values in the range from 0 to 3 , with a
resolution of 1 .) Tank and Hopfield suggest the following perfor-
mance index for the A/D conversion process:
2 2 2
1 i 1 1 2i 1
J a = --- y
2 ai 2 --- 2
2
a i a i 1 ,

i=1 i=1
where the first term represents the A/D conversion error, and the
second term forces a 1 and a 2 to take on values of 0 or 1 .
Show that this performance index can be written as the Lyapunov
function of a Hopfield network and define the appropriate weight
matrix and bias vector.
21-35
21 Hopfield Network
The first step is to expand the terms of the performance index.

2 2 2 2 2
i 1 i 1 i 1 + j 1
ai 2 = y 2y a i 2 ai aj 2
2
y + ,

i=1 i=1 j = 1i = 1
2 2 2
2i 1 2 2i 1 2i 1
2 a i a i 1 = ai 2 ai 2

i=1 i=1 i=1
If we substitute these terms back into the performance index we find
2 2 2
1 2 i 1 + j 1 2i 1 i
J a = --- y + a i a j 2 + ai 2 2 y .
2
j = 1i = 1 i=1
ij
The first term is not a function of a . Therefore, it does not affect where the
minima will occur, and we can ignore it.
We now want to show that this performance index takes the form of a high-
gain Lyapunov function:
1 T T
V a = --- a Wa b a .
2
This will be the case if
1
0 2 and b = y ---
W = 2 .
2 0
2y 2
In this Hopfield network, unlike the content-addressable memory, the in-

put to the network is the scalar y , which is then used to compute the bias
vector. In the content-addressable memory, the inputs to the network were
vector patterns, which became the initial conditions on the network out-
puts.
Note that in this network the transfer function must limit the output to the
range 0 a 1 . One transfer function that could be used is
1
f n = ----------------------
n
-.
1 e
21-36
Epilogue
Epilogue
In this chapter we have introduced the Hopfield model, one of the most in-
fluential neural network architectures. One of the reasons that Hopfield
was so influential was that he emphasized the practical considerations of
the network. He described how the network could be implemented as an
electrical circuit, and VLSI implementations of Hopfield-type networks
were built at an early stage.
Hopfield also explained how the network could be used to solve practical
problems in pattern recognition and optimization. Some of the applications
that Hopfield proposed for his networks were: content-addressable memory
[Hopf82], A/D conversion [TaHo86] and linear programming and optimiza-
tion tasks, such as the traveling salesman problem [HoTa85].
One of Hopfields key contributions was the application of Lyapunov stabil-
ity theory to the analysis of his recurrent networks. He also showed that,
for high-gain amplifiers, the Lyapunov function for his network was a qua-
dratic function, which was minimized by the network. This led to several
design procedures. The idea behind the development of the design tech-
niques was to convert a given task into a quadratic minimization problem,
which the network could then solve.
The Hopfield network is the last topic we will cover in any detail in this
text. However, we have certainly not exhausted all of the important neural
network architectures. In the next chapter we will give you some ideas
about where to go next to explore the subject further.
21
21-37
21 Hopfield Network
Further Reading
197220, 1972.
[AnSi77] J. A. Anderson, J. W. Silverstein, S. A. Ritz and R. S. Jones,
Distinctive features, categorical perception, and probabil-
ity learning: Some applications of a neural model, Psycho-
logical Review, vol. 84, pp. 413451, 1977.
This article describes the brain-state-in-a-box neural net-
work model. It combines the linear associator network with
recurrent connections to form a more powerful autoassocia-
tive system. It uses a nonlinear transfer function to contain
the network output within a hypercube.
Man and Cybernetics, vol. 13, no. 5, pp. 815826, 1983.
Cohen and Grossberg apply LaSalles Invariance Theorem
to the analysis of the stability of competitive neural net-
works. The network description is very general, and the au-
thors show how their analysis can be applied to many
different types of recurrent neural networks.
[Gros67] S. Grossberg, Nonlinear difference-differential equations
in prediction and learning theory, Proceedings of the Na-
This early work of Grossbergs discusses the storage of in-
formation in dynamically stable configurations.
21-38
Further Reading

emergent collective computational properties, Proceedings
of the National Academy of Sciences, vol. 79, pp. 2554
2558, 1982.
This is the original Hopfield neural network paper, which
signaled the resurgence of the field of neural networks. It
describes a discrete-time network that behaves as a con-
tent-addressable memory. Hopfield demonstrates that the
network evolves so as to minimize a specific Lyapunov
function.
[Hopf84] J. J. Hopfield, Neurons with graded response have collec-
tive computational properties like those of two-state neu-
rons, Proceedings of the National Academy of Sciences, vol.
81, pp. 30883092, 1984.
Hopfield demonstrates how an analog electrical circuit can
function as a model for a large network of neurons with a
graded response. The Lyapunov function for this network
is derived and is used to design a network for use as a con-
tent-addressable memory.
[HoTa85] J. J. Hopfield and D. W. Tank, Neural computation of de-
cisions in optimization problems, Biological Cybernetics,
vol. 52, pp. 141154, 1985.
This article describes the application of Hopfield networks
to the solution of optimization problems. The traveling
salesman problem, in which the length of a trip through a
number of cities with only one visit to each city is mini-
mized, is mapped onto a Hopfield network.
pendently.
21
21-39
21 Hopfield Network
[LiMi89] J. Li, A. N. Michel and W. Porod, Analysis and synthesis

of a class of neural networks: Linear systems operating on
a closed hypercube, IEEE Transactions on Circuits and
Systems, vol. 36, no. 11, pp. 14051422, November 1989.
This article investigates a class of neural networks de-
scribed by first-order linear differential equations defined
on a closed hypercube (Hopfield-like networks). Wanted
and unwanted equilibrium points fall at the corners of the
cube. The authors discuss design procedures that minimize
the number of spurious equilibrium points.
Biophysics., vol. 5, pp. 115133, 1943.
neuron in which a weighted sum of input signals is com-
ron fires.
[TaHo86] D. W. Tank and J. J. Hopfield, Simple neural optimiza-
tion networks: An A/D converter, signal decision circuit
and a linear programming circuit, IEEE Transactions on
Circuits and Systems, vol. 33, no. 5, pp. 533541, 1986.
The authors describe how Hopfield neural networks can be
designed to solve certain optimization problems. In one ex-
ample the Hopfield network implements an analog-to-digi-
tal conversion.
21-40
Exercises
Exercises
E21.1 In the Hopfield network example starting on page 18-8 we used a gain of
= 1.4 . Figure 21.3 displays the Lyapunov function for that example. The
high-gain Lyapunov function for the example is shown in Figure 21.9.
i. Show that the minima of the Lyapunov function for this example
will be located at points where n 1 = n 2 = f n 1 = f n 2 . (Use Eq.
(21.42) and set the gradient of V a to zero.)
ii. Investigate the change in location of the minima as the gain is var-
ied from = 0.1 to = 10 .
2+2
ans = iii. Sketch the contour plot for several different values of gain in this in-
4 terval. You will probably need to use MATLAB for this.
E21.2 In Problem P21.3 we used the supervised Hebb rule to design a Hopfield
network to recognize the following patterns:
p1 = 1 p2 = 1 .
1 1
If we use another design rule [LiMi89], we find the following weight matrix
and bias
W = 1 0 b = 0 .
0 10 11
i. Graph the contour plot for the high-gain Lyapunov function, if this
weight matrix and bias are used.
ii. Discuss the difference between the performance of this Hopfield
network and the one designed in Problem P21.3.
2+2
ans = iii. Write a MATLAB M-file to simulate the Hopfield network. Use the
4 ode45 routine. Plot the responses of this network for several initial
conditions.
E21.3 A Hopfield network has the following high-gain Lyapunov function:
1 2 2
V a = --- a 1 + 2a 1 a 2 + 4 a 2 + 6a 1 + 10a 2 .
2
i. Find the weight matrix and bias vector for this network.
ii. Find the gradient and Hessian for V a .
21-41
21 Hopfield Network
iii. Sketch a contour plot of V a .

iv. Find the stationary point(s) for V a . Use the corollary to LaSalles
Invariance Theorem to find as much information as you can about
basins of attraction for any stable equilibrium points.
E21.4 In Problem P21.5 we demonstrated how a Hopfield network could be de-

signed to operate as an A/D converter.
i. Sketch the contour plot of the high-gain Lyapunov function for the
two-bit A/D converter network using an input value of y = 0.5 . Lo-
cate the minimum points.
ii. Repeat part (i) for an input value of y = 2.5 .
iii. Use the answers to parts (i) and (ii) to explain how the network will
operate. Will the network perform the A/D conversion correctly?
E21.5 Assume the binary prototype vectors

T T
p1 = 1 1 1 1 , p2 = 1 1 1 1 .
i. Design a continuous-time Hopfield network (specify connection

weights and biases only) to recognize these patterns, using the Hebb
rule.
ii. Find the Hessian matrix of the high-gain Lyapunov function for this
network. What are the eigenvalues and eigenvectors of the Hessian
matrix? (This requires very little computation.)
iii. Assuming large gain, what are the stable equilibrium points for this
Hopfield network?
E21.6 Repeat Exercise E21.5 for the following prototype vectors.

T T
i. p 1 = 1 1 1 1 , p 2 = 1 1 1 1 .
T T
ii. p 1 = 1 1 1 1 , p 2 = 1 1 1 1 .
T T
iii. p 1 = 1 1 1 1 1 1 , p 2 = 1 1 1 1 1 1 .
E21.7 Consider a high-gain Hopfield network with weight matrix and bias given
by:
W = 1 3 and b = 3 .
3 9 1
21-42
Exercises
i. Sketch a contour plot of the high-gain Lyapunov function for this

network.
ii. If the network is given the following initial condition, where will the
network converge?
a0 = 0.5
0.5
E21.8 Design a high-gain Hopfield network (give the weights and the biases) with
only one stable equilibrium point:
1
1
Explain your procedure, and show all steps. (Do not use the Hebb rule.)
E21.9 Consider a high-gain Hopfield network with weight matrix and bias given
by:
W = 1 1 and b = 1 .
1 1 1
i. Sketch a contour plot of the high-gain Lyapunov function for this

network.
ii. Assuming a large gain, what are the stable equilibrium points for
this Hopfield network? What can you say about the basins of attrac-
tion for these stable equilibrium points? Explain your answers.
E21.10 Repeat E21.9 for the following weight and bias:
W = 1 0 and b = 1 .
0 1 1
2+2
ans = E21.11 In Exercise E7.11 we asked the question: How many prototype patterns
4 can be stored in one weight matrix? Repeat that problem using the
Hopfield network. Begin with the digits 0 and 1. (The digits are shown
at the end of this problem.) Add one digit at a time up to 6, and test how
often the correct digit is reconstructed after randomly changing 2, 4 and 6
pixels.
i. First use the Hebb rule to create the weight matrix for the digits 0
21-43
21 Hopfield Network
and 1. Then randomly change 2 pixels of each digit and apply the
noisy digits to the network. Repeat this process 10 times, and record
the percentage of times in which the correct pattern (without noise)
is produced at the output of the network. Repeat as 4 and 6 pixels of
each digit are modified. The entire process is then repeated when
the digits 0, 1 and 2 are used. This continues, one digit at a
time, until you test the network when all of the digits 0 through
6 are used. When you have completed all of the tests, you will be
able to plot three curves showing percentage error versus number of
digits stored, one curve each for 2, 4 and 6 pixel errors.
ii. Repeat part (i) using the pseudoinverse rule (see Chapter 7), and
compare the results of the two rules.
iii. For extra credit, repeat part (i) using the method described in
[LiMi89]. In that paper it is called Synthesis Procedure 5.1.
p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
21-44
Objectives

Objectives 22-1
Pre-Training Steps 22-3
Selection of Data 22-3
Data Preprocessing 22-5
Weight Initialization 22-13
Choice of Training Algorithm 22-14
Stopping Criteria 22-14
Choice of Performance Function 22-16
Multiple Training Runs and Committees of Networks 22-17
Post-Training Analysis 22-18
Fitting 22-18
Pattern Recognition 22-21
Clustering 22-23
Prediction 22-24
Overfitting and Extrapolation 22-27
Sensitivity Analysis 22-28
Epilogue 22-30
Objectives
Previous chapters have focused on particular neural network architectures
and training rules, with an emphasis on fundamental understanding. In
this chapter, we will discuss some practical training tips that apply to a va-
riety of networks. No derivations are provided for the techniques that are
presented here, but we have found these methods to be useful in practice.
There will be three basic sections in this chapter. The first section describes
things that need to be done prior to training a network, such as collecting
and preprocessing data and selecting the network architecture. The second
section addresses network training itself. The final section considers post-
training analysis.
22
22-1
Theory and Examples

In previous chapters, we have discussed a variety of neural network archi-
tectures and learning rules. Those chapters have placed special emphasis
on the fundamental concepts behind each network. In this chapter, we will
concentrate on practical aspects of training neural networks. Theoretical
aspects and practical aspects are not mutually exclusive. It is only by com-
bining a deep knowledge of network fundamentals with practical experi-
ence in using neural networks that we can get the most out of this
technology.
Figure 22.1 illustrates the neural network training process. It is an itera-
tive procedure that begins by collecting data and preprocessing it to make
training more efficient. At this stage, the data also needs to be divided into
training/validation/testing sets (see Chapter 13). After the data is selected,
we need to choose the appropriate network type (multilayer, competitive,
dynamic, etc.) and architecture (e.g., number of layers, number of neurons).
Then we select a training algorithm that is appropriate for the network and
the problem we are trying to solve. After the network is trained, we want
to analyze the performance of the network. This analysis may lead us to
discover problems with the data, the network architecture, or the training
algorithm. The entire process is then iterated until the network perfor-
mance is satisfactory.
Collect/Preprocess Select Network Select Training

Data Type/Architecture Algorithm
Initialize Weights
Analyze Network &
Performance Train Network
Use Network
Figure 22.1 Flowchart of Neural Network Training Process

In the remainder of this chapter, we will discuss each part of the training
process in some detail. We have divided this material into three major sec-
tions: Pre-Training Steps, Training the Network, and Post-Training Anal-
ysis.
22-2
Pre-Training Steps
Before we dig into the details of training, it is worth making a preliminary

comment. Before beginning the neural network training process, you
should first determine if a neural network is needed to solve your problem,
or if some simpler linear technique might be adequate. For example, there
is no need to use a neural network for a fitting problem, if standard linear
regression will produce a satisfactory result. The neural network tech-
niques provide additional power, but at the expense of more challenging
training requirements. When linear methods will work, they are the first
choice.
Pre-Training Steps
There are a number of steps that need to be performed before the network
is trained. They can be grouped into three categories: Selection of Data,
Data Preprocessing, and Choice of Network Type and Architecture.
Selection of Data
It is generally difficult to incorporate prior knowledge into a neural net-
work, therefore the network will only be as good as the data that is used to
train it. Neural networks represent a technology that is at the mercy of the
data. The training data must span the full range of the input space for
which the network will be used. As we discussed in Chapter 13, there are
training methods we can use to insure that the network interpolates accu-
rately throughout the range of the data provided (generalizes well). How-
ever, it is not possible to guarantee network performance when the inputs
to the network are outside the range of the training set. Neural networks,
like other nonlinear black box methods, do not extrapolate well.
It is not always easy to be sure that the input space is adequately sampled
by the training data. For simple problems, in which the dimension of the
input vector is small, and each element of the input vector can be chosen
independently, we can sample the input space using a grid. However, these
conditions are not often satisfied. For many problems, the dimension of the
input space is large, which precludes the use of grid sampling. In addition,
it is often the case that the input variables are dependent. For example,
consider Figure 22.2. The shaded area represents the range over which the
two inputs can vary. Even though each variable can range from -1 to 1, we
would not need to create a grid in which both variables vary throughout
their range, as shown by the dots in Figure 22.2. The network only needs
to fit the function in the shaded area, since this is where the network will
be used. It would be inefficient to fit the network outside the range of its
use. This is especially true when the input dimension is large.
22-3
p2
p1
Figure 22.2 Input Range With Dependent Input Variables

It may not be possible to precisely define the active region of the input
space. However, we can often collect data during standard operation of the
system we are trying to model. In some cases, we have complete control
over the design of the experiment during which data is collected. In these
cases, we must be sure that the experimental setup drives the system
through all conditions for which we plan to use the network.
How can we be sure that the input space has been adequately sampled by
the training data? This is difficult to do prior to training, and there are
many cases in which we have no control over the data collection process
and must use whatever data is available. We will come back to this subject
in the Post-Training Analysis section on page 22-18. By analyzing the
trained network, we can often tell if the training data was sufficient. In ad-
dition, we can use techniques that indicate when a network is being used
outside the range of the data with which it was trained. This will not im-
prove the network performance, but it will prevent us from using a network
in situations where it is not reliable.
After collecting the data, we generally divide it into three sets: training,
validation and testing. As we discussed in Chapter 13, the training set will
generally make up approximately 70% of the full data set, with validation
and testing making up approximately 15% each. It is important that each
of these sets be representative of the full data set that the validation and
test sets cover the same region of the input space as the training set. The
simplest method for dividing the data is to select each set at random from
the full data set. This usually produces a good result, but it is best to review
the division to check for major differences between the sets. It is also pos-
sible in the post-training analysis to detect problems in the data division.
We will have more to say about that later.
A final question we must ask about the selection of data is Do we have
enough data? This is difficult to answer, especially before we train the net-
22-4
Pre-Training Steps
work. The amount of data that is required depends on the complexity of the
underlying function that we are trying to approximate (or the complexity
of the decision boundaries that we are trying to implement). If the function
to be approximated is very complex, with many inflection points, then this
requires a large amount of data. If the function is very smooth, then the
data requirements are significantly reduced (unless the data is very noisy).
The choice of the data set size is closely related to the choice of the number
of neurons in the neural network. This is discussed in the Choice of Net-
work Architecture section on page 22-8. Of course, we generally dont know
how complex the underlying function is before we begin training the net-
work. For this reason, as we discussed earlier, the entire neural network
training process is iterative. At the completion of training we will analyze
the performance of the network. The results of that analysis can help us de-
cide if we have enough data or not.
Data Preprocessing
The main purpose of the data preprocessing stage is to facilitate network
training. Data preprocessing consists of such steps as normalization, non-
linear transformations, feature extraction, coding of discrete inputs/tar-
gets, handling of missing data, etc. The idea is to perform preliminary
processing of the data to make it easier for the neural network training to
extract the relevant information.
For example, in multilayer networks, sigmoid transfer functions are often
used in the hidden layers. These functions become essentially saturated
when the net input is greater than three ( exp 3 0.05 ). We dont want
this to happen at the beginning of the training process, because the gradi-
ent will then be very small. In the first layer, the net input is a product of
the input times the weight plus the bias. If the input is very large, then the
weight must be small in order to prevent the transfer function from becom-
ing saturated. It is standard practice to normalize the inputs before apply-
ing them to the network. In this way, initializing the network weights to
small random values guarantees that the weight-input product will be
small. Also, when the input values are normalized, the magnitudes of the
weights have a consistent meaning. This is especially important when us-
ing regularization, as described in Chapter 13. Regularization requires the
weight values to be small. However, small is a relative term; if the input
values are very small, we need large weights to produce a significant net
input. Normalizing the inputs clarifies the meaning of small weights.
Normalization There are two standard methods for normalization. The first method nor-
malizes the data so that they fall into a standard range typically -1 to 1.
This can be done with
n min max min
p = 2p p ./ p p 1, (22.1)
min
where p is the vector containing the minimum values of each element of
max
the input vectors in the data set, p contains the maximum values, ./
22-5
n
represents an element-by-element division of two vectors, and p is the re-
sulting normalized input vector.
An alternative normalization procedure is to adjust the data so that they
have a specified mean and variance typically 0 and 1. This can be done
with the transformation
n mean std
p = p p ./p , (22.2)
mean std
where p is the average of the input vectors in the data set, and p is
the vector containing the standard deviations of each element of the input
vectors.
Generally, the normalization step is applied to both the input vectors and
the target vectors in the data set.
In addition to normalization, which involves a linear transformation, non-
Nonlinear Transformation linear transformations are sometimes also performed as part of the pre-
processing stage. Unlike normalization, which is a standard process that
can be applied to any data set, these nonlinear transformations are case-
specific. For example, many economic variables show a logarithmic depen-
dence [BoJe94]. In that case, it might be appropriate to take the logarithm
of the input values before applying them to the neural network. Another
example is molecular dynamics simulation [RaMa05], in which atomic forc-
es are calculated as functions of distances between atoms. Since it is known
that the forces are inversely related to the distances, we might perform the
reciprocal transformation on the inputs, before applying them to the net-
work. This represents one way of incorporating prior knowledge into neural
network training. If the nonlinear transformation is cleverly chosen, it can
make the network training more efficient. The preprocessing will off-load
some of the work required of the neural network in finding the underlying
transformation between inputs and targets.
Feature Extraction Another data preprocessing step is called feature extraction. This generally
applies to situations in which the dimension of the raw input vectors is very
large and the components of the input vector are redundant. The idea of
feature extraction is to reduce the dimension of the input space by calculat-
ing a small set of features from each input vector, and using the features
as the input to the neural network. For example, neural networks can be
used to analyze EKG (electrocardiogram) signals to identify heart prob-
lems [HeOh97]. The EKG might involve 12 or 15 signals (leads) measured
over several minutes at a high sampling rate. This is too much data to ap-
ply directly to the neural network. Instead, we would extract certain fea-
tures from the EKG signal, such as average time intervals between certain
waveforms, average amplitudes of certain waves, etc. (See Chapter 25.)
There are also certain general-purpose feature extraction methods. One of
Principal Components these is the method of principal component analysis (PCA) [Joll02]. This
method transforms the original input vectors so that the components of the
22-6
Pre-Training Steps
transformed vectors are uncorrelated. In addition, the components of the

transformed vector are ordered such that the first component has the
greatest variance, the second component has the next greatest variance,
etc. We generally keep only the first few components of the transformed
vector, which account for most of the variance in the original vector. This
results in a large reduction in the dimension of the input vector, if the orig-
inal components are highly correlated. The drawback of using PCA is that
it only considers linear relationships between the components of the input
vector. When reducing the dimension using a linear transformation, we
might lose some nonlinear information. Since the main purpose of using
neural networks is to gain the power of their nonlinear mapping capabili-
ties, we should be careful when using principal components to reduce the
input dimension before applying the inputs to the neural network. There is
a nonlinear version of PCA, called kernel PCA [ScSm99].
Another important preprocessing step is needed whenever the inputs or
targets take on only discrete values. For example, in pattern recognition
problems, each target will represent one of a finite number of classes. In
Coding the Targets these cases we need to have a procedure for coding the inputs or targets. If
we have a pattern recognition problem in which there are four classes,
there are at least three common ways in which we could code the targets.
First, we can have scalar targets that take on four possible values (e.g., 1,
2, 3, 4). Second, we can have two-dimensional targets, which represent a
binary code of the four classes (e.g., (0,0), (0,1), (1,0), (1,1)). Third, we can
have four-dimensional targets, in which only one neuron at a time is active
(e.g., (1,0,0,0), (0,1,0,0), (0,0,1,0), (0,0,0,1)). The third method tends to yield
the best results in our experience. (Note that discrete inputs can be coded
in the same ways as discrete targets.)
When coding the target values, we also need to consider the transfer func-
tion that is used in the output layer of the network. For pattern recognition
problems, we would typically use sigmoid functions: log-sigmoid or tan-
gent-sigmoid. If we use the tangent-sigmoid in the last layer, which is more
common, then we might consider assigning target values to -1 or 1, which
represent the asymptotes of the function. However, this tends to cause dif-
ficulties for the training algorithm, which tries to saturate the sigmoid
function to meet the target value. It is better to assign target values at the
point where the second derivative of the sigmoid function is maximum (see
[LeCu98]). For the tangent-sigmoid function, this occurs when the net in-
put is -1 and 1, which corresponds to output values of -0.76 and +0.76.
Another transfer function that is used in the output layer of multilayer pat-
Softmax tern recognition networks is the softmax function. This transfer function
has the form
S
a i = f n i = exp n i exp nj . (22.3)

j=1
22
22-7
The outputs of the softmax transfer function can be interpreted as the prob-
abilities associated with each class. Each output will fall between 0 and 1,
and the sum of the outputs will equal 1. See Chapter 24 for an example ap-
plication of the softmax transfer function.
Missing Data Another practical issue to consider is missing data. It is often the case, es-
pecially when dealing with economic data, for example, that some of the
data is missing. For instance, we might have an input vector containing 20
economic variables that are collected at monthly intervals. There may be
some months in which one or two of the 20 variables were not collected
properly. The simplest solution to this problem would be to throw out the
data for any month in which any of the variables were missing. However,
we might be very limited in the amount of data available, and it might be
very expensive to collect additional data. In these cases, we would like to
make full use of any data that we have, even if it is incomplete.
There are several strategies for dealing with missing data. If the missing
data occurs in an input variable, one possibility is to replace the missing
value with the average value for that particular input variable. At the same
time, we could add an additional flag element to the input vector that
would indicate that missing data for that input variable had been replaced
with the average. This additional element of the input vector could be as-
signed the value 1 when the input variable was available, and 0 when the
input variable was missing for that case. This would provide the neural
network with information about which variables were missing. An addi-
tional flag element would be added to the input vector for every input vari-
able that contained missing points.
If the missing data occurs in an element of the target, then the performance
index can be modified so that errors associated with the missing target val-
ues are not included. All known target values will contribute to the perfor-
mance index, but missing target values will not.
Choice of Network Architecture

The next step in the network training process is the choice of network ar-
chitecture. The basic type of network architecture is determined by the
type of problem we wish to solve. Once the basic architecture is chosen, we
need to decide such specific details as how many neurons and layers we
want to use, how many outputs the network should have, and what type of
performance function we want to use for training.
Choice of Basic Architecture

The first step in choosing the architecture is to define the problem that we
are trying to solve. For this chapter, we will limit our discussion to four
types of problems: fitting, pattern recognition, clustering and prediction.
Fitting Fitting is also referred to as function approximation or regression. In fitting
problems, you want a neural network to map between a set of inputs and a
22-8
Pre-Training Steps
corresponding set of targets. For example, a realtor might want to estimate

home prices from such input variables as tax rate, pupil/teacher ratio in lo-
cal schools and crime rate. An automotive engineer might want to estimate
engine emission levels based on measurements of fuel consumption and
speed. A physician might want to predict a patient's body fat level based on
body measurements. For fitting problems, the target variable takes on con-
tinuous values. (For an example of training a neural network for a fitting
problem, see Chapter 23.)
The standard neural network architecture for fitting problems is the mul-
tilayer perceptron, with tansig neurons in the hidden layers, and linear
neurons in the output layer. The tansig transfer function is generally pre-
ferred over the logsig transfer function in the hidden layers for the same
reason that inputs are normalized. It produces outputs (which are inputs
to the next layer) that are centered near zero, whereas the logsig transfer
function always produces positive outputs. For most fitting problems, a sin-
gle hidden layer is sufficient. If the results with one hidden layer are not
satisfactory, two layers are sometimes used. It would be rare in a standard
fitting problem to use more than two hidden layers, although, for very dif-
ficult problems, deep networks, with many layers, have been used. Linear
transfer functions are used in the output layer for fitting problems, because
the target output is a continuous variable. As we saw in Chapter 11, a two
layer network with sigmoid transfer functions in the hidden layer and lin-
ear transfer functions in the output layer is a universal approximator.
Radial basis networks can also be used for fitting problems. The Gaussian
transfer function is most commonly used in the hidden layer for these net-
works, with linear transfer functions in the output layer.
Pattern Recognition Pattern recognition is also referred to as pattern classification. In pattern
recognition problems, you want a neural network to classify inputs into a
set of target categories. For example, a wine dealer might want to recognize
the vineyard that a particular bottle of wine came from, based on a chemi-
cal analysis of the wine. A physician might want to classify a tumor as be-
nign or malignant, based on uniformity of cell size, clump thickness and
mitosis.
In addition to fitting problems, multilayer perceptrons can be used for pat-
tern recognition. The main difference between a fitting network and a pat-
tern recognition network is the transfer function used in the output layer.
For pattern recognition problems, we generally use a sigmoid function in
the output layer. The radial basis function network can also be used for pat-
tern recognition.
For an example of training a neural network for a pattern recognition prob-
lem, see Chapter 25.
Clustering Clustering, or segmentation, is another use for neural networks. In cluster-
ing problems, you want a neural network to group data by similarity. For
example, businesses may wish to perform market segmentation, which is
22-9
done by grouping people according to their buying patterns. Computer sci-

entists may want to perform data mining by partitioning data into related
subsets. Biologists may wish to perform bioinformatic analyses, such as
grouping genes with related expression patterns.
Any of the competitive networks described in Chapter 16 could be used for
clustering. The self-organizing feature map (SOFM) is the most popular
network for clustering. The main advantage of the SOFM is that it allows
visualization of high-dimensional spaces.
For an example of training a neural network for a clustering problem, see
Chapter 26.
Prediction Prediction also falls under the categories of time series analysis, system
identification, filtering or dynamic modeling. The idea is that we wish to
predict the future value of some time series. An equities trader might want
to predict the future value of some security. A control engineer might want
to predict a future value of the concentration of some chemical, which is the
output of a processing plant. A power systems engineer might want to pre-
dict outages on the electric grid.
Prediction requires the use of dynamic neural networks, as discussed in
Chapter 14. The specific form of the network will depend on the particular
application. The simplest network for nonlinear prediction is the focused
time-delay neural network, which is shown in Figure 22.3. This is part of a
general class of dynamic networks, called focused networks, in which the
dynamics appear only at the input layer of a static multilayer feedforward
network. This network has the advantage that it can be trained using static
backpropagation algorithms, since the tapped-delay-line at the input of the
network can be replaced with an extended vector of delayed values of the
input.
p1(t) 1 2
T a (t) a (t)
1
D IW 1,1
n (t) LW 2,1
2
n (t) S2x1
L S1x1
1
S x (R d) 1
S2xS1
2
S x1
1 b
1 1
S x1 1 b 2
1 1
S2
1
R S x1 S S2x1
Figure 22.3 Focused Time-Delay Neural Network

For problems of dynamic modeling and control, the NARX network (Non-
linear AutoRegressive model with eXogenous input) is popular. This net-
work is shown in Figure 22.4. The input signal could represent, for
22-10
Pre-Training Steps
example, the voltage applied to a motor, and the output could represent the
angular position of a robot arm. As with the focused time-delay neural net-
work, the NARX network can be trained with static backpropagation. The
two tapped-delay-lines can be replaced with extended vectors of delayed in-
puts and targets. We can use targets, instead of feeding back the network
outputs (which would require dynamic backpropagation for training), be-
cause the output of the network should match the targets when training is
complete.This is discussed in more detail in Chapter 27.
p1(t) n1(t) a1(t) a2(t)

T
D IW1,1 LW2,1 2
L S1x1 2 1
n (t) S2x1
1
S1xR S xS
R x1 S2x1
1 b 1 1
S x1 1 b
2
1 1
S2
2
R S1x1 S S x1
T
1,3
D LW
L
Figure 22.4 NARX Neural Network

There are many other types of dynamic networks that could be used for pre-
diction, but the focused time delay network and the NARX network are the
simplest of their type.
For an example of training a neural network for a prediction problem, see
Chapter 27.
Selection of Architecture Specifics

After the basic network structure is chosen, we want to select the specifics
of the architecture (e.g., the number of layers, the number of neurons, etc.).
In some cases, the basic architecture choice will automatically determine
the number of layers. For example, if the SOFM is used for clustering, then
the network will have one layer. In the case of the multilayer network,
which can be used for fitting or pattern recognition, the number of hidden
layers is not determined by the problem, since any number of hidden layers
is possible. The standard procedure is to begin with a network with one hid-
den layer. If the performance of the two-layer network is not satisfactory,
then a three-layer network can be used. It would be unusual to use more
than two hidden layers. The training becomes more difficult when multiple
hidden layers are used. This is because each layer performs a squashing op-
eration, as sigmoid functions are used in the hidden layers. This causes the
22-11
derivatives of the performance function with respect to weights in the early

layers to be quite small, which can cause slow convergence for steepest de-
scent optimization. For very difficult problems, however, deep networks,
with several hidden layers, can be used. Typically, parallel or GPU comput-
ing is required to train deep multilayer networks within a reasonable
amount of time.
We also need to select the number of neurons in each layer. The number of
neurons in the output layer is the same as the size of the target vector. The
numbers of neurons in the hidden layers are determined by the complexi-
ties of the function that is being approximated or the decision boundaries
that are being implemented. Unfortunately, we dont normally know how
complex the problem is until we try to train the network. The standard pro-
cedure is to begin with more neurons than necessary and then to use early
stopping or Bayesian regularization to prevent overfitting, as was de-
scribed in Chapter 13.
The principal drawback to having too many neurons is that the network
may overfit the data. If we use early stopping or Bayesian regularization,
then we can prevent overfitting. However, there may be some situations in
which we are concerned with the computation time or space required by the
network (e.g., for real-time implementation on microcontrollers, VLSI or
FPGAs). In these cases we want to find the simplest network that will fit
the data. If you use Bayesian regularization, the effective number of pa-
rameters can be used to determine how many neurons to use. If, after train-
ing, the effective number of parameters is much less than the total number
of parameters in the network, then the number of neurons can be reduced,
and the network retrained. It is also possible to use pruning methods to
eliminate neurons or weights in the network.
The number of neurons in the last layer is equal to the number of elements
in the target vector. However, when there are multiple targets, we have a
choice to make. We can have one network with multiple outputs, or we can
have multiple networks, each with one output. For example, neural net-
works have been used to estimate LDL, VLDL and HDL cholesterol levels,
based on a spectral analysis of the blood. It is possible to have one neural
network with three neurons in the output layer to estimate all three cho-
lesterol levels, or we could have three neural networks, with each one esti-
mating only one of the three components. Theoretically, both methods
should work, but in practice one method may work better than another. We
generally start with one multi-output network, and then use multiple sin-
gle-output networks if the original results are not satisfactory.
Another architectural choice is the size of the input vector. This is often a
simple choice, which is determined by the training data. However, there
are times when input vectors in the training data have redundant or irrel-
evant elements. When the dimension of the potential input vector is very
large, it is sometimes advantageous to eliminate redundant or irrelevant
elements. This can reduce the required computation and can assist in pre-
22-12
Training the Network
Input Selection venting overfitting during training. The input selection process for nonlin-
ear networks can be quite difficult, and there is no perfect solution. The
Bayesian regularization method (Eq. (13.23)) can be modified to assist in
input selection. It is possible to have different parameters for different
sets of weights. For example, we can let each column of the weight matrix
in the first layer of a multilayer network have its own parameter. If a
given element of the input vector is irrelevant, then the corresponding
parameter would become large and force all elements of that column of the
weight matrix to be small. That element could then be eliminated from the
input vector.
Another technique that can assist in pruning the input vector is a sensitiv-
ity analysis of the trained network. In the Sensitivity Analysis section on
page 22-28 we discuss this technique.

After the data has been prepared, and the network architecture has been
selected, we are ready to train the network. In this section, we will discuss
some of the decisions that need to be made as part of the training process.
This includes the method for initializing the weights, the training algo-
rithm, the performance index, and the criterion for stopping training.
Weight Initialization
Before training the network, we need to initialize the weights and biases.
The method we use will depend on the type of network. For multilayer net-
works, the weights and biases are generally set to small random values
(e.g., uniformly distributed between -0.5 and 0.5, if the inputs are normal-
ized to fall between -1 and 1). As we discussed in Chapter 12, if we set the
weights and biases to zero, the initial condition may fall on a saddle point
of the performance surface. If we make the initial weights large, the initial
condition can fall on a flat part of the performance surface, caused by sat-
uration of the sigmoid transfer functions.
There is another approach to setting the initial weights and biases for a
two-layer network. It was introduced by Widrow and Nguyen [WiNg90].
The idea is to set the magnitude of the weights in the first layer so that the
1
linear region of each sigmoid function covers 1 S of the range of the input.
The biases are then randomly set, so that the center of each sigmoid func-
tion falls randomly in the input space. The details of the method are as fol-
lows (assuming the inputs to the network have been normalized to values
between -1 and 1).
1 1
Set row i of W , iw , to have a random direction and a magnitude of
1 1 1R
iw = 0.7 S .
22-13
1 1
Set b i to a uniform random value between iw and iw .
For competitive networks, the weights can also be set as small random
numbers. Another possibility is to randomly select some of the input vec-
tors in the training set to become initial rows of the weight matrix. In this
way, we can be sure that the initial weights will fall within the range of the
input vectors, so we will be less likely to have dead units, as described in
Chapter 16. For the SOM, dead units are not a problem. The initial neigh-
borhood size is set large enough so that all neurons will have the opportu-
nity to learn during the initial stages of training. This will move all weight
vectors into the appropriate region of the input space. Training can con-
verge faster, however, if rows of the weight matrix are initially placed in
the active input region.
Choice of Training Algorithm

For multilayer networks, we generally use gradient- or Jacobian-based al-
gorithms, as described in Chapter 12. These algorithms can be implement-
ed in either batch mode or sequential (also known as incremental, pattern
or stochastic) mode. For example, in the sequential form of steepest descent
(see Eq. (11.13)) we update the weights after each input is presented to the
network. In batch mode (see page 12-7), all of the inputs are presented to
the network, and the total gradient is computed by summing the gradients
for each input, before the weights are updated. In some situations, the se-
quential form is preferred for example, when on-line or adaptive opera-
tion is required. However, many of the more efficient optimization
algorithms (e.g., conjugate gradient and Newtons methods) are inherently
batch algorithms.
For multilayer networks with up to a few hundred weights and biases that
are being used for function approximation, the Levenberg-Marquardt algo-
rithm (see Eq. (12.31)) is usually the fastest training method. When the
number of weights reaches a thousand or more, the Levenberg-Marquardt
algorithm is not as efficient as some of the conjugate gradient algorithms.
This is mainly because the matrix inverse calculation scales geometrically
with the number of weights. For large networks, the Scaled Conjugate Gra-
dient algorithm of [Moll93] is very efficient. This method is also attractive
for pattern recognition problems. The Levenberg-Marquardt algorithm
does not work as well for pattern recognition, in which the sigmoid transfer
functions in the final layer are operating well outside the linear region.
Of the algorithms that can be implemented in sequential mode, the fastest
are the extended Kalman filter algorithms. These algorithms are closely re-
lated to sequential implementations of the Gauss-Newton algorithm. Un-
like the batch version of Gauss-Newton, they do not require an inversion of
the approximate Hessian matrix. The decoupled extended Kalman filter
implementation of [PuFe97] appears to be the most efficient of these types
of algorithms.
22-14
Stopping Criteria
For most applications of neural networks, the training error never converg-
es identically to zero. The error can reach zero for the perceptron network,
assuming a linearly separable problem, as we showed in Chapter 4. How-
ever, it is unlikely to happen for multilayer networks. For this reason, we
need to have other criteria for deciding when to stop the training.
We can stop the training when the error reaches some specified limit. How-
ever, it is usually difficult to know what an acceptable error level is. The
simplest criterion is to stop the training after a fixed number of iterations.
Because it is also difficult to know how many iterations will be required,
the maximum iteration number is generally set reasonably high. If the
weights have not converged after the maximum number of iterations has
been reached, we can restart training, using the final weights from the first
run as initial conditions for the restart. (We will talk more about how to tell
if a network has converged in the Post-Training Analysis section on page
22-18.)
Another stopping criterion is the norm of the gradient of the performance
index. If this norm reaches a sufficiently small threshold, then the training
can be stopped. Since the gradient should be zero at a minimum of the per-
formance index, this criterion will stop the algorithm when it gets close to
the minimum. Unfortunately, as we have seen in Chapter 12, the perfor-
mance surface for multilayer networks can have many flat regions, where
the norm of the gradient will be small. For this reason, the threshold for
6
the minimum norm should be set to a very small value (e.g., 10 for mean
square error performance indices, with normalized targets), so that the
training does not end prematurely.
We can also stop the training when the reduction in the performance index
per iteration becomes small. As with the norm of the gradient, this criterion
can stop the training too early. During the training of multilayer networks,
the performance can remain almost constant for a number of iterations be-
fore dropping suddenly. When training is complete, it is useful to view the
training performance curve on a log-log scale, as in Figure 22.5, to verify
convergence.
If we are using early stopping, as discussed in Chapter 13, to prevent over-
fitting, then we will stop the training when the performance on the valida-
tion set increases for a set number of iterations. In addition to preventing
overfitting, this stopping procedure also provides a significant reduction in
computation; for most practical problems, the validation error will increase
before any of the other stopping criteria are reached.
As shown in Figure 22.1, neural network training is an iterative process.
Even after the training algorithm has converged, post-training analysis
may suggest that the network be modified and retrained. In addition, sev-
eral training runs should be made for each potential network to ensure that
a global minimum has been reached.
22-15
3
10
2
10
Sum Square Error

1
10
0
10
1
10
2
10
3
10
0 1 2 3
10 10 10 10
Epochs
Figure 22.5 Typical Training Performance Curve

The previous stopping criteria apply mainly to gradient-based training.
When training competitive networks, like the SOFM, there is no explicit
performance index or gradient to monitor for convergence. The training
stops only when the maximum number of iterations has been reached. For
SOFMs, the learning rate and the neighborhood size are decreased over
time. Typically, the neighborhood size is reduced to zero by the completion
of training, so the maximum iteration number determines the end of train-
ing, as well as the rate of decrease in neighborhood size and learning rate.
This is therefore a very important parameter. It is generally chosen to be
more than ten times the number of neurons in the network. This is an ap-
proximate number, and the network needs to be analyzed at the completion
of training to determine if the performance is satisfactory. (This will be dis-
cussed in the Post-Training Analysis section on page 22-18.) The network
may need to be trained several times with different training lengths to
achieve a satisfactory result.
Choice of Performance Function

For multilayer networks, the standard performance index is mean square
error. When all inputs in the training set are equally likely to occur, then
this can be written
Q
1
M
T
F x = ----------
- tq aq tq aq , (22.4)
QS q = 1
or
22-16
M
Q S
1
M i q
2
F x = ----------
- t a i q . (22.5)
QS q = 1 i = 1
The scale factor that occurs outside the sum has no effect on the location of
the optimum weights. Therefore, the sum square error performance index
will produce the same weights as the mean square error performance in-
dex. However, the appropriate scaling can be useful when comparing errors
on data sets of different size.
While mean square error is the most common performance index, there are
others that have been used. For example, we could use mean absolute er-
ror. This would be similar to Eq. (22.5), except that the absolute value of
the error would be used, instead of the square of the error. This perfor-
mance index is generally less sensitive to one or two large errors in the data
set, and is therefore somewhat more robust to outliers than is the mean
square error algorithm. This concept can be extended to any power of the
absolute error, as follows
M
Q S
1
M i q
K
F x = ----------
- t a i q , (22.6)
QS q = 1 i = 1
where K = 2 corresponds to mean square error and K = 1 corresponds to

mean absolute error. The general error given by Eq. (22.6) is referred to as
the Minkowski error.
As we saw in Chapter 13, the mean square performance index can be aug-
mented with the mean square weights, to produce a regularized perfor-
mance index, which is used to prevent overfitting. The Bayesian
regularization algorithm is an excellent training method for preventing
overfitting. It uses a regularized performance index, and uses Bayesian
methods to select the regularization parameter. See Chapter 13 for details.
Mean square error works well for function approximation problems, in
which the target values are continuous. However, in pattern recognition
problems, where the targets take on discrete values, other performance in-
dices might be more appropriate. One performance index that has been
Cross-Entropy proposed for classification problems is cross-entropy [Bish95]. Cross-entro-
py is defined as
M
Q S
a i q
Fx = t i q ln -------
-.
t i q
(22.7)
q = 1i = 1
Here we assume that the target values are 0 and 1, and they identify which
of the two classes the input vector belongs to. The softmax transfer function
22-17
is generally used in the last layer of the neural network, if the cross-entro-
py performance index is used.
As a closing note on the choice of performance index, recall from Chapter
11 that the backpropagation algorithm for computing training gradients
will work for any differentiable performance index. If you change the per-
formance index, you only need to change the initialization of the sensitivi-
ties in the last layer (see Eq. (11.37)).
Multiple Training Runs and Committees of Networks

A single training run may not produce optimal performance, because of the
possibility of reaching a local minimum of the performance surface. It is
best to restart the training at several different initial conditions and select
the network that produces the best performance. Five to ten restarts will
almost always produce a global optimum [HaBo07].
There is another way to perform multiple training runs and make use of all
of the networks that have been trained. This is called the committee of net-
works [PeCo93]. For each training session, the validation set is randomly
selected from the training data, and a random set of initial weights and bi-
ases is chosen. After N networks have been trained, all of the networks are
used together to form a joint output. For function approximation networks,
the joint output can be a simple average of the outputs of each network. For
classification networks, the joint output can be the result of a vote, in which
the class that is chosen by the majority of the networks is selected as the
output of the committee. The performance of the committee will usually be
better than even the best of the individual networks. In addition, the vari-
ation in the outputs of the individual networks can be used to provide error
bars, or confidence levels, for the committee output.
Post-Training Analysis
Before using a trained neural network, we need to analyze it to determine
if the training was successful. There are many techniques for post-training
analysis. We will discuss some of the more common ones. Since these tech-
niques vary, depending on the application, we will organize them according
to these four application areas: fitting, pattern recognition, clustering and
prediction.
Fitting
One useful tool for analyzing neural networks trained for fitting problems
is a regression between the trained network outputs and the corresponding
targets. We fit a linear function of the form
a q = mt q + c + q , (22.8)
22-18
where m and c are the slope and offset, respectively, of the linear function,
t q is a target value, a q is a trained network output, and q is the residual
error of the regression.
The terms in the regression can be computed as follows:
Q
tq t aq a
= q---------------------------------------------
m
=1
-, (22.9)
Q
2
tq t
q=1
c = a m
t, (22.10)
where
Q Q
1
a = ----
Q
Q
1
a q , t = ---
- tq . (22.11)
q=1 q=1
Figure 22.6 shows an example regression analysis. The blue line represents
the linear regression, the thin black line represents the perfect match
a q = t q , and the circles represent the data points. In this example, we can
see that the match is pretty good, although not perfect. The next step would
be to investigate data points that fall far from the regression line. For ex-
Outliers ample, there are two points around t = 27 and a = 17 that seem to be out-
liers. We would investigate these points to see if there was a problem with
the data. This could be a bad data point, or it could be located far from other
training points. In the latter case, we would need to collect more data in
that region.
In addition to computing the regression coefficients, we often also compute
the correlation coefficient between the t q and a q , which is also known as
the R value:
Q
tq t aq a
q=1
R = ---------------------------------------------- , (22.12)
Q 1 s t s a
where
Q Q
st = 1
-------------
Q1
t q t and s a = 1
-------------
Q1
aq a . (22.13)
q=1 q=1
22-19
50
40
30
a
20
10
0
0 5 10 15 20 25 30 35 40 45 50
Figure 22.6 Regression Between Trained Network Outputs and Targets

The R value can generally range from -1 to 1, but we would expect it to be
close to 1 for our neural network application. If R = 1 , then all of the data
points will fall exactly on the regression line. If R = 0 , then the data will
not be concentrated around the regression line, but will be randomly scat-
tered. For the data of Figure 22.6, R = 0.965 . We can see that the data does
not fall exactly on the regression line, but the variation is relatively small.
2
The square of the correlation coefficient, R , is sometimes used instead of
2
R . R represents the proportion of the variability in a data set that is ac-
counted for by the linear regression, and is also referred to as the coefficient
2
of determination. For the data of Figure 22.6, R = 0.931 .
2
When the R or R values are significantly less than 1, then the neural net-
work has not done a good job of fitting the underlying function. A close
analysis of the scatter plot may be helpful in determining problems in the
fit. For example, we might find that when the targets are large there is
more spread in the scatter plot. (This is not the case in Figure 22.6.) We
might also notice that there are fewer data points with large targets. This
would indicate that we need to have more data points in the training set for
these target values.
Recall that the original data set was divided into training, validation (if
early stopping is used) and testing subsets. The regression analysis should
be performed on each subset individually, as well as the full data set. Dif-
ferences between the subsets would indicate overfitting or extrapolation.
For example, if the training set shows accurate fitting, but the validation
and test results are poor, then this would indicate overfitting (which can
sometimes happen, even when early-stopping is used). In this case, we
might reduce the size of the neural network and retrain. If both the train-
ing and validation results are good, but the testing results are poor, then
22-20
this could indicate extrapolation (where the testing data falls outside the
training and validation data). In this case, we need to provide more data
for training and validation. If the results for all three data sets are poor,
then it might be necessary to increase the number of neurons in the net-
work. Another choice is to increase the number of layers in the network. If
you start with a single hidden layer, and the results are poor, then a second
hidden layer could be helpful. First, try more neurons in the single hidden
layer, and then increase the number of layers.
In addition to the regression/scatter plot, another tool that can identify out-
liers is a histogram of the errors, as shown in Figure 22.7. The y-axis rep-
resents the number of errors that falls within each interval on the x-axis.
Here we can see that two errors are greater than 8. These represent the
same two errors that we identified as outliers in Figure 22.6.
140
120
100
80
60
40
20
0
8 6 4 2 0 2 4 6 8 10 12
e = ta
Figure 22.7 Histogram of Network Errors
Pattern Recognition
For pattern recognition problems, the regression analysis is not as useful
as it is for fitting problems, since the target values are discrete. However,
Confusion Matrix there is an analogous tool - the confusion (or misclassification) matrix. The
confusion matrix is a table whose columns represent the target class and
whose rows represent the output class. For example, Figure 22.8 shows a
sample confusion matrix in which there were 214 data points. There were
41 input vectors that belonged to Class 1 and were correctly classified as
Class 1. There were 162 input vectors that belonged to Class 2 and were
correctly classified as Class 2. The correctly classified inputs show in the
diagonal cells of the confusion matrix. The off-diagonal cells show misclas-
sified inputs. The lower left cell shows that four inputs from Class 1 were
misclassified by the network as Class 2. If Class 1 is considered a positive
False Negative outcome, then the lower left cell represents false negatives, which are also
22-21
called Type II errors. The upper right cell shows that one input from Class
2 was misclassified by the network as Class 1. This would be considered a
False Positive false positive or a Type I error.
Confusion Matrix
47 1 97.9%
1
22.0% 0.5% 2.1%
Output Class
4 162 97.6%
2
1.9% 75.7% 2.4%
92.2% 99.4% 97.7%

7.8% 0.6% 2.3%
1 2
Target Class
Figure 22.8 Sample Confusion Matrix

Another useful tool for analyzing a pattern recognition network is called
ROC Curve the Receiver Operating Characteristic (ROC) curve. To create this curve, we
take the output of the trained network and compare it against a threshold
which ranges from -1 to +1 (assuming a tansig transfer function in the last
layer). Inputs that produce values above the threshold are considered to be-
long to Class 1, and those with values below the threshold are considered
to belong to Class 2. For each threshold value, we count the fraction of true
positives and false positives in the data set. This pair of numbers produces
one point on the ROC curve. As the threshold is varied, we trace the com-
plete curve, as shown in Figure 22.9.
The ideal point for the ROC curve to pass through would be (0,1), which
would correspond to no false positives and all true positives. A poor ROC
curve would represent a random guess, which is represented by the diago-
nal line in Figure 22.9, which passes through the point (0.5,0.5).
22-22
0.9
0.8
0.7
True Positive Rate

0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
False Positive Rate
Figure 22.9 Receiver Operating Characteristic Curve
Clustering
The SOM is the most commonly used network for clustering. There are sev-
Quantization Error eral measures of SOM performance. One is quantization error. This is the
average distance between each input vector and the closest prototype vec-
tor. It measures the map resolution. This can be made artificially small, if
we use a large number of neurons. If there are as many neurons as input
vectors in the data set, then the quantization error could be zero. This
would represent overfitting. If the number of neurons is not significantly
smaller than the number of input vectors, then the quantization error is
not meaningful.
Topographic Error Another measure of SOM performance is topographic error. This is the pro-
portion of all input vectors for which the closest prototype vector and the
next closest prototype vector are not neighbors in the feature map topology.
Topographic error measures the preservation of the topology. In a well-
trained SOM, prototypes that are neighbors in the topology should also be
neighbors in the input space. In this case, the topographic error should be
zero.
Distortion Measure The performance of the SOM can also be assessed by the distortion mea-
sure:
Q S
h ic
2
Ed = q iw pq , (22.14)
q = 1i = 1
where h ij is the neighborhood function, and c q is the index of the prototype

that is closest to the input vector p q :
22
22-23
c q = arg min j jw p q . (22.15)
For the simplest neighborhood function, h ij is equal to 1 if prototype i is

within some pre-specified neighborhood radius of prototype j, and equal to
zero otherwise. It is also possible to have neighborhood functions that de-
crease continuously, such as the Gaussian function:
2
iw jw
h ij = exp --------------------------
- , (22.16)

2
2d
where d is the neighborhood radius.
Prediction
As we discussed earlier, one application of neural networks is the predic-
tion of the future values of some time series. For prediction problems, we
use dynamic networks, such as the focused time-delay neural network
shown in Figure 22.3. There are two important concepts that are used
when analyzing a trained prediction network:
1. the prediction errors should not be correlated in time, and
2. the prediction errors should not be correlated with the input sequence.
If the prediction errors were correlated in time, then we would be able to
predict the prediction errors and, therefore, improve our original predic-
tion. Also, if the prediction errors were correlated with the input sequence,
then we would also be able to use this correlation to predict the errors.
In order to test the correlation of the prediction errors in time, we can use
Autocorrelation Function the sample autocorrelation function:
Q
1
R e = ------------
Q e t e t + . (22.17)
t=1
White Noise If the prediction errors are uncorrelated (white noise), then we would ex-
pect R e to be close to zero, except when = 0 . To determine if R e is
close to zero, we can set an approximate 95% confidence interval [BoJe96]
using the range
2R e 0 2R e 0
- R e ----------------
---------------- -. (22.18)
Q Q
We can say that e t is white, if R e satisfies Eq. (22.18) for 0 . This

concept is illustrated in Figure 22.10 and Figure 22.11. Figure 22.10 shows
a sample autocorrelation function for the prediction errors of a network
that has not been adequately trained. We can see that the autocorrelation
22-24
function does not fall totally within the bounds defined by Eq. (22.18),
which are indicated by the dashed lines in the figure. Figure 22.11 shows
the corresponding autocorrelation function when a network has been suc-
cessfully trained. R e falls within the bounds, except at = 0 .
Correlation in the prediction errors can indicate that the length of the
tapped delay lines in the network should be increased.
0.2
0.1
0.1
40 20 0 20 40
Figure 22.10 R e for Inadequately Trained Network
0.03
0.02
0.01
40 20 0 20 40
Figure 22.11 R e for Successfully Trained Network
To test the correlation between the prediction errors and the input se-
Cross-correlation Function quence, we can use the sample cross-correlation function:
22-25
Q
1
R pe = ------------
Q p t e t + . (22.19)
t=1
If there is no correlation between the prediction errors and the input se-
quence, then we would expect R pe to be close to zero for all . To deter-
mine if R pe is close to zero, we can set an approximate 95% confidence
interval [BoJe96] using the range
2 Re 0 Rp 0 2 Re 0 Rp 0
- R pe ---------------------------------------
--------------------------------------- -. (22.20)
Q Q
This concept is illustrated in Figure 22.12 and Figure 22.13. Figure 22.12
shows a sample cross-correlation function for the prediction errors of a net-
work that has not been adequately trained. We can see that the cross-cor-
relation function does not fall totally within the bounds defined by Eq.
(22.20), which are indicated by the dashed lines. Figure 22.13 shows the
corresponding cross-correlation function when a network has been success-
fully trained. R pe falls within the bounds for all .
0.2
0.1
0.1
40 20 0 20 40
Figure 22.12 R pe for Inadequately Trained Network
When using a NARX network, correlation between the prediction error and
the input can suggest that the lengths of the tapped delay lines in the input
and feedback paths should be increased.
22-26
0.02
0.01
0.01
0.02
40 20 0 20 40
Figure 22.13 R pe for Successfully Trained Network
Overfitting and Extrapolation

Recall from Chapter 13 that the total data set is divided into three parts:
training, validation and testing. The training set is used to calculate gradi-
ents and to determine weight updates. The validation set is used to stop
training before overfitting occurs. (If Bayesian regularization is used, then
the validation set may be merged with the training set.) The test set is used
to predict future performance of the network. The test set performance is
the measure of network quality. If, after a network has been trained, the
test set performance is not adequate, then there are usually four possible
causes:
the network has reached a local minimum,
the network does not have enough neurons to fit the data,
the network is overfitting, or
the network is extrapolating.
The local minimum problem can almost always be overcome by retraining
the network with five to ten random sets of initial weights. The network
with minimum training error will generally represent a global minimum.
The other three problems can generally be distinguished by analyzing the
training, validation and test set errors. For example, if the validation error
is much larger than the training error, then overfitting has probably oc-
curred. Even though early stopping is used, it is possible to have some over-
fitting, if the training occurs too quickly. In this case, we can use a slower
training algorithm to retrain the network.
22-27
If the validation, training and test errors are all similar in size, but the er-
rors are too large, then it is likely that the network is not powerful enough
to fit the data. In this case, we should increase the number of neurons in
the hidden layer and retrain the network. If Bayesian regularization is
used, this situation is indicated by the effective number of parameters be-
coming equal to the total number of parameters. When the network is large
enough, the effective number of parameters should remain below the total
number of parameters.
If the validation and training errors are similar in size, but the test errors
are significantly larger, then the network may be extrapolating. This indi-
cates that the test data fall outside the range of the training and validation
data. In this case, we need to get more data. You can merge the test data
into the training/validation data and then collect new test data. You should
continue to add data until the results on all three data sets are similar.
If training, validation and test errors are similar, and the errors are small
enough, then we can put the multilayer network to use. However, we still
need to be careful about the possibility of extrapolation. If the multilayer
network inputs are outside the range of the data with which it was trained,
then extrapolation will occur. It is difficult to guarantee that training data
will encompass all future uses of a neural network.
One method for detecting extrapolation is to train a companion competitive
network to cluster the input vectors in the multilayer network training set.
Then, when an input is applied to the multilayer network, the same input
is applied to the companion competitive network. When the distance of the
input vector to the nearest prototype vector of the competitive network is
larger than the distance from the prototype to the most distant member of
its cluster of inputs in the training set, we can suspect extrapolation. This
Novelty Detection technique is referred to as novelty detection.
Sensitivity Analysis
After a multilayer network has been trained, it is often useful to assess the
importance of each element of the input vector. If we can determine that a
given element of the input vector is unimportant, then we can eliminate it.
This can simplify the network, reduce the amount of computation and help
prevent overfitting. There is no one method that can absolutely determine
the importance of each input, but a sensitivity analysis can be helpful in
this regard. A sensitivity analysis computes the derivatives of the network
response with respect to each element of the input vector. If the derivative
with respect to a certain input element is small, then that element can be
eliminated from the input vector.
Because the multilayer network is nonlinear, the derivative of the network
output with respect to an input element will not be constant. For each input
vector in the training set, the derivatives will be different. For this reason,
we cant use a single derivative to determine sensitivity. One option would
22-28
be to take the average of the absolute derivatives, or else the rms deriva-
tives, over the entire training set. Another option would be to compute the
derivative of the sum square error with respect to each element of the input
vector. Each of these will compute a single derivative for each element of
the input vector. The last computation can be performed with a simple vari-
ation of the backpropagation algorithm (see Eq. (11.44) to Eq. (11.47)). Re-
call from Eq. (11.32) that
m F
s i --------m- , (22.21)
n i
where F is a single error squared. We want to convert this to a derivative

with respect to an element of the input vector, using the chain rule:
1 1
S 1 S 1
F F n n

1
------- = -------1- -------i- = si -------i- . (22.22)
p j
i=1
n i p j i=1
p j
We know that
R
w i j pj + b i ,
1 1 1
ni = (22.23)
j=1
therefore, Eq. (22.22) becomes

1 1
S 1 S
F F n i
------- si w i j .
1 1
------- = - -------- = (22.24)
p j n
1 p j
i=1 i i=1
In matrix form, we can write this as
F 1 T 1
------ = W s . (22.25)
p
This will be the derivative for a single squared error. To get the derivative
of the sum square error, we sum the individual derivatives for each single
squared error. The resulting vector will contain the derivatives of the sum
square error for each element of the input vector. If we find that some of
these derivatives are much smaller than the maximum derivative, then we
can consider removing those inputs. After removing the potentially irrele-
vant inputs, we retrain the network and compare the performance with the
original network. If the performance is similar, then we accept the simpli-
fied network.
22-29
Epilogue
While previous chapters have focused on the fundamentals of particular
network architectures and training rules, this chapter has discussed some
practical aspects of neural network training. Neural network training is an
iterative process involving data collection and preprocessing, network ar-
chitecture selection, network training and post-training analysis.
The next five chapters will demonstrate some of these practical aspects, as
we present some real-world case studies. The case studies will cover a va-
riety of applications, including function fitting, density estimation, pattern
recognition, clustering and prediction.
22-30
Further Reading
Further Reading
[Bish95] C.M. Bishop, Neural Networks for Pattern Recognition, Ox-
ford University Press,1995.
This well-written and well-organized textbook presents
neural networks from a statistical perspective.
[BoJe94] G.E.P. Box, G.M. Jenkins, and G.C. Reinsel, Time Series
Analysis: Forecasting and Control, 4th Edition, John Wiley
& Sons, 2008.
This is a classic text on time series analysis. It focuses on
practical aspects, rather than theoretical derivations.
[HaBo07] L. Hamm, B. W. Brorsen and M. T. Hagan, Comparison of
Stochastic Global Optimization Methods to Estimate Neu-
ral Network Weights, Neural Processing Letters, Vol. 26,
No. 3, December 2007.
This paper demonstrates that using multiple restarts of a
local optimization procedure, like steepest descent or con-
jugate gradient, produces results that are comparable to
global optimization methods, but with less computation.
[HeOh97] B. Hedn, H. hlin, R. Rittner, L. Edenbrandt, Acute My-
ocardial Infarction Detected in the 12-Lead ECG by Artifi-
cial Neural Networks, Circulation, vol. 96, pp. 17981802,
1997.
Describes the use of neural networks in detecting myocar-
dial infarctions, using the electrocardiogram.
[Joll02] I.T. Jolliffe, Principal Component Analysis, Springer Series
in Statistics, 2nd ed., Springer, NY, 2002.
The most popular text on principal component analysis.
[LeCu98] Y. LeCun, L. Bottou, G. B. Orr, K.-R. Mueller, Efficient
BackProp, Lecture Notes in Comp. Sci., vol. 1524, 1998.
This paper presents practical tips that improve the train-
ing of multilayer networks.
[Moll93] M. Moller, A scaled conjugate gradient algorithm for fast
supervised learning, Neural Networks, vol. 6, pp. 525-533,
1993.
The scaled conjugate gradient algorithm presented in this
paper converges quickly, and with a minimum amount of
memory requirements.
22
22-31

2126, July 1990.
This paper describes a procedure for setting the initial
weights and biases for the backpropagation algorithm. It
uses the shape of the sigmoid transfer function and the
range of the input variables to determine how large the
weights should be, and then uses the biases to center the
sigmoids in the operating region. The convergence of back-
propagation is improved significantly by this procedure.
[PeCo93] M. P. Perrone and L. N. Cooper, When networks disagree:
Ensemble methods for hybrid neural networks, in Neural
Networks for Speech and Image Processing, R. J. Mam-
mone, Ed., Chapman-Hall, pp. 126-142, 1993.
This paper describes how you can combine the outputs of a
committee of networks to produce results that are more ac-
curate than any of the individual networks.
[PuFe97] G.V. Puskorius and L.A. Feldkamp, Extensions and en-
hancements of decoupled extended Kalman filter training,
Proceedings of the 1997 International Conference on Neural
Networks, vol. 3, pp. 1879-1883, 1997.
The extended Kalman filter algorithm described in this pa-
per is one of the faster sequential algorithms for neural net-
work training.
[RaMa05] L.M. Raff, M. Malshe, M. Hagan, D.I. Doughan, M.G. Rock-
ley, and R. Komanduri, Ab initio potential-energy surfaces
for complex, multi-channel systems using modified novelty
sampling and feedforward neural networks, The Journal
of Chemical Physics, vol. 122, 2005.
This paper describes how neural networks can be used to
model molecular interactions.
[ScSm99] B. Schlkopf, A. Smola, K.-R. Muller, Kernel Principal
Component Analysis, in B. Schlkopf, C. J. C. Burges, A.
J. Smola (Eds.), Advances in Kernel Methods-Support Vec-
tor Learning, MIT Press Cambridge, MA, USA, pp. 327-
352, 1999.
This paper introduces a nonlinear version of principal com-
ponent analysis using a kernel method.
22-32
Objectives
23 Case Study 1:
Objectives 23-1
Description of the Smart Sensor System 23-2
Validation 23-7
Data Sets 23-10
Epilogue 23-11
Objectives
This chapter represents the first of a series of case studies with neural net-
works. Neural networks can be used for a wide variety of applications, and
it would be impossible to provide case studies for each application. We will
limit our presentations to five important application areas: function ap-
proximation (aka, nonlinear regression), density function estimation, pat-
tern recognition (aka, pattern classification), clustering and prediction
(aka, time series analysis, system identification, or dynamic modeling). For
each case study, we will step through the neural network design/training
process.
In this chapter, we present a function approximation problem. For function
approximation problems, the training set consists of a set of dependent
variables (response variables) and one or more independent variables (ex-
planatory variables). The neural network learns to create a mapping be-
tween the explanatory variables and the response variables. In the case
study we consider in this chapter, the system in question is a smart sensor.
A smart sensor consists of one or more standard sensors that are coupled
with a neural network to produce a calibrated measurement of a single pa-
rameter. In this chapter, we will consider a smart position sensor, which
uses the voltages coming from two solar cells to produce an estimate of the
location of an object in one dimension.
23-1
23 Case Study 1: Function Approximation
Theory and Examples

This chapter presents a case study in using neural networks for function
approximation. Function approximation consists of defining a mapping be-
tween a set of input variables and a corresponding set of output variables.
For example, we might want to estimate the price of a home, based on char-
acteristics of the neighborhood, such as tax rate, pupil/teacher ratio in local
schools and crime rate. Another example would be estimating octane num-
ber of a gasoline product at an oil refinery, based on measurements of re-
actor temperatures and pressures [FoGi07]. In the case study presented in
this chapter, we will consider a smart position sensor system.
Description of the Smart Sensor System

Figure 23.1 illustrates the sensor arrangements for this case study. An ob-
ject is suspended between a light source and two solar cells. The object
casts a shadow on the solar cells, which causes the voltage out of the solar
cells to decrease.
y
v1
Shadow
v2
Figure 23.1 Position Sensor Arrangement

As the object position y increases, first the voltage v 1 decreases, then the
voltage v 2 decreases, then v 1 increases, and finally v 2 increases. This is
demonstrated in Figure 23.2. Our objective is to determine the object posi-
tion from measurements of the two voltages. Clearly this is a very nonlin-
ear relationship, so a multilayer network will be needed to learn the
mapping. This is a classic type of function approximation problem, in which
we are trying to learn the inverse of a function. The forward function is the
23-2
Data Collection and Preprocessing
mapping from y to v 1 and v 2 . We want to learn the mapping from v 1 and

v 2 to y .
v2
v1
Figure 23.2 Example Solar Cell Outputs vs. Object Position

In order to collect data for this process, we took measurements of the two
solar cell voltages at a number of calibrated positions of the object. The ob-
ject we used for these experiments was a table tennis ball. The data is dis-
played in Figure 23.3. There are a total of 67 sets of measurements. Each
circle represents a voltage measurement at a calibrated position. The units
of position are inches, and the units of voltage are volts. The flat regions at
0 volts for each curve occur where the shadow of the ball completely covers
a sensor. If the shadow were large enough to cover both cells at the same
time, we would not be able to recover ball position from cell voltages.
5 v2
v1
1
0
0 1 2 3
y
Figure 23.3 Data Collected from Solar Cells
23-3
The next step is to divide the data into training, validation and test sets. In
this case, because we will be using the Bayesian regularization training
technique, we do not need to have a validation set. We did set aside 15% of
the data for testing purposes. To perform the division, we arranged the
data in order, according to object position, and then selected every sixth or
seventh point for testing. This resulted in 10 testing points. The testing
points are not used in any way for training the network, but after the net-
work has been completely trained, we will use the testing data as an indi-
cator of future network performance.
The input vector for network training will consist of the solar cell voltages
v1
p = , (23.1)
v2
and the target will be ball position
t = y. (23.2)
The data were scaled using Eq. (22.1), so that both the inputs and the tar-
gets were in the range [-1,1]. The resulting scaled data is shown in Figure
23.4
S
v2
0.5
S
0.5 v1
1 0.5 0 0.5 1
S
y
Figure 23.4 Scaled Data
Selecting the Architecture

Because the mapping between the solar cell voltages and the ball position
is highly nonlinear, we will use a multilayer network architecture to learn
the mapping. We know that there will be two elements in the input vector,
23-4
which is defined in Eq. (23.1). The single target for the network is the ball
position, given in Eq. (23.2).
Figure 23.5 shows the network architecture. We are using the tan-sigmoid
transfer function in the hidden layer, and a linear output layer. This is the
standard network for function approximation. As we discussed in Chapter
11, this network has been shown to be a universal approximator. There are
cases in which two hidden layers are used, but we normally try first with
1
one hidden layer. The number of neurons in the hidden layer, S , will de-
pend on the function to be approximated. This is something that cannot
generally be known before training. We will have more to say about this in
the next section.
Inputs Tan-Sigmoid Layer Linear Layer
p1 a1 a2
2x1
W1 1
S x1
W2 1x1
1
S x2
n1 1xS
1 n2
S1x1 1x1
1 b 1
1 b2
2 S1x1 S1 1x1 1
1 1 1 2 2 1 2
a = tansig(W p + b ) a = purelin(W a + b )
Figure 23.5 Network Architecture

Before beginning the training, we initialized the network weights using the
method of Widrow and Nguyen described in Chapter 22. Then we used
Bayesian regularization to train the network. Bayesian regularization,
which we discussed in Chapter 13, is a very effective algorithm for training
multilayer networks to perform function approximation. This algorithm is
designed to train networks so that they generalize well, without the need
for a validation set. Because the validation set can then be added to the
training set, the performance is often better than that obtained with early
stopping. (In the next chapter, we will give an example of using early stop-
ping, with a validation set.)
Figure 23.6 illustrates the sum square error versus iteration number,
while using the Bayesian regularization training algorithm. We used a net-
1
work with 10 neurons in the hidden layer ( S = 10 ) for this case. The net-
work was trained for 100 iterations, at which time the performance was
changing very little.
23-5
2
10
1
10
0
10
SSE
1
10
2
10
3
10 0 1 2
10 10 10
Iterations
1
Figure 23.6 Sum Squared Error vs. Iteration Number ( S = 10 )
Training has converged after 100 iterations, but we want to ensure that we
have not fallen into a local minimum. For this reason, we want to retrain
the network several times, using different initial weights and biases. (We
use the Nguyen-Widrow initialization method described in Chapter 22.)
Table 23.1. shows the final validation SSE for each of five different training
runs. We can see that all of the errors are similar, although the errors are
slightly smaller for runs 2, 4 and 5. Any of the weights from these five cases
would produce a satisfactory network. We will discuss this in more detail
in the next section.
1.121e-003 8.313e-004 1.068e-003 8.672e-004 8.271e-004
Table 23.1. Final Training SSE for Five Different Initial Conditions
Recall from Chapter 13 that the Bayesian regularization algorithm com-

putes a parameter , which indicates the effective number of parameters
that are being used by the network. In Figure 23.7, we can see the variation
of during training. It eventually converges to 17.4. There are a total of 41
parameters in this 2-10-1 network, so we are only using about 40% of the
weights and biases. For each of the five training runs discussed above,
converged to values between 17 and 20. This indicates that we might be
able to use a smaller network, if we are concerned about the amount of com-
putation required to compute a network response.
23-6
Validation
20
15
10
0
0 20 40 60 80 100
Iterations
1
Figure 23.7 Effective Number of Parameters ( S = 10 )
To see whether or not a smaller network would be satisfactory, we trained

several networks with different numbers of hidden neurons. Since the ef-
fective number of parameters is near 20, we would expect that a network
with five hidden neurons (21 weights and biases) might provide an ade-
quate fit. Our experiments produced the results shown in Table 23.2. We
can see that the performances of all five networks are roughly equivalent,
1
except for the case where S = 3 , where the total number of parameters in
the network is only 13.
1 1 1 1 1
S = 3 S = 5 S = 8 ( S = 10 ) S = 20
4.406e-003 9.227e-004 8.088e-004 8.672e-004 8.096e-004
Table 23.2. Final Training SSE for Five Different Hidden Layer Sizes
The Bayesian regularization method allows us to train a network of almost

arbitrary size, and yet insure that only the required number of parameters
is effectively used. If we were concerned about the amount of time required
to compute the network output (e.g., for real-time applications), then we
1
would want to use the network with S = 5 . Otherwise, the original net-
1
work with S = 10 is satisfactory. We dont need to spend a lot of time find-
ing the optimal number of neurons. The training algorithm will insure that
we do not overfit.
Validation
An important tool for network validation is a scatter plot of network out-
puts versus targets, as shown in Figure 23.8 (in normalized units). We ex-
23-7
pect that for a well trained network the points in the scatter plot will fall
close to the 45 output=target line. In this case, the fit is excellent. The fig-
ure on the left shows the training data, while the figure on the right shows
the testing data. Because the testing data fit is as good as the training data
fit, we can be confident that the network did not overfit.
1 1
Training Testing
0.5 0.5
a 0 a 0
0.5 0.5
1 1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
t t
Figure 23.8 Scatter Plots of Network Outputs vs. Targets - Training and Testing Sets
Another useful plot is a histogram of network error, as shown in Figure
23.9. This gives us an idea of the accuracy of the network. For this histo-
gram, we have converted the network output back into units of inches. This
is done by applying the reverse of the target preprocessing function to the
network output. The reverse of Eq. (22.1), for the targets, is given by
max min
n t t min
a = a + 1 .* ----------------------------- + t (23.3)
2
n
where a is the original network output, which was trained to match the
normalized target, and .* represents an element-by-element multiplica-
tion of two vectors. After the postprocessing operation of Eq. (23.3), the re-
sulting un-normalized output is subtracted from the raw targets, to
produce an error in inches. Figure 23.9 shows the distribution of these er-
rors for both training and testing sets. We can see that almost all errors are
within one hundredth of an inch. This is within the accuracy of the original
measurements, so we cannot expect to do better.
23-8
Validation
15
10
0
0.02 0.01 0 0.01 0.02
Figure 23.9 Histogram of Position Errors (in Inches)

Because this network has only two inputs, we can plot the trained network
response, which is shown in Figure 23.10. (This figure shows the response
from original unscaled inputs, in volts, to original unscaled output, in inch-
es.) The blue circles indicate the path that is taken by the voltages as the
ball is moved. Notice that the Bayesian regularization training has pro-
duced a smooth network response, even though the response is highly non-
linear.
2.5
2
y 1.5
0.5
0
0
2
6
v2 4
4
2 v1
0
Figure 23.10 Network Response (Original Units)

Another thing to notice about Figure 23.10 is that training data only falls
along the blue circles. The form of the network response in other regions is
of no importance for the operation of the smart sensor system, since the
23-9
network will never be used there. If the network were retrained, the shape
of the response away from the blue circles might be very different, even
though the response near the blue circles will always be the same. This con-
cept is very important for many neural network applications. Often, during
normal network operation, only a small portion of the input space will be
accessed. The network only has to fit the underlying function in these re-
gions where the network will be used. This means that the size of the data
set can be modest, even when the input dimension is large. Of course, in
these cases, it is critical that the training data span the full range of poten-
tial network operation.
Data Sets
There are two data files associated with this case study:
ball_p.txt contains the input vectors in the original data set
ball_t.txt contains the target vectors in the original data set
They can be found with the demonstration software, which is described in
Appendix C.
23-10
Epilogue
23
Epilogue
This chapter has demonstrated the use of multilayer neural networks for
function approximation. This case study is representative of a large class
of neural network applications that could be termed soft sensors or
smart sensors. The idea is to use a neural network to fuse several raw
sensor outputs into a calibrated measurement of some key variable of in-
terest.
A multilayer network with sigmoid transfer functions in the hidden layers
and linear transfer functions in the output layer is well suited to this type
of application, and Bayesian regularization is an excellent training algo-
rithm to use in this situation.
In the next chapter, we will look at another neural network application
probability estimation. We will also use multilayer neural networks for
that application, but we will change the transfer function in the output lay-
er.
23-11
Further Reading
[FoGi07] L. Fortuna, P. Giannone, S. Graziani, M. G. Xibilia, Virtu-
al Instruments Based on Stacked Neural Networks to Im-
prove Product Quality Monitoring in a Refinery, IEEE
Transactions on Instrumentation and Measurement, vol.
56, no. 1, pp. 95101, 2007.
This paper describes the use of neural networks as soft sen-
sors in a refinery. Measurements of reactor temperatures
and pressures are used to predict octane number in a gaso-
line product.
23-12
Objectives
24 Case Study 2:
Probability Estimation 24
Objectives 24-1
Description of the CVD Process 24-2
Validation 24-9
Data Sets 24-12
Epilogue 24-13
Objectives
This chapter represents the second of a series of case studies with neural
networks. The previous chapter demonstrated the use of neural networks
for function approximation. In this chapter we use a neural network to es-
timate a probability function.
Probability estimation is a special case of function approximation. In func-
tion approximation we want the neural network to map between a set of in-
put variables and a set of response variables. However, in the case of
probability estimation the response variables correspond to a set of proba-
bilities. Since probabilities have certain special properties they must al-
ways be positive, and they must sum to 1 we want the neural network
to enforce these conditions.
In the case study we consider in this chapter, the system in question is
chemical vapor deposition of diamond. A carbon dimer (a bound pair of car-
bon atoms) is projected toward a diamond surface. We want to determine
the probabilities for various reactions based on characteristics of the pro-
jected dimer. The input variables consist of such properties as translational
energy and incidence angle, and the response variables consist of the prob-
abilities of the potential reactions, such as chemisorption and scattering.
24-1
24 Case Study 2: Probability Estimation
Theory and Examples

This chapter presents a case study in using neural networks for probability
estimation. Probability estimation consists of determining the probabilities
of certain events, based on a set of input variables. For example, we might
want to know the probabilities associated with a patient having a certain
disease, based on a set of laboratory tests. Another example would be de-
termining the probability of a financial instrument going up in price, based
on a set of market conditions.
For this probability estimation case study, we will train a neural network
to estimate reaction rates in a chemical process. Chemical vapor deposition
CVD (CVD) of diamond is a process for making synthetic diamond. The idea is to
cause carbon atoms in a gas to settle on a substrate in crystalline form. In
order to study this process, scientists are often interested in reaction rates,
which will determine how quickly the diamond can be created. In this case
study, we will train a neural network to compute reaction rates as a carbon
dimer (a bound pair of carbon atoms) interacts with the crystalline dia-
mond substrate.
We will begin by describing the CVD process and how simulated data can
be collected for this process. Then, we will show how a neural network can
be trained to learn the reaction probabilities. The details of the procedure
are described in [AgSa05].
Description of the CVD Process

During the CVD process, a carbon dimer is projected toward a diamond
substrate. For the purpose of this study, we will assume that the dimer can
react with the substrate in one of three ways: chemisorption (the atoms in
the dimer become bound to the substrate), scattering (the atoms bounce off
the substrate), or desorption (the atoms become bound to the substrate for
a period of time, but are then released). There is another possible reaction
that occurs with very small probability, but we will ignore it for this study.
(See [AgSa05] for a full discussion.) We will train a neural network to esti-
mate the probabilities of each of the reactions, based on various character-
istics of the carbon dimer, which will be described below.
The notation we will use to define this interaction is illustrated in Figure
24.1. The black circle represents the carbon dimer, and the corresponding
directed line represents the direction of the initial velocity vector. The blue
star represents the location of the central carbon atom in the diamond sub-
strate. The angle denotes the angle of incidence, i.e., the angle between
the direction of the initial velocity vector of the carbon dimer and the per-
pendicular on the surface (the z direction). The impact parameter b is de-
fined as the distance between the location of the central atom and the point
of intersection of the initial velocity vector and the diamond surface (indi-
24-2
cated by the origin of the axes in Figure 24.1). The angle represents the
angle between the x axis and the line from the origin to the central atom.
z
y
b
f
Figure 24.1 Notation for Carbon Dimer/Diamond Substrate Interaction

Data for training the neural network are obtained by molecular dynamics
(MD) simulations. In MD, the motion of atoms and molecules in a material
under a given force are simulated, using known laws of physics to calculate
the forces on individual atoms [RaMa05]. For this case study, we use a total
of 324 atoms to model the CVD system. Out of these, 282 atoms of diamond
substrate are used to model the crystalline face with 40 atoms of hydrogen
on the top layer of the diamond surface, and 2 atoms in the C2 dimer. In
Figure 24.1, the (x,y) plane represents the location of the diamond sub-
strate. Each of the carbon atoms on the top layer of the substrate, except
the central atom and boundary atoms, is capped by a hydrogen atom. Any
reactions will occur near the central atom.
For this study, we want to determine the dependence of probabilities for
chemisorption, scattering, and desorption on b, , , rotational velocity
( v rot ), and translational velocity ( v trans ) of the C2 dimer. The initial C2 vibra-
tional energy is set equal to the zero-point energy and the temperature of
the lattice is maintained constant at 600 K [RaMa05].
This is an interesting function approximation problem, in that we dont
have access to the true underlying reaction probabilities, which are un-
24-3
known. We will obtain estimates of these probabilities by running Monte

Carlo simulation experiments. We will need some notation to help us keep
track of the various probabilities that we will work with. First, we will in-
dicate the true underlying reaction probabilities by P X p , where X refers
to the reaction process, and p is the vector that characterizes the C2 dimer:

p = b . (24.1)
v trans
v rot
The reaction process can be chemisorption ( X = C ), scattering ( X = S ), or

desorption ( X = D ). The probability estimates produced by the neural net-
NN
work will be indicated by P X p . The probability estimates obtained from
MC
the Monte Carlo simulations will be indicated by P X p .
The Monte Carlo estimates are obtained by
MC N
P X p = -----X- , (24.2)
NT
where N X is the number of MD trajectories that resulted in reaction X and

N T is the total number of trajectories computed. The results of a given tra-
jectory depend upon a multitude of input variables. These include the pa-
rameters included in p , as well as the initial orientation of the C2 dimer,
the angle defining the C2 rotational plane, the initial C2 vibrational energy
and its phase, the temperature of the system, and all of the variables that
define the vibrational phases of the diamond surface. Because we are only
interested in the effect of p on the reaction probabilities, the other vari-
ables are randomly set for each MD simulation, except that the initial C2
vibrational energy is set equal to the zero-point energy and the tempera-
ture of the lattice is maintained constant at 600 K. Eq. (24.2) averages over
the trajectories to estimate the underlying true probabilities P X p . (As a
note of clarification here, we use the term Monte Carlo to refer to the set of
simulations that are obtained by setting a number of the variables to ran-
dom values for each trajectory. We refer to the simulation of a single tra-
jectory as an MD simulation, since the principles of molecular dynamics
are used to perform the computations.)
This is a standard method used by chemists to estimate reaction probabil-
ities. If they want to determine the effect of , for example, on the proba-
bilities, they must run a series of Monte Carlo simulations at each value of
that is of interest. This can be extremely time consuming. The required
number of Monte Carlo trials can be quite large, if an accurate reaction
probability is required. Our objective in this case study is to train a neural
24-4
network to learn the true reaction probabilities as a function of the param-

eters in p .
To train a neural network, we need a set of target outputs. Since we do not
know the true underlying probabilities P X p , we will use the estimates ob-
MC
tained from the Monte Carlo simulations P X p . We can think of these
Monte Carlo probabilities as being noisy versions of the true probabilities.
The neural network will need to interpolate these noisy values to produce
an accurate estimate of P X p without overfitting. This is a good applica-
tion for our generalization procedures, which were discussed in Chapter 13.
MC
The data set consists of 2000 different p P X p input/target pairs. Out
of these 2000 data points, 1400 (70%) were randomly selected for training,
300 (15%) for validation, and 300 for testing. For each trajectory, the p
were generated randomly, using physically-appropriate distributions for
each variable [RaMa05]. A total of N T = 50 different trajectories were run
MC
to obtain each P X p . This means that 2000x50 trajectories were run to
create the entire data set.
The original units of the inputs are radians for and , angstroms for b,
angstroms per picosecond for v trans and radians per picosecond for v rot . Be-
fore presenting the input data to the network for training, they are scaled
using Eq. (22.1), so that each element of the input vector ranges from -1 to
1. The targets have values that are always in the range 0 to 1, since they
represent probabilities. In the next section, we will describe a network ar-
chitecture, in which the softmax transfer function of Eq. (22.3) is used in
the final layer. This transfer function produces outputs that range from 0
to 1, so the original unscaled targets will work fine.

We will use a multilayer network for this application. We know that there
will be five elements in the input vector, which is defined in Eq. (24.1). The
target for the network can be a vector with three elements:
MC
PC p
t = P MC p , (24.3)
S
MC
PD p
MC
or we can use three different networks, each with a different P X p as a
target. We have tried both possibilities, and the results are similar.
In this case, there is an advantage to using the single network, with three
elements in the output vector. The three targets represent probabilities.
Therefore, they are always in the range 0 to 1, and they always sum to 1.
24-5
This is an ideal situation for using the softmax transfer function of Eq.
(22.3), which is repeated here:
S
a i = f n i = exp n i exp nj . (24.4)

j=1
This transfer function is different from others we have used, in that each
neuron output a i is affected by all of the net inputs n j . (In the other trans-
fer functions, the net input n i affected only the neuron output a i .) This
does not cause any substantial difficulties in network training. The back-
propagation algorithm of Eq. (11.44) and Eq. (11.45) can still be used to
compute the gradient. However, the derivative of the transfer function is
no longer a diagonal matrix. The derivative of the softmax function has the
following form:
m
S
a 1m a im a 1m a 1m a 2m a 1m a Smm

i = 1
m
S
F m n m = a 2m a 1m a 2 a im a 2m
m a 2m a Smm
(24.5)
i = 1

m
S
a Smm a 1m a Smm a 2m a Smm a im a Smm

i = 1
The complete network architecture is shown in Figure 24.2. The input vec-
tor of Eq. (24.1) has 5 elements. The output vector has 3 elements, which is
consistent with the target vector of Eq. (24.3). The transfer function in the
hidden layer is the hyperbolic tangent sigmoid, and the softmax transfer
function is used in the output layer. The number of neurons in the hidden
1
layer, S , is yet to be determined. It depends on the complexity of the func-
tion that we are trying to approximate, but we do not know at this point
how complex the function is. In general, the size of the hidden layer must
1
be determined as part of the training process. We must choose S so that
the network provides an accurate fit to the training data, without overfit-
ting. We will discuss this selection in the next section.
24-6
Inputs Tan-Sigmoid Layer Softmax Layer
1 1 2
p 1
a 2
a
5x1
W S1x1
W 3x1
1
S x5
n1 3xS1
n2
S1x1 3x1
1 b 1
1 b2
5 S1x1 S1 3x1 3
1 1 1 2 2 1 2
a = tansig(W p + b ) a = softmax(W a + b )

We trained the network using the scaled conjugate gradient algorithm of
[Mill93]. Many other conjugate gradient or Levenberg-Marquardt algo-
rithms, such as those discussed in Chapter 12, would have also worked
well. The targets for this problem have a significant amount of noise, so we
are not expecting extreme accuracy in the final fit. We used early stopping,
as described in Chapter 13, to prevent overfitting. We stopped the training
if the error on the validation set failed to improve over 25 iterations. A typ-
ical training session is illustrated in Figure 24.3, which shows training and
validation MSE. The minimum of the validation performance was reached
at iteration 69. The algorithm continued for 25 more iterations, until iter-
ation 94. Since the validation error was not reduced during those 25 itera-
tions, the weights from iteration 69 were saved as the final trained values.
1
10
Training
Validation
MSE
2
10
3
10 0 1 2
10 10 10
Iterations
1
Figure 24.3 Training and Validation Mean Square Error ( S = 10 )
24-7
The results shown in Figure 24.3 represent a network with 10 neurons in

1
the hidden layer ( S = 10 ). We need to verify that this is a reasonable num-
ber. One indicator is a comparison of training and validation performance.
Table 24.1. shows the training and validation root mean square error
(RMSE) for the trained network. We can see that training and validation
errors are roughly the same. The validation data was randomly selected
and was selected independently of the training set. Because the errors were
approximately the same on both sets, it appears that the network fit is con-
sistent throughout the relevant input range, and no overfitting occurs.
Training RMSE Validation RMSE
PC p 0.0496 0.0439
PS p 0.0634 0.0659
PD p 0.0586 0.0604
1
Table 24.1. Comparison of Training and Validation RMSE for S = 10
It is also important to determine if the errors are as small as possible and

if the fit is adequate. We will have more to say about that in the next sec-
tion, but at this point we can try fitting networks with different numbers
of hidden neurons. Table 24.2. shows the results of fitting a network with
two hidden neurons. Again, the training and validation errors are consis-
tent, which indicates lack of overfitting, but the errors are higher than
1
those for S = 10 .
PC p 0.0634 0.0627
PS p 0.0669 0.0704
PD p 0.0617 0.0618
1
1
Table 24.3. shows the results for S = 20 . The validation error is slightly
higher than the training error, which might indicate some overfitting. The
main point is that neither training nor validation errors are significantly
1 1
smaller for S = 20 than for S = 10 . This indicates that ten hidden neu-
rons are sufficient for this problem. We will investigate this further in the
next section.
24-8
Validation
PC p 0.0432 0.0444
PS p 0.0603 0.0643
PD p 0.0569 0.0595
1
There is one further step that we want to make as part of the training pro-
cess. We want to ensure that we have not fallen into a local minimum. For
this reason, we want to retrain the network several times, using different
initial weights and biases. (We use the Nguyen-Widrow initialization
method described in Chapter 22.) Table 24.4. shows the final validation
MSE for each of five different training runs. We can see that all of the er-
rors are similar, so we have reached a global minimum at each run. If one
error was significantly lower than the others, then we would use the
weights that obtained the lowest error.
3.074e-003 2.953e-003 3.031e-003 3.105e-003 3.050e-003
Table 24.4. Final Validation MSE for Five Different Initial Conditions
We have determined that a neural network with ten neurons in the hidden
layer produces a reasonable response without overfitting. The next step is
to analyze the performance of the network. Depending on the results of
that analysis, we might adjust the network architecture or training data
and retrain the network.
Validation
An important tool for network validation is a scatter plot of network out-
puts versus targets, as shown in Figure 24.4. Here we can see that there is
a strong linear relationship between the targets and the network outputs,
but there appears to be quite a bit of variation. We might expect that for a
well trained network the points in the scatter plot would fall exactly on the
outputs=target line. Why do we have so much variation in this plot? The
reason is that the targets of the network are not the true reaction probabil-
MC
ities, P X p , but the Monte Carlo estimates P X p . There is noise in the
targets.
24-9
1 1 1
0.8 0.8 0.8
0.6 0.6 0.6

NN NN NN
PC PS PD
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
MC MC MC
PC PS PD
Figure 24.4 Scatter Plot of Network Outputs vs. Targets ( N T = 50 )
MC
The relationship between P X and P X is such that ~95% of the time we ex-
pect to have
MC
P X 2 P X P X + 2 , (24.6)
where
PX 1 PX
= ---------------------------- . (24.7)
NT
This relationship is illustrated in Figure 24.5 for N T = 50 .
0.8
0.6
MC
PX
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
PX
MC
Figure 24.5 Expected Statistical Spread of P X for N T = 50
24-10
Validation
By comparing Figure 24.4 with Figure 24.5, we can see that the spread in
MC
the data is explained by the statistical variations in P X . To further verify
this observation, we generated additional testing data, in which 500 Monte
MC
Carlo trials were run to obtain each P X p (i.e., N T = 500 ). We applied
this testing data to the network that was trained on the original data set
with N T = 50 . The resulting scatter plots are shown in Figure 24.6. Here
we can see that the spread has decreased dramatically from Figure 24.4,
even though the network has not changed. This means that the neural net-
work is fitting the true probabilities P X p and not the statistical fluctua-
MC
tions in P X p .
1 1 1
0.8 0.8 0.8
0.6 0.6 0.6

NN NN NN
PC PS PD
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
MC MC MC
PC PS PD
Figure 24.6 Scatter Plot of Network Outputs vs. Targets ( N T = 500 )
After the neural network has been trained, it becomes a simple matter to
investigate the effect of the input parameters on the reaction probabilities.
In Figure 24.7, we see the effect of the impact parameter b on the reaction
probabilities, as determined by the neural network. As the impact param-
eter is increased, the probability of chemisorption decreases, while the
probabilities of scattering and desorption increase. (For this study, we have
set to 5.4 radians, to 0.3 radians, v rot to 0.004 radians per femtosecond,
and v trans to 0.004 angstroms per femtosecond.)
With standard methods, a study such as that shown in Figure 24.7 would
take thousands of simulations. The trained neural network has fully cap-
tured the relationships between the parameters in p and the reaction prob-
abilities. Therefore, we can perform arbitrary studies by simply computing
the network responses at a varying set of input points. Note that the net-
work interpolated smoothly through a noisy set of data points to capture
the true underlying function. By using the early stopping technique, we
prevented the network from overfitting the noise in the data.
24-11
NN
PC 0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
NN
PS 0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
NN
PD 0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
Figure 24.7 Reaction Probabilities vs. Impact Parameter
Data Sets
There are four data files associated with this case study:
cvd_p.txt contains the input vectors in the original data set
cvd_t.txt contains the target vectors in the original data set
cvd_p500.txt contains the input vectors in the N T = 500 test set
cvd_t500.txt contains the target vectors in the N T = 500 test set
Appendix C.
24-12
Epilogue
Epilogue
This chapter has illustrated the use of neural networks for probability es-
timation on a chemical vapor deposition problem. Monte Carlo simulations
were used to provide estimates of the reaction probabilities. These esti-
24
mates were used as targets for the neural network. The network was able
to capture the true underlying probability function without overfitting to
the errors in the Monte Carlo estimates. This was accomplished by using
the early stopping procedure, which stops network training if the error on
an independent validation set increases.
In the next chapter, we apply neural networks to a pattern recognition
problem. We will also use multilayer neural networks for that application.
24-13
Further Reading
[AgSa05] P.M. Agrawal, A.N.A. Samadh, L.M. Raff, M. Hagan, S. T.
Bukkapatnam, and R. Komanduri, Prediction of molecu-
lar-dynamics simulation results using feedforward neural
networks: Reaction of a C2 dimer with an activated dia-
mond (100) surface, The Journal of Chemical Physics 123,
224711, 2005.
This paper describes the details of training a neural net-
work to predict the reaction probabilities for chemical va-
por deposition of diamond.
[Mill93] M.F. Miller, A scaled conjugate gradient algorithm for fast
1993.
The scaled conjugate gradient algorithm is a fast batch
training algorithm for neural networks that requires a
minimum of memory and computation at each iteration.
of Chemical Physics, 122, 084104, 2005.
This paper explains how neural networks can be used for
molecular dynamics simulations.
24-14
Objectives
25 Case Study 3:
Pattern Recognition
Objectives 25-1
Description of Myocardial Infarction Recognition 25-2
Validation 25-7
Data Sets 25-10
Epilogue 25-11
Objectives
This chapter presents a case study in using neural networks for pattern
recognition. In pattern recognition problems, you want a neural network to
classify inputs into a set of target categories, e.g., recognize the vineyard
that a particular bottle of wine came from, based on a chemical analysis, or
classify a tumor as benign or malignant, based on uniformity of cell size,
clump thickness, mitosis.
In this chapter we will demonstrate the application of multilayer neural
networks to the recognition of heart disease from a reading of the electro-
cardiogram. We will show each of the steps in the pattern recognition pro-
cess: data collection, feature extraction, architecture selection, network
training and network validation.
25-1
25 Case Study 3: Pattern Recognition
Theory and Examples

In pattern recognition (pattern classification) problems we are trying to
categorize network inputs into their corresponding classes. Here are a few
examples of pattern recognition problems:
recognition of handwritten zip codes
spoken word recognition
disease recognition from a list of symptoms
fingerprint recognition
white blood cell classification
In the case study presented in this chapter, we will be looking for patterns
in electrocardiogram signals that indicate the presence of a myocardial in-
farction (heart attack).
Description of Myocardial Infarction Recognition

An electrocardiogram (EKG) is a recording of the electrical activity of the
heart over time. It generally consists of an array of different signals record-
ed at the same time. An EKG can consist of a single signal (also called a
lead), although the standard EKG that is used for detailed interpretation
consists of 12 leads. EKGs with as many as 15 leads are sometimes used.
Each lead represents the electrical activity across two points on the body.
The 12-lead EKG is determined from 10 electrodes that are placed on spe-
cific locations on the body. The calculation of the 12-lead potentials from
the 10 electrodes is a somewhat complex calculation, and beyond the scope
of this case study. The interested reader is referred to [Dubi00] for a more
complete discussion of the EKG.
Through a careful analysis of the EKG, a physician can often determine the
health of the heart. The shapes of the signals indicate the path of electrical
flow in the heart as various muscles are contracted in a coordinated way to
pump blood in and out. If a part of the heart muscle has been damaged be-
cause of a lack of blood flow through the coronary arteries (called a myocar-
dial infarction (MI), or heart attack), then the path of electrical flow
changes. A well-trained physician can discern from the changes in the
EKG, if the heart has been damaged, and where the damage has occurred.
For this case study, we will train a neural network to recognize MIs, using
information obtained from a 15 lead EKG.
25-2

The EKG signals used for this case study were obtained from the Physio-
Net database [MoMa01]. Data were extracted from the QT data set for
healthy patients and patients with MIs. Each EKG consists of 15 leads.
The leads are labeled I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6, VX,
VY, VZ. Figure 25.1 shows a small portion of the lead I signal for one of the
healthy patients.
0.6
0.4
0.2
0.2
0 0.5 1 1.5 2
Time (sec)
Figure 25.1 Example EKG Signal

Our data set has a total of 447 EKG records. Of these, 79 represent healthy
patients, and the remaining 368 have an MI diagnosis. A diagnosis for each
record was provided by a physician, but it is possible for the diagnosis to be
in error. We will have more to say about this when we come to the valida-
tion of the network.
Each EKG consists of 15 leads measured at a rate of 1000 Hz for a period
of several minutes. This is an enormous amount of data, and it would be
impractical to use the entire EKG as an input to the neural network. As
with many pattern recognition problems, we need to perform a feature ex-
traction step before using the neural network to execute the pattern recog-
nition step. Feature extraction involves mapping the high-dimensional
input space into a space with fewer dimensions, in order to simplify and
make more robust the pattern recognition step.
There are a number of general methods for dimensionality reduction. This
includes linear methods, like the principal components method that we
mentioned in Chapter 22, and nonlinear methods, like manifold learning
[TeSi00]. For this case study, instead of using general methods to generate
the low-dimensional feature space, we will extract features that are com-
25-3
monly used by physicians to detect abnormalities in the EKG. The first step
is to consider a typical cycle of an EKG signal, as shown in Figure 25.2.
R
PR ST
Segment Segment
P T
Q S
PR
Interval QRS
Duration
QT Interval
Figure 25.2 Prototype Cycle of an Electrocardiogram Signal

William Einthoven, in the early 1900s, was the first to carefully measure
and analyze the EKG. He assigned the letters P, Q, R, S and T to the vari-
ous deflections shown in the prototype cycle of Figure 25.2. He also de-
scribed the electrocardiographic features of a number of cardiovascular
disorders. He won the Nobel Prize in Medicine in 1924 for his discoveries.
His features are still used to this day.
For this study, we have used some of the features that are in standard use
by physicians, as well as other features that are related to the prototype cy-
cle [Raff06]. The descriptions of the 47 features that we used are listed be-
low. (If a description refers to the amplitude lead, this refers to the square
root of the sum of squares of the VX, VY and VZ leads.)
Input Features to the Neural Network
1. age in years
2. gender, -1=female, 1=male
3. maximum heart rate in beats/min
4. minimum heart rate in beats/min
5. average time between heart beats in sec
6. rms deviation of the mean heart rate in beats/sec
7. full width at half maximum for the heart rate distribution
8. average qt interval for lead with max t wave
9. average qt interval for all leads
10. average corrected qt interval for lead with max t wave
11. average corrected qt interval for all leads
12. average qrs interval for all leads
13. average pr interval for lead with maximum p wave
14. rms deviation of pr intervals from average-max p lead
25-4
15. average pr interval for all leads

16. rms deviation for pr interval from average-all leads
17. percentage of negative p waves-max p lead
18. average percentage of negative p waves for all leads
19. maximum amplitude of any t wave
20. rms deviation of qt intervals
21. rms deviation of corrected qt intervals
22. average st segment length
23. rms deviation of st segment lengths 25
24. average heart rate in beats/min
25. rms deviation of heart rate distribution in beats/min
26. average rt angle averaged over all amplitude beats
27. number of missed r waves (beats)
28. % total qt intervals not analyzed or missing
29. % total pr intervals not analyzed or missing
30. % total st intervals not analyzed or missing
31. average number of maxima between t wave end and q
32. rms deviation of rt angle for all beats
33. ave qrs from amplitude lead
34. rms deviation of qrs from amplitude lead
35. ave st segment from amplitude lead
36. rms deviation of st segment from amplitude lead
37. ave qt interval from amplitude lead
38. rms deviation of qt interval from amplitude lead
39. ave bazetts corrected qt interval from amplitude lead
40. rms deviation of corrected qt interval from amplitude lead
41. ave r-r interval from amplitude lead
42. rms deviation of r-r interval from amplitude lead
43. average area under qrs complexes
44. average area under s-t wave end
45. average ratio of qrs area to s-t wave area
46. rms deviation of rt angle within each beat averaged over all beats in amplitude signal
47. st elevation at the start of the st interval for amplitude signal
To summarize, the data set contains 447 records. Each record has 47 input
variables, and one target value. The target is 1 for a healthy diagnosis and
-1 for an MI diagnosis.
One of the problems with the data set is that there are only 79 records for
the healthy diagnosis, while there are 368 records for the MI diagnosis. If
we train the network using the sum square error performance index, where
all of the errors are weighted equally, the network will be biased to indicate
25-5
the MI diagnosis. The ideal solution to this problem would be to collect

more data from healthy patients. Lets say that this is not possible in this
case, and we need to do what we can with the data available. One possibil-
ity is to use a weighted sum square error as the performance index, where
errors for healthy patients would be weighted higher than errors for MI pa-
tients, so that overall healthy and MI contributions would be equal if each
error were equal. Another simple approach is to repeat the healthy records
in the data set, so that the total number of healthy records is equal to the
number of MI records. This requires extra computation, but that is not a
problem in this case study. Since it is the simplest solution, we will use it
here.
After the data has been collected, the next step is to divide the data into
training, validation and test sets. In this case, we randomly set aside 15%
of the data for validation and 15% for testing. For the validation and test-
ing sets, we did not include multiple entries of the healthy records. This
was done only for the training set.
The data were normalized using Eq. (22.1), so that the inputs were in the
range [-1,1]. Since the tangent-sigmoid transfer function will be used in the
output layer of the neural network, the targets were set to values of -0.76
and +0.76, instead of -1 and 1, to prevent training difficulties caused by sat-
uration of the sigmoid functions, as discussed in Chapter 22.

Figure 25.3 shows the network architecture. We are using the tangent-sig-
moid transfer function in both layers. This is the standard network for pat-
tern recognition. There are cases in which two hidden layers are used, but
we normally try first with one hidden layer. The number of neurons in the
1
hidden layer, S , will depend on the complexity of the decision boundaries
needed for the pattern recognition task. This is something that cannot gen-
erally be known before training. We will start with 10 neurons in the hid-
den layer, and then test the network performance after training.
Inputs Tan-Sigmoid Layer Tan-Sigmoid Layer
1 2
p 1
a 2
a
47 x 1
W 1
S x1
W 1x1
1
S x 47
n1 1xS1
n2
S1x1 1x1
1 b 1
1 b2
47 S1x1 S1 1x1 1
1 1 1 2 2 1 2
a = tansig(W p + b ) a = tansig(W a + b )
25-6

We trained the network using the scaled conjugate gradient algorithm of
[Mill93]. This algorithm is very efficient for pattern recognition problems.
We used early stopping to prevent network overfitting.
Figure 25.4 illustrates the mean squared error versus iteration number.
The blue line shows the validation error, and the black line shows the train-
ing error. We used a network with 10 neurons in the hidden layer
1
( S = 10 ). The minimum validation error occurred at iteration 16, as indi-
cated by the circle in Figure 25.4, and the network parameters were saved
at this point. Note that the validation error curve does not always fall at
each iteration, and it may rise before falling to a lower value. We tested
that the validation error was not reduced over 40 iterations before we final-
ly stopped the training.
1
10
0
10
Validation
MSE
1
10
Training
2
10
0 10 20 30 40 50
Iterations
1
Figure 25.4 Mean Squared Error vs. Iteration Number ( S = 10 )
Validation
As we discussed in previous chapters, an important tool for network vali-
dation in function approximation problems is a scatter plot of network out-
puts versus targets. For pattern recognition problems, the network outputs
and targets are discrete variables, so a scatter plot is not particularly use-
ful. Instead of the scatter plot, we use the confusion matrix, which was dis-
cussed in Chapter 22. Figure 25.5 shows the confusion matrix for our
trained network on the test data. The upper left cell shows that 13 of the
14 healthy EKGs in the test set were classified correctly, while the 2,2 cell
shows that 66 of the 71 MI EKGs were classified correctly. A total of 92.9%
of the test data were classified correctly. The largest number of mistakes
were for MI records that were classified as healthy (5), as shown in cell 1,2.
25-7
Confusion Matrix
13 5 72.2%
1
15.3% 5.9% 27.8%
Output Class
1 66 98.5%
2
1.2% 77.6% 1.5%
92.9% 93.0% 92.9%

7.1% 7.0% 7.1%
1 2
Target Class
Figure 25.5 Confusion Matrix for Test Data (One Data Division)
Another useful validation tool for pattern recognition problems is the Re-
ceive Operating Characteristic (ROC) curve, described in Chapter 22. Fig-
ure 25.6 shows the ROC curve (blue line) for the test data. The ideal curve
would follow the path from 0,0 to 0,1 and then to 1,1. The curve for this test
set is close to the ideal path.
0.9
0.8
0.7
True Positive Rate
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
False Positive Rate
Figure 25.6 Receiver Operating Characteristic Curve (Test Set)

The results shown in Figure 25.5 and Figure 25.6 represent one division of
the data into training/validation/testing sets. Because the data set is fairly
small, especially in terms of healthy diagnoses, we might wonder how sen-
sitive the results are to the data division. To investigate this sensitivity, we
performed a Monte Carlo simulation. The data were divided 1,000 different
25-8
Validation
times. For each division of the data, a neural network was trained with dif-
ferent random initial weights. The results of the 1,000 trials were averaged
together, and the results are shown in Figure 25.7.
Confusion Matrix
9.11 3.69 71.2%

1
14.0% 5.7% 28.8%
Output Class
2.51 49.83 95.2%
2
3.9% 76.5% 4.8%
78.4% 93.1% 90.5%

21.6% 6.9% 9.5%
1 2
Target Class
Figure 25.7 Average Test Confusion Matrix for 1,000 Monte Carlo Runs
Figure 25.7 represents the average results over the 1,000 different net-
works and data divisions. There were approximately 12 healthy patients
(on average) in each test set. Of these, more than 9 were correctly diag-
nosed. There were approximately 54 sick patients in each test set, and ap-
proximately 50 were correctly diagnosed. The average testing error was
approximately 9.5%. Note that none of the patients in the test set were
used to train the neural network, so these numbers should be conservative
estimates of how the network should perform on new patients.
The average test results for the Monte Carlo simulation are similar to our
original test results. However, in addition to knowing the average results,
it is also helpful to look at the distribution of errors. Figure 25.8 shows a
histogram of the percentage errors. The average percent error is 9.5%, but
there is a significant spread in the distribution of errors. The standard de-
viation of the percent error is 3.5.
25-9
160
120
80
40
0
0 5 10 15 20 25
Percent Error
Figure 25.8 Percent Error Histogram (1000 Monte Carlo Trials)

This Monte Carlo process can be helpful in validating the data and the
training process. For example, it is possible to identify which patients are
misclassified in each Monte Carlo run. Those patients who are consistently
misclassified (regardless of the division of the data) can be carefully inves-
tigated. These cases can be helpful in two areas. First, they can enable us
to refine the data base. If, after reevaluation by clinicians, it is determined
that a particular patient was mislabeled in the original data set, the data
can be corrected. Second, if we find upon review that the patient was cor-
rectly labeled in the original data set, then we can use that patient to help
improve the operation of the neural network classification. This may in-
volve identifying new features which capture the key characteristics of the
EKG, or it may involve obtaining more data with similar characteristics, in
order to reinforce the training of the neural network.
The Monte Carlo process can be also used to help to improve the network
performance. By combining the individual networks obtained from the
Monte Carlo trials, we can often obtain a more accurate classification. The
same input is applied to all of the networks, and the outputs can be com-
bined through a voting procedure. We choose the class that is selected by
the largest number of networks.
Data Sets
ekg_p.txt contains the input vectors in the ekg data set
ekg_t.txt contains the targets in the ekg data set
Appendix C.
25-10
Epilogue
Epilogue
pattern recognition. In this case study, the pattern recognition network
was used to classify EKG records into healthy and myocardial infarction di-
agnoses.
Most pattern recognition tasks involve a feature extraction step, in which
the original data set is reduced in dimension. The features that were ex-
tracted from the EKG data consisted of characteristics of the prototype
25
EKG cycle.
A Monte Carlo procedure was used as part of the network validation. The
data were randomly divided a number of times into training/validation/test
sets, and for each division a neural network was trained with random ini-
tial weights. The performances of all of the networks were analyzed to de-
termine expected future performance. In addition, those records that were
misclassified by most of the networks, regardless of the data division, were
analyzed to assist in refining the data set and improving the pattern recog-
nition.
In the next chapter, we apply neural networks to a clustering problem. We
will use a self-organizing feature map network for that application.
25-11
Further Reading
[Dubi00] D. Dubin, Rapid Interpretation of EKG's, Sixth Edition,
Tampa, FL: COVER, 2000.
This book describes the EKG in very clear terms, and leads
you through the interpretation in a step-by-step way.
[MoMa01] G.B. Moody, R.G. Mark, and A.L. Goldberger, PhysioNet:
a Web-based resource for the study of physiologic signals,
IEEE Transactions on Engineering in Medicine and Biolo-
gy, vol. 20, no. 3, pp: 70-75, 2001.
This paper describes the PhysioNet data base that contains
a large variety of recorded physiologic signals. The data
base can be found at http://www.physionet.org/.
[Raff06] The features in the data set described in this chapter were
designed and extracted by Dr. Lionel Raff, Regents Profes-
sor of Chemistry at Oklahoma State University.
[TeSi00] J. B. Tenenbaum, V. de Silva, J. C. Langford, A Global
Geometric Framework for Nonlinear Dimensionality Re-
duction, Science, vol. 290, pp. 2319-2323, 2000.
There are several different approaches to manifold learn-
ing, in which data is mapped from a high-dimensional
space to a lower-dimensional manifold. This paper intro-
duces a method called Isomap.
25-12
Objectives
26 Case Study 4: Clustering

Objectives 26-1
Description of the Forest Cover Problem 26-2
Validation 26-7
Data Sets 26-10
Epilogue 26-11
Objectives
This chapter presents a case study in using neural networks for clustering.
In clustering problems, you want a neural network to group data by simi-
larity. For example, market segmentation can be done by grouping people
according to their buying patterns, data mining can be done by partitioning
data into related subsets, and bioinformatic analysis can be done by group-
ing genes with related expression patterns.
In this chapter, we will apply clustering to a problem in forestry, in which
we would like to analyze forest cover types. We will use the self-organizing
feature map network of Chapter 16 to perform the clustering, and we will
demonstrate a variety of visualization tools that can be used in conjunction
with the SOFM.
26-1
Theory and Examples

This chapter presents a case study in using neural networks for clustering.
In clustering problems, we generally dont have a set of network targets
available, so clustering networks are trained by unsupervised training al-
gorithms. Instead of training a network to produce a desired response, we
want to analyze a data set to look for hidden patterns. There are many ap-
plication areas for clustering. It is widely used in data mining, in which we
analyze large data sets to identify similarities within subsets of the data.
It is used in city planning, when town councils apportion regions of the city
into areas of similar home type and land usage. It is used in image compres-
sion, in which a small set of prototype sub-images are identified and com-
bined to represent a large collection of images. It is used in speech
recognition systems, in which speakers are clustered into categories in or-
der to simplify the problem of speaker-independent recognition. Clustering
is used by marketers to identify distinct groups in their customer bases. It
has also been used to organize large bibliographic data bases so that relat-
ed material can be quickly accessed.
The neural network that we will use in this application is the self-organiz-
ing feature map (SOFM), which we introduced in Chapter 16. This cluster-
ing network has a unique attribute that enables us to visualize large data
sets in many dimensions. We will focus on that visualization capability in
this case study.
Description of the Forest Cover Problem

An important job of the forest service is to maintain accurate natural re-
source inventory information. One key characteristic that is recorded is the
type of forest cover found in wilderness areas. This type of data can be ex-
pensive to collect, since it generally requires on-site inspection or estima-
tion from remotely sensed data. [BlDe99] describes how forest cover type
can be predicted from independent variables that can be more easily ob-
tained. In this chapter, we will use the data described in that paper to per-
form a clustering analysis. We will demonstrate how an analysis of the
data using an SOFM can allow us to visualize the high-dimensional space
of independent variables and identify relationships between the forest cov-
er types.
Ten independent variables that can indicate forest cover type are shown in
Table 26.1. (There were 12 variables used in [BlDe99], but for ease of pre-
sentation, we selected only the first 10 for this case study.) These variables
can be measured or estimated much more easily than forest cover type. We
want to use the SOFM to find out whether these variables can be used to
cluster the data in such a way as to separate regions with different forest
cover types.
26-2
Description of the Forest Cover Problem
Variable Number Description Units
1 Elevation in meters meters
2 Aspect in degrees azimuth azimuth
3 Slope in degrees degrees
4 Horz Dist to nearest surface water meters
5 Vert Dist to nearest surface water meters
6 Horz Dist to nearest roadway meters
7 Hillshade index at 9am, summer solstice 0 to 255 index
8 Hillshade index at noon, summer solstice 0 to 255 index
9 Hillshade index at 3pm, summer solstice 0 to 255 index
10 Horz Dist to nearest wildfire ignition points meters
Table 26.1 Description of Independent Variables
The forest cover types of interest in [BlDe99] are shown in Table 26.1. The
data set we will use for this case study contains information on cover type,
but we will not use this as part of the training process. We will use it to test
the clustering ability of the SOM.
Label Name
0 Krummholz
1 Spruce/Fir
2 Lodgepole Pine
3 Ponderosa Pine
4 Cottonwood/Willow
5 Aspen
6 Douglas-fir
Table 26.2 Forest Cover Types
26-3

The data used in this study came originally from [HeBa99]. It contains the
forest cover type for 30 x 30 meter cells obtained from US Forest Service
(USFS) Region 2 Resource Information System (RIS) data. The original
data set contained 581,012 observations of 12 independent variables and
the forest cover type. We have used the first 20,000 observations, and we
have used only the first 10 independent variables, which are described in
Table 26.1. The forest cover types are given in Table 26.2. As mentioned
previously, we did not use these for training.
For supervised learning, as demonstrated in the previous three chapters,
after the data is collected, the next step is to divide the data into training,
validation and test sets. For unsupervised learning, we dont generally di-
vide the data in this way, because there is no need for a validation set to
stop the training. Competitive training is typically performed for a fixed
number of iterations. We use the entire data set for training.
The next step is to normalize the data. The data were scaled using Eq.
(22.1), so that the inputs were in the range [-1,1]. (For the SOFM, data can
also be scaled using Eq. (22.2), so that the input variables have a mean of
0 and a variance of 1.)
Before proceeding to train the network, it is often useful to view the input
data. One convenient format for this is the scatter plot. Figure 26.1 illus-
trates a set of scatter plots among the input variables 7, 8 and 9. The diag-
onal plots in this figure are histograms for these three input variables, and
the off-diagonal plots are the scatter plots. (We only show three of the vari-
ables in this figure because of the limits of the page size.)
7
8000 1 1
6000 0.5 0.5
4000 0 0
8
2000 0.5 0.5
0 1 1
1 0 1 1 0 1 1 0 1
7 7
8
8000 1
6000 0.5
4000 0
9
2000 0.5
0 1
1 0 1 1 0 1
8
9
6000
4000
2000
0
1 0 1
Figure 26.1 Scatter Plots for Input Variables 7, 8, 9.
26-4
There are several things we can look for in Figure 26.1. First, we want to
see how well the data is scattered throughout the range. If some variables
show little or no variation, then we would remove them from the analysis.
We also look for correlation between the variables. For example, if the
points in the scatter plot fell exactly along a line, then we would know that
the two variables were linearly related. There would be no need to use both
variables in the analysis. From Figure 26.1 we can see that there is some
correlation between the variables, but they are not linearly dependent.

We will use the SOFM network, described in Chapter 16, to perform the
clustering for this case study. The specific architecture is often selected
based, in part, on the number of data points, so that there will be a reason-
able amount of data associated with each prototype vector. (Recall that
each row of the weight matrix represents a prototype vector. An input is as-
sociated with the prototype vector to which it is closest.) As the data set size
increases, the number of neurons should increase as well, although not as
rapidly. A rule of thumb is to have the number of neurons increase as the
square root of the number of data points.
Figure 26.2 shows the architecture of the network we selected. We have 10
input variables (defined in Table 26.1), and we are using 150 neurons. The
feature map is 15x10, and it uses an hexagonal arrangement of neurons.
This means that each internal neuron will have six neighbors.
Hexagonal Feature Map
Inputs Competitive Layer
W
150 x 10
p n a
10 x 1
||dist||
150 x 1
C 150 x 1
10 150
ni = ||iw-p||
a = compet(n)
15 x 10
Figure 26.2 SOM Network Architecture
After the network has been trained, we will analyze the results to deter-
mine if the network architecture is satisfactory. In practical cases, we often
try several different architectures. Unlike with supervised training, in
which we have a clear performance measure (normally sum square error),
there is no firm criterion for best performance in SOFM networks. Often
26-5
what we are looking for is insight into the data set. It is somewhat of an art
to selecting the best architecture and training regime for SOFM networks.
This will become more clear when we analyze the results of the trained net-
work in a later section.

Before beginning the training, the weight vectors (rows of W) were initial-
ized using what is called linear initialization [Koho95]. First, a covariance
matrix of the input vectors was computed. Then, the two eigenvectors of
this matrix having the two largest eigenvalues were found. The rows of W
were then assigned by taking the average of the input vectors and adding
linear combinations of the two eigenvectors. This places all of the initial
weight vectors in the space spanned by the two eigenvectors. This initial-
ization process produces quicker training convergence than when using a
purely random weight initialization. (It is also possible to randomly select
input vectors from the training set to be the initial weight vectors.)
Recall from Eq. (16.21) the SOFM learning rule, which we repeat here:
i N i d (26.1)
= 1 iw q 1 + p q
where i is the index of the winning neuron, and
N i d = j d ij d (26.2)
defines the neuron neighborhood. For this case study, we have used a batch
form of the algorithm, in which all of the inputs in the training set are ap-
plied to the network before the weights are updated. To develop the batch
form, we can first modify the sequential form of Eq. (26.1) to
iw q = iw q 1 + h i i p q iw q 1 , (26.3)
where h i i is the neighborhood function. The neighborhood function that
would produce Eq. (26.1) is
i N i d
h i i = (26.4)
0 i N i d
Using this definition of neighborhood function, we can define a batch ver-

sion of Eq. (26.1):
Q
hi q i p q
iw k
q=1
-,
= ------------------------------------
Q
(26.5)
h i q i
q=1
26-6
Validation
where k is the iteration number and i q is the winning neuron for input
p q . Note that for the batch algorithm we have to distinguish between the
iteration number and the input number, since all inputs are applied to the
network at each iteration. This is in contrast to the sequential algorithm of
Eq. (26.1), where there is one iteration for each input. Also, notice that the
learning rate does not affect the batch algorithm, since it would appear in
both the numerator and the denominator of Eq. (26.5).
For the neighborhood function of Eq. (26.4), this batch algorithm has the
effect of assigning each weight to the average of the input vectors for which
it is in the neighborhood of the winner. As with the sequential algorithm,
the neighborhood size is decreased during training. The neighborhood size
is set large at the beginning of training until all weights move into the re-
gion of the input space where the data lies. Then the neighborhood size is
reduced, to fine-tune the position of the weights.
The batch algorithm requires many fewer iterations than the sequential al-
gorithm, although each iteration requires much more computation. For
this case study, we used two iterations of the batch algorithm. During the
first iteration the neighborhood size was 4, and during the second iteration
the neighborhood size was reduced to 1.
Validation
We will consider two numerical measures of the quality of a trained SOM:
resolution and topology preservation (see page 22-23). One measure of
SOM resolution is the quantization error, which is the average distance be-
tween each data vector and its winning neuron. If the average distance is
too large, then there are many input vectors that are not adequately repre-
sented by any of the prototypes.
A measure of SOM topology preservation is the topographic error. This is
the proportion of all input vectors for which the closest (winning) neuron
and the next closest neuron are not adjacent to each other in the feature
map topology. When this number is small, it means that the neurons that
are neighbors in the topology are also neighbors in the input space. It is im-
portant that this topology be preserved, so that the visualization tools we
will discuss later can provide valid insight into the data set.
For our trained SOM, the final quantization error was 0.535, and the final
topographic error was 0.037. This means that for less than 4% of all input
vectors, the winning neuron and the next closest neuron were not adjacent
to each other. It appears that the SOM has achieved the correct topology by
the completion of the training.
There are a number of visualization methods that can be used to assess the
trained SOM network. One of the key tools is called the unified distance
matrix, or u-matrix. This is a figure that shows the distance between neigh-
boring neurons in the feature map. The u-matrix has a cell for each neuron
26-7
in the feature map and an additional cell between each pair of neurons. The
cells between neurons are color-coded with the distance between the corre-
sponding weight vectors. The cells that represent the neurons are coded
with the mean of the surrounding values. Figure 26.3 shows the u-matrix
for our trained SOM.
0.6
0.5
0.4
0.3
0.2
0.1
Figure 26.3 U-Matrix for Trained SOM

In Figure 26.3, the light-colored cells represent large distances between
neurons. We can see that there is a string of light colored cells on the left
side of the feature map. This indicates that the clusters associated with the
neurons on the left side of the map are significantly different than those in
the middle and right sides of the map. For this data set we actually know
the forest cover types for each data point. We can label the feature map
cells with the cover type that is associated with the closest input vector to
that cluster center. The resulting labeled map is shown in Figure 26.4.
By comparing Figure 26.4 with Figure 26.3, we can see that the forest cover
type 2 (see Table 26.2) is associated with the left edge of the feature map.
As we move from left to right across the map, we see type 0 and 1 coded into
the center section, followed by types 5, 3 and 4, with type 6 located mainly
in the upper right section of the map. It is clear that the SOM has learned
to cluster the data according to forest cover type.
To get more insight into how the SOM has clustered the data, we can pro-
duce a hit histogram. For this graph, we count how many times each neu-
ron was the winning neuron for the entire data set. Since our data is
labeled with forest cover type, we can also see where each type falls on the
feature map. Such a graph is displayed in Figure 26.5. In each cell you can
see a hexagram with a certain gray-scale. The size of the hexagrams indi-
cate how many times the corresponding neuron was the winning neuron.
The gray level of the hexagram indicates the forest cover type. The darkest
hexagons correspond to type 0 forest covers, and the lightest hexagons cor-
respond to type 6 forest covers. We can see that the various regions of the
26-8
Validation
map have consistent colors. The left side has medium gray levels, corre-
sponding to type 2 cover. The darkest levels are in the center-left region of
the map, which corresponds to cover types 0 and 1. The lighter levels,
which correspond to types 5 and 6, are in the center-right region, and the
median levels of gray, corresponding to types 3 and 4, are on the right edge.
2 2 1 1 1 1 6 6 6 3
2 0 1 0 1 5 6 6 6 3
2 0 1 1 0 1 5 6 6 4
0 0 1 1 0 5 5 3 3 4
2 2 0 1 0 2 5 3 3 4
2 0 0 0 0 5 5 3 3 4
2 2 0 0 0 0 5 5 3 4
2 2 1 0 1 0 5 5 3 4
2 2 1 0 0 0 5 5 3 4
2 2 1 0 0 0 5 5 4 4
2 2 0 0 0 0 2 5 4 4
2 2 0 0 0 1 5 6 4 4
2 2 2 0 1 0 1 5 6 4
1 2 2 1 0 2 2 5 6 4
2 2 2 2 0 0 1 5 5 5
Figure 26.4 Labeled SOM

For many problems, we would not be able to label each input vector. The
point here is that the SOM has been able to cluster the data into similar
cover types, without knowing what the actual cover types were. This means
that the 10 variables making up the input vectors have enough correlation
with cover type to allow the SOM to make a useful clustering of the data.
Another tool that is useful in analyzing the trained SOM is the component
plane. A component plane is a figure that represents a column of the weight
matrix of the SOM. Each column corresponds to one element of the input
vector; the jth element of column i represents the connection from input i
to neuron j. In a component plane, each element of the weight is represent-
ed by a cell in the feature map topology at the location of the neuron to
which it is connected. The gray level of the cell represents the magnitude
of that element of the weight vector.
The ten component planes (one for each column of the weight matrix - each
element of the input vector) for the trained SOM are shown in Figure 26.6.
The first thing that we notice is that each of the columns is distinct. There
are no two columns that have the same pattern. We can also see that input
variables 1, 4, 5, 6 and 10 seem to be important in separating type 2 cover
types from the rest of the data. They show patterns in which a boundary
appears on the left edge of the feature map, where the type 2 cover types
are clustered. By going back to Table 26.1, we can then locate the appropri-
ate variables to see if we can deduce their connection to type 2 cover.
26-9
Figure 26.5 Hit Histogram for Trained SOM
1 2 3 4
0.4
0.5 0.2
0
0 0.4 0.6
0.4 0.5 0.6 0.8
5 6 7 8
0.5
0.4 0.8 0.8
0.45 0 0.6
0.4
0.5 0.5 0.4
9 10
0.4 0.5
0 0
0.5
0.4
Figure 26.6 Component Planes for the Trained SOM
Data Sets
cover_p.txt contains the input vectors in the data set
cover_t.txt contains the targets (labels) in the data set
Appendix C.
26-10
Epilogue
Epilogue
This chapter has demonstrated the use of SOM networks for clustering, in
which input vectors in a data set are arranged so that similar vectors are
placed in the same cluster. In this case study, the SOM was used to cluster
forestry data. The idea was to cluster land into similar forest cover types.
One of the principal advantages of the SOM network, in addition to its abil-
ity to efficiently cluster a data set, is its ability to enable visualization of
high dimensional data sets.
In the next chapter, we apply neural networks to a prediction problem. We
will use a Nonlinear Autoregressive model with eXogenous inputs (NARX)
network for that application. 26
26-11
Further Reading
[BlDe99] J. A. Blackard and D. J. Dean, Comparative Accuracies of
Artificial Neural Networks and Discriminant Analysis in
Predicting Forest Cover Types from Cartographic Vari-
ables, Computers and Electronics in Agriculture, vol. 24,
pp. 131-151, 1999.
This study compared neural networks and discriminant
analysis for predicting forest cover types from cartographic
variables. The study evaluated four wilderness areas in the
Roosevelt National Forest, located in the Front Range of
northern Colorado.
[HeBa99] S. Hettich and S. D. Bay, The UCI KDD Archive [http://
kdd.ics.uci.edu], Irvine, CA: University of California, De-
partment of Information and Computer Science, 1999.
The UCI Knowledge Discovery in Databases Archive. This
is an online repository of large data sets which encompass-
es a wide variety of data types, analysis tasks, and applica-
tion areas. It is maintained by the University of California,
Irvine.
[Koho93] T. Kohonen, Things you haven't heard about the Self-Or-
ganizing Map, Proceedings of the International Conference
on Neural Networks (ICNN), San Francisco, pp. 1147-1156,
1993.
This paper describes the batch form of the SOM learning
rule, as well as other variations on the SOM.
[Koho95] T. Kohonen, Self-Organizing Map, 2nd ed., Springer-Ver-
lag, Berlin, 1995.
This text describes the theory and practical operation of the
Self-Organizing Map in detail. It also has a chapter on the
Learning Vector Quantization algorithms.
26-12
Objectives

Objectives 27-1
Description of the Magnetic Levitation System 27-2
Validation 27-8
Data Sets 27-13
Epilogue 27-14
Objectives
This chapter presents a case study in using neural networks for prediction.
Prediction is a kind of dynamic filtering, in which past values of one or
more time series are used to predict future values. Dynamic networks, such
as those described in Chapter 10 and Chapter 14, are used for filtering and
prediction. Unlike the previous case studies, the input to these dynamic
networks is a time sequence.
There are many applications for prediction. For example, a financial ana-
lyst might want to predict the future value of a stock, bond, or other finan-
cial instrument. An engineer might want to predict the impending failure
of a jet engine. Predictive models are also used for system identification (or
dynamic modeling), in which we build dynamic models of physical systems.
These dynamic models are important for analysis, simulation, monitoring
and control of a variety of systems, including manufacturing systems,
chemical processes, robotics and aerospace systems. In this chapter we will
demonstrate the development of predictive models for a magnetic levita-
tion system.
27-1
Theory and Examples

This chapter presents a case study in using neural networks for prediction.
In this case study, the predictor neural network is used to model a dynamic
system. This dynamic modeling from data is referred to as system identifi-
cation. System identification can be applied to a variety of systems: eco-
nomic, aerospace, biological, transportation, communication,
manufacturing, chemical process, etc. For this case study, we will consider
a simple magnetic levitation system. Magnetic levitation has been used for
many years in transportation systems. In our simple maglev system, we
will suspend a magnet above an electromagnet. A maglev train works on a
similar principle.
Description of the Magnetic Levitation System

The objective of this maglev system is to control the position of a magnet
suspended above an electromagnet, where the magnet is constrained so
that it can only move in the vertical direction, as shown in Figure 27.1.
y(t)
+
i(t)
-
Figure 27.1 Magnetic Levitation System

The equation of motion for this system is
2 2
d yt i t sgn i t dy t
- = g + ----- --------------------------------- ----- ------------
-------------- (27.1)
dt
2 M yt M dt
where y t is the distance of the magnet above the electromagnet, i t is

the current flowing in the electromagnet, M is the mass of the magnet, and
g is the gravitational constant. The parameter is a viscous friction coef-
ficient that is determined by the material in which the magnet moves, and
is a field strength constant that is determined by the number of turns of
wire on the electromagnet and the strength of the magnet. For our case
study, the parameter values are set to = 12 , = 15 , g = 9.8 , M = 3 .
27-2
The objective of the case study will be to develop a dynamic neural network
model that can predict the next value of the magnet position, based on pre-
vious values of the magnet position and the input current. Once the model
has been developed, it can be used to find a controller that can determine
the correct current to apply to the electromagnet, so as to move the magnet
to some desired position. We wont go in to the control design in this case
study, but the reader is referred to [HaDe02] and [NaMu97].

For this case study, we did not build the maglev system of Figure 27.1. In-
stead, we created a computer simulation to implement Eq. (27.1). We used
Simulink to implement the simulation, but any simulation tool could be
used. For our simulations, the current was allowed to range from -1 to 4
amps. The data were collected every 0.01 seconds.
To develop an accurate model, we need to be sure that the system inputs
and outputs cover the operating range for which the system will be used.
For system identification problems, we often collect training data while ap-
plying random inputs which consist of a series of pulses of random ampli-
tude and duration. (This form of input is sometimes called a skyline
function, because of its resemblance to a city skyline.) The duration and
amplitude of the pulses must be chosen carefully to produce accurate iden-
tification. Figure 27.2 shows the input current, and the corresponding mag-
net position for our data set. A total of 4000 data points were collected.
4
Current (A)
0 10 20 30 40
Position (m)
0
0 10 20 30 40
Time (Sec)
Figure 27.2 Magnetic Levitation Data

The skyline form of input function has the advantage that it can explore
both transient and steady state operation of the system. Because some of
the pulses are long, the system will approach steady state at the end of
27-3
those pulses. Shorter width pulses explore transient operation of the sys-
tem.
When steady state performance is poor, it is useful to increase the duration
of the input pulses. Unfortunately, within a training data set, if we have
too much data in steady state conditions, the training data may not be rep-
resentative of typical plant behavior. This is due to the fact that the input
and output signals do not adequately cover the region that is going to be
controlled. This will result in poor transient performance. We need to
choose the training data so that we produce adequate transient and steady
state performance. This can be done by using an input sequence with a
range of pulse widths and amplitudes.
After the data has been collected, the next step is to divide the data into
training, validation and test sets. In this case, because we will be using the
Bayesian regularization training technique, we do not need to have a vali-
dation set. We did set aside 15% of the data for testing purposes. When the
input is a time sequence, it is useful to have the testing sequence consist of
a contiguous segment of the original data set. For our tests, we used the
last 15% of the data as the testing set.
The data were scaled using Eq. (22.1), so that both the inputs and the tar-
gets were in the range [-1,1]. The resulting scaled data is shown in Figure
27.3
0.5
Current
0.5
1
0 10 20 30 40
1
Position
0.5
0.5
1
0 10 20 30 40
Time (Sec)
Figure 27.3 Scaled Data

There are many dynamic network architectures that can be used for pre-
diction. A popular architecture is the nonlinear autoregressive network
27-4
with exogenous inputs (NARX) network, which was discussed in Chapter

22. The NARX network is a recurrent dynamic network, with feedback con-
nections enclosing several static layers of the network. The NARX model is
based on the linear ARX model, which is commonly used in time series
modeling.
The defining equation for the NARX model is
y t = f y t 1 y t 2 y t n y u t 1 u t 2 u t n u , (27.2)
where the next value of the dependent output signal y t is regressed on

previous values of the output signal and previous values of an independent
(exogenous) input signal u t . (For our application, y t is the position of
the magnet, and u t is the current going into the electromagnet.) We can
implement the NARX model by using a feedforward neural network to ap-
proximate the function f . A diagram of the resulting network is shown
in the Figure 27.4, where a two-layer feedforward network is used for the
approximation. The output of the last layer of the network is the prediction
of the next value of the magnet position. The network input is the current
into the electromagnet.
We are using the tan-sigmoid transfer function in the hidden layer, and a
linear output layer. As with the standard multilayer network, the number
1
of neurons in the hidden layer, S , will depend on the complexity of the sys-
tem being approximated. We will discuss this choice in the next section.
1 1 1 2
p (t) n (t) a (t) a (t)
T
D IW1,1 LW
2,1
2
L 1
1
S x1 1
n (t) 1x1
S x1 1xS
1x1 2
S x1
1 b1 S1x1 1 b2
1 S1x1 S1 1x1 1
T
1,3
D LW
L
Figure 27.4 NARX Network Architecture

To define the architecture, we also need to set the length of the tapped-de-
lay lines. The TDL for the inputs will contain the variables
u t 1 u t n u , and the TDL for the outputs will contain the variables
y t 1 y t n y . The TDL lengths n u and n y need to be defined. Be-
27-5
cause the defining differential equation in Eq. (27.1) is second order, we

will start with n y = n u = 2 . Later, we will investigate other possibilities.
Before demonstrating the training of the NARX network, we need to

present an important configuration that is useful in training. We can con-
sider the output of the NARX network to be an estimate of the output of the
nonlinear dynamic system that we are trying to model. The output is fed
back to the input of the feedforward neural network, as part of the stan-
dard NARX architecture, as shown on the left side of Figure 27.5. Since the
true output is available during the training of the network, we could create
a series-parallel architecture (see [NaPa90]), in which the true output is
used instead of feeding back the estimated output, as shown on the right
side of Figure 27.5. This has two advantages. The first is that the input to
the feedforward network will be more accurate. The second is that the re-
sulting network has a purely feedforward architecture, and static back-
propagation can be used for training.
T T
u(t) D u(t) D
L
Feed ^y(t) L
Feed ^y(t)
Forward Forward
T Network T Network
D y(t) D
L L
Parallel Architecture Series-Parallel Architecture
Figure 27.5 Parallel and Series-Parallel Forms

Using the series-parallel form, we can actually use a standard multilayer
network to implement the NARX model. We can create an input vector that
consists of previous system inputs and outputs:
ut 1
p = ut 2 . (27.3)
yt 1
yt 2
The target is then the next value of the output:
t = yt . (27.4)
27-6

We used the Bayesian regularization training algorithm, described in
Chapter 13, to train the NARX network, after the weights were initialized
using the Widrow/Nguyen method (see page 22-13). The prediction problem
is similar to the function approximation problem, which was demonstrated
in Chapter 23, and the Bayesian regularization method is effective for both
applications.
Because we have 4000 data points, and the number of network weights and
biases will be less than 100 (as we will see later), the chances of overfitting
are very small. We do not need to use Bayesian regularization (or early
stopping) in this case. However, because it can tell us the effective number
of parameters, we like to use it whenever it is appropriate.
Figure 27.6 illustrates the sum square error versus iteration number,
while using Bayesian regularization. We used a network with 10 neurons
1
in the hidden layer ( S = 10 ) for this case. The network was trained for
1000 iterations, at which time the performance was changing very little.
Several different networks were trained with different initial conditions,
and the final SSE was similar for each, so we can be confident that we did
not reach a local minimum.
5
10
4
10
3
10
2
10
1
10
SSE
0
10
1
10
2
10
3
10
4
10 0 1 2 3
10 10 10 10
Iterations
1
Figure 27.6 Sum Squared Error vs. Iteration Number ( S = 10 )
In Figure 27.7, we can see the variation of the effective number of param-
eters during training. It eventually converges to 39. There are a total of
61 parameters in this 4-10-1 network, so we are effectively using less than
2/3 of the weights and biases. If the effective number of parameters was
close to the total number of parameters, then we would increase the num-
ber of hidden neurons and retrain the network. That is not the case here.
27-7
There is no need to decrease the number of neurons, since the network com-
putation time is not critical for this application. The only other reason for
reducing the number of neurons would be to prevent overfitting. In terms
of preventing overfitting, having 39 effective parameters is equivalent to
having 39 total parameters. That is the beauty of using the Bayesian reg-
ularization technique. This method chooses the correct number of parame-
ters for each problem, as long as we have a sufficient number of potential
parameters in the network.
40
30
20
10
0
0 200 400 600 800 1000
Iterations
1
Figure 27.7 Effective Number of Parameters ( S = 10 )
Validation
As we discussed in previous chapters, an important tool for network vali-
dation is a scatter plot of network outputs versus targets, as shown in Fig-
ure 27.8 (in normalized units). The figure on the left shows the training
data, while the figure on the right shows the testing data. Because the test-
ing data fit is as good as the training data fit, we can be confident that the
network did not overfit.
For prediction problems, there is another set of tools for model validation.
These tools are based on two basic properties of accurate prediction models.
The first property is that the prediction errors,
2
e t = y t y t = y t a t , (27.5)
should be uncorrelated with each other from one time step to the next. The
second property is that prediction errors should be uncorrelated with the
input sequence u t . (See [BoJe86].)
27-8
Validation
If the prediction errors were correlated with each other, then we could use
that correlation to improve the predictions. For example, if prediction er-
rors one time step apart had a positive correlation, then a large positive
prediction error at the current time point would suggest that the prediction
error at the next time point would also be positive. By lowering our next
prediction, we could then reduce the next prediction error.
1 1
Training Testing
0.5 0.5
a 0 a 0
0.5 0.5
1 1
1 0.5 0 0.5 1 1 0.5 0 0.5 1
t t
Figure 27.8 Scatter Plots of Network Outputs vs. Targets - Training and Testing Sets
The same argument holds for correlation between the input sequence and
the prediction error. For accurate prediction models, there should be no cor-
relation between the input and the prediction error. If there was correla-
tion, then we could use this correlation to improve the predictor.
To measure the correlation in a time sequence, we use the autocorrelation
function, which can be estimated by
Q
1
R e = ------------
Q e t e t + . (27.6)
t=1
The autocorrelation function of the prediction error (in normalized units)

of our trained network for the maglev problem is shown in Figure 27.9.
For the prediction error to be uncorrelated (termed white noise), the au-
tocorrelation function should be an impulse at = 0 , with all other values
equal to zero. Because Eq. (27.6) provides only an estimate of the true au-
tocorrelation function, the values at 0 will never be exactly equal to ze-
ro. If the error sequence is white noise, then we can find confidence
intervals for the R e (see [BoJe86]) defined by
27-9
Re 0
2 ------------- . (27.7)
Q
8
x 10
6
Re
2
50 0 50
1
Figure 27.9 R e ( n y = n u = 2 , S = 10 )
The dashed blue lines in Figure 27.9 indicate these confidence bounds. We
can see that the estimated autocorrelation function for the prediction er-
rors falls outside these bounds at a number of points. This indicates that
we may need to increase n y and n u .
To measure correlation between the input sequence u t and the prediction
error e t , we use the cross-correlation function, which can be estimated by
Q
1
R ue = ------------
Q u t e t + . (27.8)
t=1
The cross-correlation function R ue (in normalized units) of our trained

network for the maglev problem is shown in Figure 27.9.
As with the estimated autocorrelation function, we can define confidence
intervals to determine if the cross-correlation function is near zero
Re 0 Ru 0
2 ------------------------------------ . (27.9)
Q
The dashed blue lines in Figure 27.10 represent these confidence bounds.
The cross-correlation function remains within these bounds, so it does not
indicate a problem.
27-10
Validation
5
x 10
0.8
0.4
R ue 0
0.4
0.8
50 0 50
1
Figure 27.10 R ue ( n y = n u = 2 , S = 10 )
Because the autocorrelation function of the prediction errors in Figure 27.9

indicated correlation in the errors, we increased the n y and n u values from
2 to 4 and retrained our neural network predictor. The resulting estimated
autocorrelation function is shown in Figure 27.11. Here we can see that
R e falls within the confidence bounds, except at = 0 , which indicates
that our model is performing correctly.
8
x 10
Re
50 0 50

1
Figure 27.11 R e ( n y = n u = 4 , S = 10 )
The estimated cross-correlation function, with the increased delay order, is

shown in Figure 27.12. All of the points are well within the zero confidence
27-11
interval. There is no significant correlation between the errors and the in-
put.
6
x 10
R ue 0
50 0 50

1
Figure 27.12 R ue ( n y = n u = 4 , S = 10 )
With n y = n u = 4 , we have white prediction errors, and there is no signif-

icant correlation between the prediction errors and the model input. It ap-
pears that we have an accurate prediction model.
The errors for the final prediction model are shown in Figure 27.13. We can
see that the errors are very small. However, because of the series-parallel
configuration, these are errors for only a one-step-ahead prediction. A more
stringent test would be to rearrange the network into the original parallel
form and then to perform an iterated prediction over many time steps. We
will now demonstrate the parallel operation.
Figure 27.14 illustrates the iterated prediction. The solid line is the actual
position of the magnet, and the dashed line is the position predicted by the
NARX neural network. The network prediction is very accurate - even 600
time steps ahead.
27-12
Data Sets
4
x 10
Normalized Prediction Error

2
0 10 20 30 40
Time (Sec)
Figure 27.13 Prediction Errors vs. Time
1
27
0.5
Normalized Position
0.5
1
0 2 4 6
Time (Sec)
Figure 27.14 Iterated Prediction for the Maglev NARX Network
Data Sets
maglev_u.txt contains the input sequence in the original data set
maglev_y.txt contains the output sequence in the original data set
Appendix C.
27-13
Epilogue
prediction, in which a future value of a time series is predicted from past
values of that series and potentially other series. In this case study, the
prediction network was used as a model of a magnetic levitation system.
This modeling of dynamic systems is referred to as system identification.
A Nonlinear Autoregressive model with eXogenous inputs (NARX) network
is well suited to this problem, and Bayesian regularization is an excellent
training algorithm to use in this situation.
27-14
Further Reading
Further Reading
[BoJe94] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series
Analysis: Forecasting and Control, Fourth Edition, Wiley,
2008.
A classic textbook on time series analysis and the develop-
ment of prediction models.
[HaDe02] M. Hagan, H. Demuth, O. De Jesus, An Introduction to the
Use of Neural Networks in Control Systems, International
Journal of Robust and Nonlinear Control, vol. 12, no. 11,
pp. 959-985, 2002.
This survey paper describes some practical aspects of using
neural networks for control systems. Three neural network
controllers are demonstrated: model predictive control,
NARMA-L2 control, and model reference control.
[NaMu97] Narendra, K.S.; Mukhopadhyay, S., Adaptive control us- 27
ing neural networks and approximate models, IEEE
Transactions on Neural Networks, vol. 8, no. 3, pp. 475 -
485, 1997.
This paper introduced the NARMA-L2 model and control-
ler. Once the NARMA-L2 model is trained, it can be easily
inverted to form a controller for the identified system.
[NaPa90] K. S. Narendra and K. Parthasarathy, Identification and
control of dynamical systems using neural networks,
IEEE Transactions on Neural Networks, vol. 1, no. 1, pp. 4
27, 1990.
Classic early paper on the use of neural networks for the
identification and control of dynamical systems.
27-15
A Bibliography
A Bibliography
[AgSa05] P.M. Agrawal, A.N.A. Samadh, L.M. Raff, M. Hagan, S. T.
Bukkapatnam, and R. Komanduri, Prediction of molecu-
lar-dynamics simulation results using feedforward neural
networks: Reaction of a C2 dimer with an activated dia-
mond (100) surface, The Journal of Chemical Physics 123,
224711, 2005. (Chapter 24)
[Albe72] A. Albert, Regression and the Moore-Penrose Pseudoin-
verse, New York: Academic Press, 1972. (Chapter 7)
[AmMu97] S. Amari, N. Murata, K.-R. Muller, M. Finke, and H. H.
Yang, Asymptotic Statistical Theory of Overtraining and
Cross-Validation, IEEE Transactions on Neural Net-
works, vol. 8, no. 5, 1997. (Chapter 13)
[Ande72] J. A. Anderson, A simple neural network generating an in-
teractive memory, Mathematical Biosciences, vol. 14, pp.
197220, 1972. (Chapter 1, 15, 21)
[AnRo88] J. A. Anderson and E. Rosenfeld, Neurocomputing: Foun-
dations of Research, Cambridge, MA: MIT Press, 1989.
(Chapter 1, 10)
[AnSi77] J. A. Anderson, J. W. Silverstein, S. A. Ritz and R. S. Jones,
Distinctive features, categorical perception, and probabil-
ity learning: Some applications of a neural model, Psycho-
logical Review, vol. 84, pp. 413451, 1977. (Chapter 21)
[Barn92] E. Barnard, Optimization for training neural nets, IEEE
Transactions on Neural Networks, vol. 3, no. 2, pp. 232
240, 1992. (Chapter 12)
[BaSu83] A. R. Barto, R. S. Sutton and C. W. Anderson, Neuronlike
adaptive elements that can solve difficult learning control
problems, IEEE Transactions on Systems, Man, and Cy-
bernetics, vol. 13, pp. 834846, 1983. (Chapter 4)
[Batt92] R. Battiti, First and second order methods for learning:
Between steepest descent and Newtons method, Neural
Computation, vol. 4, no. 2, pp. 141166, 1992. (Chapter 9,
12)
[Bish91] C. M. Bishop, Improving the generalization properties of
radial basis function neural networks, Neural Computa-
tion, vol. 3, no. 4, pp. 579-588, 1991. (Chapter 17)
A-1
A Bibliography
[Bish95] C.M. Bishop, Neural Networks for Pattern Recognition, Ox-

ford University Press,1995. (Chapter 22)
[BlDe99] J. A. Blackard and D. J. Dean, Comparative Accuracies of
Artificial Neural Networks and Discriminant Analysis in
Predicting Forest Cover Types from Cartographic Vari-
ables, Computers and Electronics in Agriculture, vol. 24,
pp. 131-151, 1999. (Chapter 26)
[BoJe94] G.E.P. Box, G.M. Jenkins, and G.C. Reinsel, Time Series
Analysis: Forecasting and Control, 4th Edition, John Wiley
& Sons, 2008. (Chapter 22, 27)
[BrLo88] D.S. Broomhead and D. Lowe, Multivariable function in-
terpolation and adaptive networks, Complex Systems,
vol.2, pp. 321-355, 1988. (Chapter 17)
Cliffs, NJ: Prentice-Hall, 1991. (Chapter 4, 5, 6, 8, 9, 20)
[CaGr87a] G. A. Carpenter and S. Grossberg, A massively parallel ar-
chitecture for a self-organizing neural pattern recognition
machine, Computer Vision, Graphics, and Image Process-
ing, vol. 37, pp. 54115, 1987. (Chapter 19)
[CaGr87b] G. A. Carpenter and S. Grossberg, ART2: Self-organiza-
tion of stable category recognition codes for analog input
patterns, Applied Optics, vol. 26, no. 23, pp. 49194930,
1987. (Chapter 19)
[CaGr90] G. A. Carpenter and S. Grossberg, ART3: Hierarchical
search using chemical transmitters in self-organizing pat-
tern recognition architectures, Neural Networks, vol. 3,
no. 23, pp. 129152, 1990. (Chapter 19)
[CaGrMa92] G. A. Carpenter, S. Grossberg, N. Markuzon, J. H. Rey-
nolds and D. B. Rosen, Fuzzy ARTMAP: A neural network
architecture for incremental learning of analog multidi-
mensional maps, IEEE Transactions on Neural Networks,
vol. 3, pp. 698713, 1992. (Chapter 19)
[CaGrRe91] G.A. Carpenter, S. Grossberg and J. Reynolds, ARTMAP:
Supervised real-time learning and classification of nonsta-
tionary data by a self-organizing neural network, Neural
Networks, vol. 4, pp. 565588, 1991. (Chapter 19)
[Char92] C. Charalambous, Conjugate gradient algorithm for effi-
cient training of artificial neural networks, IEEE Proceed-
ings, vol. 139, no. 3, pp. 301310, 1992. (Chapter 12)
A-2
A Bibliography
[ChCo91] S. Chen, C.F.N. Cowan, and P.M. Grant, Orthogonal least

squares learning algorithm for radial basis function net-
works, IEEE Transactions on Neural Networks, vol.2,
no.2, pp.302-309, 1991. (Chapter 17)
[ChCo92] S. Chen, P. M. Grant, and C. F. N. Cowan, Orthogonal
least squares algorithm for training multioutput radial ba-
sis function networks, Proceedings of the Institute of Elec-
trical Engineers, vol. 139, Pt. F, no. 6, pp. 378384, 1992.
(Chapter 17)
[ChCh96] S. Chen, E. S. Chng, and K. Alkadhimi, Regularised or-
thogonal least squares algorithm for constructing radial
basis function networks, International Journal of Control,
vol. 64, no. 5, pp. 829837, 1996. (Chapter 17)
[ChCo99] S. Chen, C.F.N. Cowan, and P.M. Grant, Combined Genet-
ic Algorithm Optimization and Regularized Orthogonal
Least Squares Learning for Radial Basis Function Net-
works, IEEE Transactions on Neural Networks, vol.10,
no.5, pp.302-309, 1999. (Chapter 17)
Man, and Cybernetics, vol. 13, no. 5, pp. 815826, 1983.
(Chapter 20, 21)
[DARP88] DARPA Neural Network Study, Lexington, MA: MIT Lin-
coln Laboratory, 1988. (Chapter 1)
[DeHa07] O. De Jess and M. Hagan, Backpropagation Algorithms
for a Broad Class of Dynamic Networks, IEEE Transac-
tions on Neural Networks, vol. 18, no. 1, pp., 2007. (Chapter
14)
[Dubi00] D. Dubin, Rapid Interpretation of EKG's, Sixth Edition,
Tampa, FL: COVER, 2000. (Chapter 25)
[Fahl89] S. E. Fahlman, Fast learning variations on back-propaga-
tion: An empirical study, in Proceedings of the 1988 Con-
nectionist Models Summer School, D. Touretzky, G. Hinton
and T. Sejnowski, eds., San Mateo, CA: Morgan Kaufmann,
pp. 3851, 1989. (Chapter 12)
[FoGi07] L. Fortuna, P. Giannone, S. Graziani, M. G. Xibilia, Virtu-
al Instruments Based on Stacked Neural Networks to Im-
prove Product Quality Monitoring in a Refinery, IEEE
Transactions on Instrumentation and Measurement, vol.
56, no. 1, pp. 95101, 2007. (Chapter 23)
A-3
A Bibliography
[FoHa97] D. Foresee and M. Hagan, Gauss-Newton Approximation

to Bayesian Learning, Proceedings of the 1997 Interna-
tional Joint Conference on Neural Networks, vol. 3, pp.
1930 - 1935, 1997. (Chapter 13)
[FrSk91] J. Freeman and D. Skapura, Neural Networks: Algorithms,
Applications, and Programming Techniques, Reading, MA:
Addison-Wesley, 1991. (Chapter 16)
[Gill81] P. E. Gill, W. Murray and M. H. Wright, Practical Optimi-
zation, New York: Academic Press, 1981. (Chapter 8, 9)
[GoLa98] C. Goutte and J. Larsen, Adaptive Regularization of Neu-
ral Networks Using Conjugate Gradient, Proceedings of
the IEEE International Conference on Acoustics, Speech
and Signal Processing, vol. 2, pp. 1201-1204, 1998. (Chap-
ter 13)
[GrMi89] S. Grossberg, E. Mingolla and D. Todorovic, A neural net-
work architecture for preattentive vision, IEEE Transac-
tions on Biomedical Engineering, vol. 36, no. 1, pp. 6584,
Jan. 1989. (Chapter 18)
[Gros67] S. Grossberg, Nonlinear difference-differential equations
in prediction and learning theory, Proceedings of the Na-
(Chapter 21)
[Gros68] S. Grossberg, Some physiological and biochemical conse-
quences of psychological postulates, Proceedings of the Na-
(Chapter 15)
134, 1976. (Chapter 1, 18, 19)
[Gros80] S. Grossberg, How does the brain build a cognitive code?
Psychological Review, vol. 88, pp. 375407, 1980. (Chapter
1)
Publishing Co., 1982. (Chapter 15, 18, 19)
[Gros90] S. Grossberg, Neural networks: From foundations to ap-
plications, Short-Course Notes, Boston University, Bos-
ton, MA, May 611, 1990. (Chapter 18)
[HaBo07] L. Hamm, B. W. Brorsen and M. T. Hagan, Comparison of
Stochastic Global Optimization Methods to Estimate Neu-
A-4
A Bibliography
ral Network Weights, Neural Processing Letters, vol. 26,

no. 3, December 2007. (Chapter 22)
[HaDe02] M. Hagan, H. Demuth, O. De Jesus, An Introduction to the
Use of Neural Networks in Control Systems, International
Journal of Robust and Nonlinear Control, vol. 12, no. 11,
pp. 959-985, 2002. (Chapter 27)
[HaMe94] M. T. Hagan and M. Menhaj, Training feedforward net-
works with the Marquardt algorithm, IEEE Transactions
on Neural Networks, vol. 5, no. 6, pp. 989993, 1994. (Chap-
ter 12)
[HeBa99] S. Hettich and S. D. Bay, The UCI KDD Archive [http://
kdd.ics.uci.edu], Irvine, CA: University of California, De-
partment of Information and Computer Science, 1999.
(Chapter 26)
[Hebb 49] D. O. Hebb, The Organization of Behavior, New York:
Wiley, 1949. (Chapter 1, 7, 15)
[Hech90] R. Hecht-Nielsen, Neurocomputing, Reading, MA: Addi-
son-Wesley, 1990. (Chapter 16)
[HeOh97] B. Hedn, H. hlin, R. Rittner, L. Edenbrandt, Acute My-
ocardial Infarction Detected in the 12-Lead ECG by Artifi-
cial Neural Networks, Circulation, vol. 96, pp. 17981802,
1997. (Chapter 22)
York: McGraw-Hill, 1972. (Chapter 8, 9)
emergent collective computational properties, Proceedings
of the National Academy of Sciences, vol. 79, pp. 2554
2558, 1982. (Chapter 1, 21)
[Hopf84] J. J. Hopfield, Neurons with graded response have collec-
tive computational properties like those of two-state neu-
rons, Proceedings of the National Academy of Sciences, vol.
81, pp. 30883092, 1984. (Chapter 21)
[HoTa85] J. J. Hopfield and D. W. Tank, Neural computation of de-
cisions in optimization problems, Biological Cybernetics,
vol. 52, pp. 141154, 1985. (Chapter 21)
[HoSt89] K. M. Hornik, M. Stinchcombe and H. White, Multilayer
feedforward networks are universal approximators, Neu-
ral Networks, vol. 2, no. 5, pp. 359366, 1989. (Chapter 11)
A-5
A Bibliography
[Hube88] D. H. Hubel, Eye, Brain, and Vision, New York: Scientific

American Library, 1988. (Chapter 18)
[Jaco88] R. A. Jacobs, Increased rates of convergence through
learning rate adaptation, Neural Networks, vol. 1, no. 4, pp
295308, 1988. (Chapter 12)
[John01] G. L. Johnson, Contributions to the comparative anatomy
of the mammalian eye, chiefly based on ophthalmoscopic
examination, Philosophical Transactions of the Royal So-
ciety of London, Series B., vol. 194, pp. 182, Plate 1, 1901.
(Chapter 18)
[Joll02] I.T. Jolliffe, Principal Component Analysis, Springer Series
in Statistics, 2nd ed., Springer, NY, 2002. (Chapter 22)
actions on Computers, vol. 21, pp. 353359, 1972. (Chapter
1, 15, 21)
2nd Ed., Berlin: Springer-Verlag, 1987. (Chapter 15, 16)
[Koho93] T. Kohonen, Things you haven't heard about the Self-Or-
ganizing Map, Proceedings of the International Conference
on Neural Networks (ICNN), San Francisco, pp. 1147-1156,
1993. (Chapter 26)
[Koho95] T. Kohonen, Self-Organizing Map, 2nd ed., Springer-Ver-
lag, Berlin, 1995. (Chapter 26)
[LaSa67] J. P. LaSalle, An invariance principle in the theory of sta-
bility, in Differential Equations and Dynamic Systems, J.
K. Hale and J. P. Lasalle, eds., New York: Academic Press,
pp. 277286, 1967. (Chapter 20)
[LeCu85] Y. Le Cun, Une procedure dapprentissage pour reseau a
seuil assymetrique, Cognitiva, vol. 85, pp. 599604, 1985.
(Chapter 11)
[LeCu98] Y. LeCun, L. Bottou, G. B. Orr, K.-R. Mueller, Efficient
BackProp, Lecture Notes in Computer Science, vol. 1524,
1998. (Chapter 22)
[Leib90] D. Lieberman, Learning, Behavior and Cognition, Bel-
mont, CA: Wadsworth, 1990. (Chapter 15)
A-6
A Bibliography
[LiMi89] J. Li, A. N. Michel and W. Porod, Analysis and synthesis

of a class of neural networks: Linear systems operating on
a closed hypercube, IEEE Transactions on Circuits and
Systems, vol. 36, no. 11, pp. 14051422, November 1989.
(Chapter 21)
[Lowe89] D. Lowe, Adaptive radial basis function nonlinearities,
and the problem of generalization, Proceedings of the First
IEE International Conference on Artificial Neural Net-
works, pp. 171 - 175, 1989. (Chapter 17)
[MacK92] D. J. C. MacKay, Bayesian Interpolation, Neural Compu-
tation, vol. 4, pp. 415-447, 1992. (Chapter 13)
[MaNe99] J.R. Magnu and H. Neudecker, Matrix Differential Calcu-
lus, John Wiley & Sons, Ltd., Chichester, 1999. (Chapter
14)
[MaGa00] E. A. Maguire, D. G. Gadian, I. S. Johnsrude, C. D. Good, J.
Ashburner, R. S. J. Frackowiak, and C. D. Frith, Naviga-
tion-related structural change in the hippocampi of taxi
drivers, Proceedings of the National Academy of Sciences,
vol. 97, no. 8, pp. 4398-4403, 2000. (Chapter 1)
Biophysics, vol. 5, pp. 115133, 1943. (Chapter 1, 4, 21)
[Mill90] A.J. Miller, Subset Selection in Regression. Chapman and
Hall, N.Y., 1990. (Chapter 17)
[Mill93] M.F. Miller, A scaled conjugate gradient algorithm for fast
1993. (Chapter 24)
[MoDa89] J. Moody and C.J. Darken, Fast Learning in Networks of
Locally-Tuned Processing Units, Neural Computation,
vol. 1, pp. 281294, 1989. (Chapter 17)
[Moll93] M. Moller, A scaled conjugate gradient algorithm for fast
1993. (Chapter 22)
[MoMa01] G.B. Moody, R.G. Mark, and A.L. Goldberger, PhysioNet:
a Web-based resource for the study of physiologic signals,
IEEE Transactions on Engineering in Medicine and Biolo-
gy, vol. 20, no. 3, pp: 70-75, 2001. (Chapter 25)
MIT Press, 1969. (Chapter 1, 4)
A-7
A Bibliography
[NaMu97] Narendra, K.S.; Mukhopadhyay, S., Adaptive control us-

ing neural networks and approximate models, IEEE
Transactions on Neural Networks, vol. 8, no. 3, pp. 475 -
485, 1997. (Chapter 27)
[NaPa90] K. S. Narendra and K. Parthasarathy, Identification and
control of dynamical systems using neural networks,
IEEE Transactions on Neural Networks, vol. 1, no. 1, pp. 4
27, 1990. (Chapter 27)
2126, July 1990. (Chapter 12, 22)
[OrHa00] M. J. Orr, J. Hallam, A. Murray, and T. Leonard, Assess-
ing rbf networks using delve, IJNS, 2000. (Chapter 17)
[Park85] D. B. Parker, Learning-logic: Casting the cortex of the hu-
man brain in silicon, Technical Report TR-47, Center for
Computational Research in Economics and Management
Science, MIT, Cambridge, MA, 1985. (Chapter 11)
[PaSa93] J. Park and I.W. Sandberg, Universal approximation us-
ing radial-basis-function networks, Neural Computation,
vol. 5, pp. 305-316, 1993. (Chapter 17)
[PeCo93] M. P. Perrone and L. N. Cooper, When networks disagree:
Ensemble methods for hybrid neural networks, in Neural
Networks for Speech and Image Processing, R. J. Mam-
mone, Ed., Chapman-Hall, pp. 126-142, 1993. (Chapter 22)
[PhHa13] M. Phan and M. Hagan, Error Surface of Recurrent Net-
works, IEEE Transactions on Neural Networks and
Learning Systems, vol. 24, no. 11, pp. 1709 - 1721, October,
2013. (Chapter 14)
[Powe87] M.J.D. Powell, Radial basis functions for multivariable in-
terpolation: a review, Algorithms for Approximation, pp.
143-167, Oxford, 1987. (Chapter 17)
[PuFe97] G.V. Puskorius and L.A. Feldkamp, Extensions and en-
hancements of decoupled extended Kalman filter training,
Proceedings of the 1997 International Conference on Neural
Networks, vol. 3, pp. 1879-1883, 1997. (Chapter 22)
of Chemical Physics, vol. 122, 2005. (Chapter 22, 24)
A-8
A Bibliography
[RiIr90] A. K. Rigler, J. M. Irvine and T. P. Vogl, Rescaling of vari-

ables in back propagation learning, Neural Networks, vol.
3, no. 5, pp 561573, 1990. (Chapter 12)
logical Review, vol. 65, pp. 386408, 1958. (Chapter 1, 4)
[Rose61] F. Rosenblatt, Principles of Neurodynamics, Washington
DC: Spartan Press, 1961. (Chapter 4)
[RuHi86] D. E. Rumelhart, G. E. Hinton and R. J. Williams, Learn-
ing representations by back-propagating errors, Nature,
vol. 323, pp. 533536, 1986. (Chapter 11)
Cognition, vol. 1, Cambridge, MA: MIT Press, 1986. (Chap-
ter 1, 11, 16)
[Sarle95] W. S. Sarle, Stopped training and other remedies for over-
fitting, In Proceedings of the 27th Symposium on Interface,
1995. (Chapter 13)
York: Springer-Verlag, 1985. (Chapter 8, 9, 12)
[ScSm99] B. Schlkopf, A. Smola, K.-R. Muller, Kernel Principal
Component Analysis, in B. Schlkopf, C. J. C. Burges, A.
J. Smola (Eds.), Advances in Kernel Methods-Support Vec-
tor Learning, MIT Press Cambridge, MA, USA, pp. 327-
352, 1999. (Chapter 22)
[Shan90] D. F. Shanno, Recent advances in numerical techniques
for large-scale optimization, in Neural Networks for Con-
trol, Miller, Sutton and Werbos, eds., Cambridge, MA: MIT
Press, 1990. (Chapter 12)
[SjLj94] J. Sjoberg and L. Ljung, Overtraining, regularization and
searching for minimum with application to neural net-
works, Linkoping University, Sweden, Tech. Rep. LiTH-
ISY-R-1567, 1994. (Chapter 13)
[SlLi91] J.-J. E. Slotine and W. Li, Applied Nonlinear Control, En-
glewood Cliffs, NJ: Prentice-Hall, 1991. (Chapter 20)
[StDo84] W. D. Stanley, G. R. Dougherty and R. Dougherty, Digital
Signal Processing, Reston VA: Reston Publishing Co.,
1984. (Chapter 10)
A-9
A Bibliography

Academic Press, 1980. (Chapter 5, 6)
[TaHo86] D. W. Tank and J. J. Hopfield, Simple neural optimiza-
tion networks: An A/D converter, signal decision circuit
and a linear programming circuit, IEEE Transactions on
Circuits and Systems, vol. 33, no. 5, pp. 533541, 1986.
(Chapter 21)
[TeSi00] J. B. Tenenbaum, V. de Silva, J. C. Langford, A Global
Geometric Framework for Nonlinear Dimensionality Re-
duction, Science, vol. 290, pp. 2319-2323, 2000. (Chapter
25)
[Tikh63] A. N. Tikhonov, The solution of ill-posed problems and the
regularization method, Dokl. Acad. Nauk USSR, vol. 151,
no. 3, pp. 501-504, 1963. (Chapter 13)
[Toll90] T. Tollenaere, SuperSAB: Fast adaptive back propagation
with good scaling properties, Neural Networks, vol. 3, no.
5, pp. 561573, 1990. (Chapter 12)
[vanT75] H. F. J. M. van Tuijl, A new visual illusion: Neonlike color
spreading and complementary color induction between
subjective contours, Acta Psychologica, vol. 39, pp. 441
445, 1975. (Chapter 18)
[VoMa88] T. P. Vogl, J. K. Mangis, A. K. Zigler, W. T. Zink and D. L.
Alkon, Accelerating the convergence of the backpropaga-
tion method, Biological Cybernetics, vol. 59, pp. 256264,
Sept. 1988. (Chapter 12)
[vond73] C. von der Malsburg, Self-organization of orientation sen-
sitive cells in the striate cortex, Kybernetic, vol. 14, pp. 85
100, 1973. (Chapter 18)
[WaVe94] C. Wang, S. S. Venkatesh, and J. S. Judd, Optimal Stop-
ping and Effective Machine Complexity in Learning, Ad-
vances in Neural Information Processing Systems, J. D.
Cowan, G. Tesauro, and J. Alspector, Eds., vol. 6, pp. 303-
310, 1994. (Chapter 13)
[Werbo74] P. J. Werbos, Beyond regression: New tools for prediction
and analysis in the behavioral sciences, Ph.D. Thesis,
Harvard University, Cambridge, MA, 1974. Also published
as The Roots of Backpropagation, New York: John Wiley &
Sons, 1994. (Chapter 11)
[Werb90] P. J. Werbos, Backpropagation through time: What it is
and how to do it, Proceedings of the IEEE, vol. 78, pp.
15501560, 1990. (Chapter 14)
A-10
A Bibliography
[WeTe84] J. F. Werker and R. C. Tees, Cross-language speech per-

ception: Evidence for perceptual reorganization during the
first year of life, Infant Behavior and Development, vol. 7,
pp. 49-63, 1984. (Chapter 1)
[WhSo92] D. White and D. Sofge, eds., Handbook of Intelligent Con-
trol, New York:Van Nostrand Reinhold, 1992. (Chapter 4)
[WiHo60] B. Widrow, M. E. Hoff, Adaptive switching circuits,1960
IRE WESCON Convention Record, New York: IRE Part 4,
pp. 96104, 1960. (Chapter 1, 10)
[WiSt 85] B. Widrow and S. D. Stearns, Adaptive Signal Processing,
Englewood Cliffs, NJ: Prentice-Hall, 1985. (Chapter 10)
[WiWi 88] B. Widrow and R. Winter, Neural nets for adaptive filter-
ing and adaptive pattern recognition, IEEE Computer
Magazine, March 1988, pp. 2539. (Chapter 10)
[WiZi89] R. J. Williams and D. Zipser, A learning algorithm for con-
tinually running fully recurrent neural networks, Neural
Computation, vol. 1, pp. 270280, 1989. (Chapter 14)
A-11
B Notation
B Notation
Basic Concepts
Scalars: small italic letters.....a,b,c
Vectors: small bold nonitalic letters.....a,b,c
Matrices: capital BOLD nonitalic letters.....A,B,C
Language
Vector means a column of numbers.
Row vector means a row of a matrix used as a vector (column).
General Vectors and Transformations (Chapters 5 and 6)

x = Ay
Weight Matrices
Scalar Element
k
w i j t
i - row, j - column, k - layer, t - time or iteration
Matrix
k
W t
Column Vector
k
wj t
Row Vector
k
iw t
Bias Vector
Scalar Element
k
bi t B
Vector
k
b t
B-1
B Notation
Input Vector
Scalar Element
pi t
As One of a Sequence of Input Vectors

pt
As One of a Set of Input Vectors

pq
Net Input Vector

Scalar Element
k k
n i t or n i q
Vector
k k
n t or n q
Output Vector
Scalar Element
k k
a i t or a i q
Vector
k k
a t or a q
Transfer Function
Scalar Element
k k k
ai = f ni
Vector
k k k
a = f n
Target Vector
Scalar Element
t i t or t i q
B-2
B Notation
Vector
t t or t q
Set of Prototype Input/Target Vectors

{p 1, t 1} {p 2 , t 2} {p Q, t Q}
Error Vector
Scalar Element
e i t = t i t a i t or e i q = t i q a i q
Vector
e t or e q
Sizes and Dimensions

Number of Layers, Number of Neurons per Layer
k
M, S
Number of Input Vectors (and Targets), Dimension of Input Vector

Q, R
Parameter Vector (includes all weights and biases)

Vector
x
At Iteration k
x k or x k
Norm
x
Performance Index
F x
Gradient and Hessian

F x k = g k and 2F x k = A k
B
B-3
B Notation
Parameter Vector Change

x k = x k + 1 x k
Eigenvalue and Eigenvector

i and z i
Approximate Performance Index (single time step)

F x
Transfer Function Derivative

Scalar
d
fn = f n
dn
Matrix
f n 1
m m
0 0
m m
F n =
m m 0 f n2 0

m m
0 0 f n Sm
Jacobian Matrix
Jx
Approximate Hessian Matrix

T
H = J J
Sensitivity Vector
Scalar Element
m F
s i --------m-
n i
Vector
m F
s --------m-
n
B-4
B Notation
Marquardt Sensitivity Matrix

Scalar Element
m v h e k q
s i h ----------
m
- = -----------
m
n i q n i q
Partial Matrix (single input vector p q ) and Full Matrix (all inputs)
m m
S q and S = S m S m S m
1 2 Q
Dynamic Networks
Sensitivity
e u
u m ak t
s k i t ----------------
m
-
n i t
Weight Matrices
m l
IW d - input weight between input l and layer m at delay d
m l
LW d - layer weight between layer l and layer m at delay d
Index Sets
DL m l - delays in the tapped delay line between Layer l and Layer m.
DI m l - delays in the tapped delay line between Input l and Layer m.
I m - indices of input vectors that connect to layer m.
f
L m - indices of layers that directly connect forward to layer m.
b
L m - indices of layers that are directly connected backwards to layer
m (or to which layer m connects forward) and that contain no de-
lays in the connection.
U x u
E LW x = u U LW d 0 d 0
X u x
ES u = x X S 0
u x
E S u = x S 0
X
E LW u = x X LW
x u
d 0 d 0 B
U u x
ES x = u U S 0
B-5
B Notation
Definitions
Input Layer (X) - has an input weight, or contains any delays with
any of its weight matrices
Output Layer (U) - its output will be compared to a target during
training, or it is connected to an input layer through a matrix that
has delays associated with it.
Parameters for Backpropagation and Variations

Learning Rate and Momentum
and
Learning Rate Increase, Decrease and Percentage Change

, and
Conjugate Gradient Direction Adjustment Parameter

k
Marquardt Parameters
and
Generalization
Regularization Parameters

, and = ---


Selected Model
M
Sum Squared Error and Sum Squared Weights

ED , EW
Maximum Likelihood and Most Probable Weights

ML MP
x ,x
B-6
B Notation
Feature Map Terms

Distance Between Neurons
d ij - distance between neuron i and neuron j
Neighborhood
N i d = j d ij d
Grossberg and ART Networks

On-Center and Off-Surround Connection Matrices
1 0 0 0 1 1
0 1 0 1
W = 0 1
+ 1 - 1
and W =

0 0 1 1 1 0
Excitatory and Inhibitory Biases

+ -
b and b
Time Constant

Relative Intensity
1
S
p
p i = ----i where P =
P pj
j=1
Instar and Outstar Weight Matrices

1:2 2:1
W and W
Orienting Subsystem Parameters

, and = --- (vigilance)

ART1 Learning Law Parameter

B
B-7
B Notation
Lyapunov Stability
Lyapunov Function
Va
Zero Derivative Set, Largest Invariant Set and Closure

Z , L and L
Bounded Lyapunov Function Set

= a:V a
Hopfield Network Parameters

Circuit Parameters
T i j , C , R i , I i ,
Amplifier Gain

B-8
Introduction
C Software
Introduction
We have used MATLAB, a numeric computation and visualization soft-
ware package, in this text. However, MATLAB is not essential for using
this book. The computer exercises can performed with any available pro-
gramming language, and the Neural Network Design Demonstrations,
while helpful, are not critical to understanding the material covered in this
book.
MATLAB is widely available and, because of its matrix/vector notation and
graphics, is a convenient environment in which to experiment with neural
networks. We use MATLAB in two different ways. First, we have included
a number of exercises for the reader to perform in MATLAB. Many of the
important features of neural networks become apparent only for large scale
problems, which are computationally intensive and not feasible for hand
calculations. With MATLAB, neural network algorithms can be quickly im-
plemented, and large scale problems can be tested conveniently. If MAT-
LAB is not available, any other programming language can be used to
perform the exercises.
The second way in which we use MATLAB is through the Neural Network
Design Demonstrations, which can be downloaded from the website
hagan.okstate.edu/nnd.html. These interactive demonstrations illustrate
important concepts in each chapter. The icon to the left identifies referenc-
es to these demonstrations in the text.
MATLAB, or the student edition of MATLAB, version 2010a or later,
should be installed on your computer in a a folder named MATLAB. To cre-
ate this directory or folder and complete the MATLAB installation process,
follow the instructions given in the MATLAB documentation. Take care to
follow the guidelines given for setting the path.
After the Neural Network Design Demonstration software has been loaded
into the MATLAB directory on your computer (or if the MATLAB path has
been set to include the directory containing thedemonstration software),
the demonstrations can be invoked by typing nnd at the MATLAB prompt.
All demonstrations are easily accessible from a master menu.
C
C-1
C Software
Overview of Demonstration Files

Running the Demonstrations
You can run the demonstrations directly by typing their names at the
MATLAB prompt. Typing help nndesign brings up a list of all the demos you
can choose from.
Alternatively, you can run the Neural Network Design splash window (nnd)
and then click the Contents button. This will take you to a graphical Table
of Contents. From there you can select chapters with buttons at the bottom
of the window and individual demonstrations with popup menus.
Sound
Many of the demonstrations use sound. In many cases the sound adds to
the understanding of a demonstration. In other cases it is there simply for
fun. If you need to turn the sound off you can give MATLAB the following
command and all demonstrations will run quietly:
nnsound off
To turn sound back on:

nnsound on
You may note that demonstrations that utilize sound often run faster when
sound is off. In addition, on some machines which do not support sound er-
rors can occur unless the sound is turned off.
List of Demonstrations
General
nnd - Splash screen.
nndtoc - Table of contents.
nnsound - Turn Neural Network Design sounds on and off.
Chapter 2, Neuron Model and Network Architectures

nnd2n1 - One-input neuron.
nnd2n2 - Two-input neuron.
Chapter 3, An Illustrative Example

nnd3pc - Perceptron classification.
nnd3hamc - Hamming classification.
nnd3hopc - Hopfield classification.
C-2
Chapter 4, Perceptron Learning Rule

nnd4db - Decision boundaries.
nnd4pr - Perceptron rule.
Chapter 5, Signal and Weight Vector Spaces

nnd5gs - Gram-Schmidt.
nnd5rb - Reciprocal basis.
Chapter 6, Linear Transformations for Neural Networks

nnd6lt - Linear transformations.
nnd6eg - Eigenvector game.
Chapter 7, Supervised Hebbian Learning

nnd7sh - Supervised Hebb.
Chapter 8, Performance Surfaces and Optimum Points

nnd8ts1 - Taylor series #1.
nnd8ts2 - Taylor series #2.
nnd8dd - Directional derivatives.
nnd8qf - Quadratic function.
Chapter 9, Performance Optimization

nnd9sdq - Steepest descent for quadratic function.
nnd9mc - Method comparison.
nnd9nm - Newton's method.
nnd9sd - Steepest descent.
Chapter 10, Widrow-Hoff Learning

nnd10nc - Adaptive noise cancellation.
nnd10eeg - Electroencephalogram noise cancellation.
nnd10lc - Linear pattern classification.
Chapter 11, Backpropagation

nnd11nf - Network function.
nnd11bc - Backpropagation calculation.
nnd11fa - Function approximation.
nnd11gn - Generalization.
C
C-3
C Software
Chapter 12, Variations on Backpropagation

nnd12sd1- Steepest descent backpropagation #1.
nnd12sd2 - Steepest descent backpropagation #2.
nnd12mo - Momentum backpropagation.
nnd12vl - Variable learning rate backpropagation.
nnd12ls - Conjugate gradient line search.
nnd12cg - Conjugate gradient backpropagation.
nnd12ms - Maquardt step.
nnd12m - Marquardt backpropagation.
Chapter 13, Generalization

nnd13es - Early stoppinng.
nnd13reg - Regularization.
nnd13breg - Bayesian regularization.
nnd13esr - Early stopping/regularization.
Chapter 14, Dynamic Networks

nnd14fir - Finite impulse response network.
nnd14iir - Infinite impulse response network.
nnd14dynd - Dynamic derivatives.
nnd14rnt - Recurrent network training.
Chapter 15, Associative Learning

nnd15uh - Unsupervised Hebb.
nnd15edr - Effect of decay rate.
nnd15hd - Hebb with decay.
nnd15gis - Graphical instar.
nnd15is - Instar.
nnd15os - Outstar.
Chapter 16, Competitive Networks

nnd16cc - Competitive classification.
nnd16cl - Competitive learning.
nnd16fm1 - 1-D feature map.
nnd16fm2 - 2-D feature map.
nnd16lv1 - LVQ1.
nnd16lv2 - LVQ2.
Chapter 17, Radial Basis Networks

nnd17nf - Network function.
nnd17pc - Pattern classification.
nnd17lls - Linear least squares.
nnd17ols - Orthogonal least squares.
nnd17no - Nonlinear optimization.
C-4
Chapter 18, Grossberg Network

nnd18li - Leaky integrator.
nnd18sn - Shunting network.
nnd18gl1 - Grossberg layer 1.
nnd18gl2 - Grossberg layer 2.
nnd18aw - Adaptive weights.
Chapter 19, Adaptive Resonance Theory

nnd19al1 - ART1 layer 1.
nnd19al2 - ART1 layer 2.
nnd19os - Orienting subsystem.
nnd19a1 - ART1 algorithm.
Chapter 20, Stability

nnd20ds - Dynamical system.
Chapter 21, Hopfield Network

nnd21hn - Hopfield network.
C
C-5
Index
Index
A Layer 1 19-4
Abbreviated notation 2-8 Layer 2 19-10
ADALINE network 10-2 learning law
decision boundary 10-4 L1 - L2 19-17
mean square error 10-4 L2 - L1 19-20
Adaptive filtering 10-13 orienting subsystem 19-13
Adaptive noise cancellation 10-15 resonance 19-17
Adaptive resonance theory (ART) 19-2 subset/superset dilemma 19-17
Amacrine cells 18-4 summary 19-21
Amari, S. 18-2 vigilance 19-15
AND gate 4-7 ART2 19-23
Anderson, J.A. 1-2, 1-3, 15-2, 18-2 ART3 19-23
Angle 5-7 ARTMAP 19-23
Apple and Orange Example 3-2 Associative learning
Hamming network solution 3-8 Hebb rule 7-4
Hopfield solution 3-12 instar rule 15-11
perceptron 3-3 Kohonen rule 15-15
perceptron solution 3-5 outstar rule 15-17
problem statement 3-2 pseudoinverse rule 7-7
Applications of neural networks 1-5 unsupervised Hebb rule 15-5
aerospace 1-5 Associative memory 7-3
automotive 1-5 autoassociative memory 7-10
banking 1-5 Hopfield network 21-5
defense 1-6 linear associator 7-3
electronics 1-6 Associative networks 15-3
entertainment 1-6 instar 15-9
financial 1-6 outstar 15-16
insurance 1-6 Attractors 21-11
manufacturing 1-6 Autoassociative memory 7-10
medical 1-6 Autocorrelation function 22-24
oil and gas 1-6
robotics 1-7 B
securities 1-7 Backpropagation 11-7
speech 1-7 batching 12-7
telecommunications 1-7 CGBP 12-15
transportation 1-7 choice of network architecture 11-18
ART1 conjugate gradient 12-14
fast learning 19-19 convergence 11-20
Index-1
Index
delta-bar-delta 12-13 C
drawbacks 12-3, 14-3 Carpenter, G. 19-2
example 11-14 Center 17-6
generalization 11-22 CGBP 12-15
initial weights 12-6 Chain rule 11-9
Jacobian matrix 12-23 Change of basis 6-6
Levenberg-Marquardt 12-19, 12-21 similarity transformation 6-8
Jacobian calculation 12-22 Chemical vapor deposition 24-2
Marquardt sensitivity 12-24 Choice of network architecture 11-18
LMBP 12-25 Circular hollow 8-16
MOBP 12-11 Clustering 22-9
performance index 11-8 Coding the targets 22-7
performance surface 12-3 Committee of networks 22-18, 25-10
Quickprop 12-14 Competitive learning 16-7
SDBP 12-2 adaptive resonance theory 19-2
sensitivity 11-10 ART1 19-4
summary 11-13 ART2 19-23
SuperSAB 12-14 ART3 19-23
VLBP 12-12 ARTMAP 19-23
Backpropagation order 14-4 Fuzzy ARTMAP 19-23
Backpropagation-through-time (BPTT) 14-2, 14- instar rule 16-7
11, 14-22 Kohonen rule 16-7
Backward elimination 17-18 learning rate 16-9
Basis set 5-5 LVQ2 16-21
Batch SOFM learning algorithm 26-6 problems 16-9
Batch training 11-17 Competitive networks 16-5
Batching 12-7 ART1 19-4
Bayes rule 13-10 Grossberg 18-13
Bayes Theorem 13-10 Hamming network 16-3
Bayesian analysis 13-12 Lateral inhibition 16-5
effective number of parameters 13-16 learning vector quantization 16-16
evidence 13-13 self-organizing feature map 16-12
Gauss-Newton approximation to Bayesian winner-take-all 16-5
regularization 13-17 Conditioned stimulus 15-3
likelihood function 13-13 Cones 18-3
posterior density 13-13 Confusion matrix 22-21, 25-7
prior density 13-13 Conjugate directions 9-16
Bayesian regularization 13-12, 23-6 Conjugate gradient 9-15, 12-14
effective number of parameters 23-6 golden section search 12-17
Baysian analysis interval location 12-16
most probable 13-14 interval reduction 12-16
Biological inspiration of neural networks 1-8 Content-addressable memory 21-16
Bipolar cells 18-3 Contour plot 8-8
Brightness constancy 18-8 Contrast enhance 18-18
Index-2
Index
control Featural filling-in 18-6

electromagnet 27-2 Feature extraction 22-6, 25-3
Correlation matrix 10-6 Finite impulse response 14-6
Cross-correlation function 22-25 Fitting 22-8
Cross-entropy 22-17 Forest cover 26-2
Cross-validation 13-6 Forward selection 17-18
Curse of dimensionality 17-11 Fovea 18-5
Fukushima, K. 18-2
D Function approximation 11-4, 22-8
Decay rate 15-7 Fuzzy ARTMAP 19-23
Decision boundary 4-5, 10-4, 11-4
Delay 2-13 G
Delta rule 7-13, 10-7 Ganglion cells 18-4
Delta-bar-delta 12-13 Gauss-Newton algorithm 12-21
Descent direction 9-3 Jacobian matrix 12-20
Diagonalization 6-13 Gauss-Newton approximation to Bayesian regu-
Directional derivative 8-5 larization 13-17
Distortion measure 22-23 Generalization 11-22, 13-2
Domain 6-2 Golden Section search 12-17
Dynamic networks 14-2 Gradient 8-5
Gradient descent 9-2
E Gram-Schmidt orthogonalization 5-8
Early stopping 13-7, 13-19, 25-7 Grossberg competitive network 18-13
Echo cancellation 10-21 choice of transfer function 18-20
EEG 10-15 Layer 1 18-13
Effective number of parameters 13-16, 13-23, 23- Layer 2 18-17
6 learning law 18-22
Eigenvalues 6-10 relation to Kohonen law 18-24
Eigenvectors 6-10 Grossberg, S. 1-3, 15-2, 18-2, 19-2
Electrocardiagram 25-2
electromagnet 27-2 H
Elliptical hollow 8-17 Hamming network 3-8, 16-3
Emergent segmentation 18-6 feedforward layer 3-8, 16-3
Equilibrium point 20-4 recurrent layer 3-9, 16-4
Euclidean space 5-3 Hebb rule 7-4, 21-18
evidence 13-13 decay rate 15-7
Excitatory 18-10 performance analysis 7-5
Extended Kalman filter algorithm 22-14 supervised 7-4
Extrapolation 13-3, 22-21, 22-27 unsupervised 7-4, 15-5
with decay 7-12
F Hebb, D.O. 1-3, 7-2
Fahlman, A.E. 12-14 Hebbs postulate 7-2
False negative 22-21 Hebbian learning 7-2
False positive 22-22 variations 7-12
Index-3
Index
Hessian 8-5 K
eigensystem 8-13 Kohonen rule 15-15, 16-7
Hidden layer 2-11 graphical representation 16-7
High-gain Lyapunov function 21-13 Kohonen, T. 1-3, 15-2, 18-2
Hinton, G.E. 11-2
Histogram of errors 22-21 L
History of neural networks 1-2 LaSalles corollary 20-14
Hoff, M.E. 1-3, 10-2, 11-2 LaSalles invariance theorem 20-13
Hopfield model 21-3 invariant set 20-13
Hopfield network 3-12, 6-2, 21-5 Set
attractors 21-11 L 20-13
design 21-16 Z 20-12
content-addressable memory 21-16 Lateral inhibition 16-5
effect of gain 21-12 Layer 2-9
example 21-7 competitive 16-5
Hebb rule 21-18 problems 16-9
high-gain Lyapunov function 21-13 hidden 2-11
Lasalles invariance theorem 21-7 output layer 2-11
Lyapunov function 21-5 superscript 2-11
Lyapunov surface 21-22 Layer weight 14-3
spurious patterns 21-20 Layered Digital Dynamic Network (LDDN) 14-3
Hopfield, J.J. 1-4 Le Cun, Y. 11-2
Horizontal cells 18-4 Leaky integrator 18-9
Hubel, D.H. 16-2, 18-12 Learning rate 9-3, 10-8
competitive learning 16-9
I stable 9-6, 10-10
Illusions 18-4 Learning rules 4-2
Incremental training 11-17 ART1 19-21
Infinite impulse response 14-7 backpropagation 11-7
Inhibitory 18-10 competitive learning 16-7
Inner product 5-6 delta rule 7-13
Input selection 22-12 Grossberg competitive network 18-22
Input weight 14-3 Hebb rule 7-4
Instar 15-9 Hebbian learning 7-2
Instar rule 15-11, 16-7 learning vector quantization 16-16
Integrator 2-13 LMS algorithm 10-7
Interpolation 13-3 local learning 15-5
Interval location 12-16 perceptron 4-8, 4-13
Interval reduction 12-16 proof of convergence 4-15
Invariant set 20-13 performance learning 8-2
pseudoinverse rule 7-7
J reinforcement learning 4-3
Jacobian matrix 12-20 supervised learning 4-3
Jacobs, R.A. 12-13 unsupervised learning 4-3
Index-4
Index
Widrow-Hoff 7-13 Mean squared error 22-16

Learning vector quantization (LVQ) 16-16 Memory
subclass 16-17 associative 7-3
Levenberg-Marquardt algorithm 12-19, 12-21, autoassociative 7-10
22-14, 24-7 Mexican-hat function 16-11
Jacobian calculation 12-22 Minima 8-7
Jacobian matrix 12-20 first-order conditions 8-10
Likelihood function 13-13 global minimum 8-7
Linear associator 7-3 necessary conditions 8-9
Linear independence 5-4 second-order conditions 8-11
Linear initialization 26-6 strong minimum 8-7
Linear least squares 17-11 sufficient condition 8-11
Linear separability 4-19 weak minimum 8-7
Linear transformation 6-2 Minkowski error 22-17
change of basis 6-6 Minsky, M. 1-3, 4-2
domain 6-2 Missing data 22-8
matrix representation 6-3 MOBP 12-11
change of basis 6-6 Molecular dynamics 24-3
range 6-2 Momentum 12-9, 15-7
Linear vector spaces 5-2 Monte Carlo simulation 25-9
LMBP 12-25 Most probable 13-14
LMS algorithm 10-2, 10-7 Multilayer perceptron 11-2
adaptive filtering 10-13 Myocardial infarction 25-2
adaptive noise cancellation 10-15
analysis of convergence 10-9 N
learning rate 10-8 NARX network 22-10
stable learning rate 10-10 Negative definite matrix 8-11
Local learning 15-5 Negative semidefinite 8-11
Long term memory (LTM) 18-12, 18-22 Neighborhood 16-12
LVQ2 16-21 Network architectures 2-9
Lyapunov function 20-12 layer 2-9
Lyapunov stability theorem 20-6 multiple layers 2-10
Neural Network Toolbox for MATLAB 1-5
M Neuron model 2-2
Mach, E. 1-2 multiple input neuron 2-7
magnet 27-2 single input neuron 2-2
Marquardt algorithm 12-19 transfer functions 2-3
Marquardt sensitivity 12-24 Newtons method 9-10
Matrix representation 6-3 Nguyen-Widrow weight initialization 22-13, 24-9
change of basis 6-6 Nilsson, N. 16-2
diagonalization 6-13 Noise cancellation
McClelland, J.L. 1-4, 11-2 adaptive 10-15
McCulloch, W.S. 1-3, 4-2 echo cancellation 10-21
Mean square error 10-4, 11-8 Norm 5-7
Index-5
Index
Normalization 22-5, 25-6, 26-4 single-neuron 4-5

Novelty detection 22-28 test problem 4-9
training multiple-neuron perceptrons 4-13
O two-input case 3-4
Ockhams razor 13-2 unified learning rule 4-12
On-center/off-surround 16-11, 18-14 Performance Index 11-8
Optic disk 18-5 Performance index 8-2, 22-16
Optimality cross-entropy 22-17
first-order conditions 8-10 mean squared error 22-16
necessary conditions 8-9 Minkowski error 22-17
second-order conditions 8-11 quadratic function 8-12
sufficient condition 8-11 Performance learning 8-2
Optimization Pitts, W.H. 1-3, 4-2
conjugate gradient 9-15, 12-14 Positive definite 20-5
descent direction 9-3 Positive definite matrix 8-11
Gauss-Newton 12-21 Positive semidefinite 8-11, 20-5
Levenberg-Marquardt 12-19, 12-21 posterior density 13-13
Newtons method 9-10 Post-training analysis 22-18
quadratic termination 9-15 Prediction 22-10, 22-24
steepest descent 9-2 Preprocessing 22-5
stable learning rates 9-6 coding the targets 22-7
Oriented receptive field 18-20 feature extraction 22-6
Orienting subsystem 19-13 normalization 22-5, 25-6, 26-4
Orthogonal least squares 17-18 principal component analysis 22-6
Orthogonality 5-7 Principal component analysis 22-6
Orthonormal 5-9 Prior density 13-13
Outliers 22-19 Probability estimation 24-2
Outstar 15-16 Projection 5-8
Outstar rule 15-17 Prototype patterns 21-16
Overfitting 13-3, 22-27 Pseudoinverse rule 7-7
P Q
Papert, S. 1-3, 4-2 Quadratic function 8-12
Parker, D.B. 11-2 circular hollow 8-16
Pattern classification 11-3, 22-9, 25-2 elliptical hollow 8-17
Pavlov, I. 1-2 Hessian
Perceptron 3-3 eigensystem 8-13
architecture 4-3 saddle point 8-18
constructing learning rules 4-10 stationary valley 8-19
decision boundary 4-5 Quadratic termination 9-15
learning rule 4-8, 4-13 Quantization error 22-23
proof of convergence 4-15 Quickprop 12-14
multilayer 11-2
multiple-neuron 4-8
Index-6
Index
R Short term memory (STM) 18-12, 18-17

R value 22-20 Shunting model 18-10
Radial basis network 17-2 Similarity transform 6-8
backpropagation 17-25 Simulation order 14-4
center 17-6 Smart sensor 23-2
pattern classification 17-6 Softmax 22-7, 24-6
Range 6-2 Spanning a space 5-5
RBF 17-2 Spurious patterns 21-20
Real-time recurrent learning (RTRL) 14-2, 14-11, Stability
14-12 asymptotically stable 20-3, 20-5
Receiver operating characteristic (ROC) curve concepts 20-3
22-22, 25-8 equilibrium point 20-4
Reciprocal basis vectors 5-10 in the sense of Lyapunov 20-3, 20-4
Recurrent 14-2 LaSalles corollary 20-14
Recurrent network 2-13, 2-14, 20-2 LaSalles invariance theorem 20-13
regression 22-8 Lyapunov function 20-12
Regression/scatter plot 22-20 Lyapunov stability theorem 20-6
Regularization 13-8, 13-19, 13-21 pendulum example 20-6
Reinforcement learning 4-3 Stability-plasticity dilemma 19-2
Resonance 19-17 Stationary point 8-10
Retina 18-3 minima 8-7
Rods 18-3 saddle point 8-8
Rosenblatt, F. 1-3, 4-2, 10-2, 11-2, 16-2 Stationary valley 8-19
Rosenfeld, E. 1-2 Steepest descent 9-2
Rumelhart, D.E. 1-4, 11-2 learning rate 9-3
minimizing along a line 9-8
S stable learning rates 9-6
Saddle point 8-8, 8-18 Stimulus-response 15-2
Scaled conjugate gradient algorithm 22-14 conditioned stimulus 15-3
SDBP 12-2 unconditioned stimulus 15-3
Segmentation 22-9 Stopping criteria 22-15
Self-organizing feature map (SOFM) 16-12, 22- Subclass 16-17
16, 26-2 Subset selection 17-18
distortion measure 22-23 Subset/superset dilemma 19-17
neighborhood 16-12 SuperSAB 12-14
quantization error 22-23 Supervised learning 4-3
topographic error 22-23 Hebb rule 7-4
Sensitivity 11-10 performance learning 8-2
backpropagation 11-11 target 4-3
Sensitivity analysis 22-28 training set 4-3
Set
L 20-13 T
Z 20-12 Tapped delay line 10-13
Shakespeare, W. 1-5 Target 4-3
Index-7
Index
Taylor series expansion 8-2 Vision normalization 18-8

vector case 8-4 Visual cortex 18-4
Test set 13-6 VLBP 12-12
Thomas Bayes 13-10 von der Malsburg, C. 16-2, 18-12
Tikhonov 13-8
Time constant 18-9 W
Tollenaere, T. 12-14 Weight indices 2-7
Topographic error 22-23 Weight initialization 22-13, 24-9
Training process 22-2 Weight matrix 2-7
Training set 4-3 Werbos, P.J. 11-2
sequence 15-5 White noise 22-24
Transfer functions 2-3, 2-6 Widrow, B. 1-3, 10-2, 11-2
competitive 2-6 Widrow-Hoff algorithm 7-13, 10-7
global vs local 17-9 adaptive filtering 10-13
hard limit 2-3, 2-6 Wiesel, T. 16-2, 18-12
hyperbolic tangent sigmoid 2-6 Williams, R.J. 11-2
linear 2-4, 2-6 Winner-take-all 16-5
log-sigmoid 2-4, 2-6
positive linear 2-6
saturating linear 2-6
softmax 22-7, 24-6
symmetric saturating linear 2-6
symmetrical hard limit 2-6
table 2-6
Type I error 22-22
U
Unconditioned stimulus 15-3
Unsupervised learning 4-3
Hebb rule 7-4, 15-5
V
Validation set 13-7
Vector expansion 5-9
reciprocal basis vectors 5-10
Vector space 5-2
angle 5-7
basis set 5-5
orthonormal 5-9
projection 5-8
spanning 5-5
vector expansion 5-9
Vigilance 19-15
Vision 18-3
Index-8
This book provides a clear and detailed coverage of fundamental neural network
architectures and learning rules. In it, the authors emphasize a coherent presentation of the
principal neural networks, methods for training them and their applications to practical
problems.
Features
Extensive coverage of training methods for both feedforward networks (including
multilayer and radial basis networks) and recurrent networks. In addition to conjugate
gradient and Levenberg-Marquardt variations of the backpropagation algorithm, the
text also covers Bayesian regularization and early stopping, which ensure the
generalization ability of trained networks.
Associative and competitive networks, including feature maps and learning vector
quantization, are explained with simple building blocks.
A chapter of practical training tips for function approximation, pattern recognition,
clustering and prediction, along with five chapters presenting detailed real-world case
studies.
Detailed examples and numerous solved problems. Slides and comprehensive
demonstration software can be downloaded from hagan.okstate.edu/nnd.html.
About the Authors
Martin T. Hagan (Ph.D. Electrical Engineering, University of Kansas) has taught and
conducted research in the areas of control systems and signal processing for the last 35 years.
For the last 25 years his research has focused on the use of neural networks for control,
filtering and prediction. He is a Professor in the School of Electrical and Computer
Engineering at Oklahoma State University and a co-author of the Neural Network Toolbox
for MATLAB.
Howard B. Demuth (Ph.D. Electrical Engineering, Stanford University) has twenty-three

years of industrial experience, primarily at Los Alamos National Laboratory, where he helped
design and build one of the world's first electronic computers, the "MANIAC." Demuth has
fifteen years teaching experience as well. He is co-author of the Neural Network Toolbox for
MATLAB and currently teaches a Neural Network course for the University of Colorado at
Boulder.
Mark Hudson Beale (B.S. Computer Engineering, University of Idaho) is a software engineer
with a focus on artificial intelligence algorithms and software development technology. Mark
is co-author of the Neural Network Toolbox for MATLAB and provides related consulting
through his company, MHB Inc., located in Hayden, Idaho.
Orlando De Jess (Ph.D. Electrical Engineering, Oklahoma State University) has twenty-
four years of industrial experience, with AETI C.A. in Caracas, Venezuela, Halliburton in
Carrollton, Texas and is currently working as Engineering Consultant in Frisco, Texas.
Orlandos dissertation was a basis for the dynamic network training algorithms in the
Neural Network Toolbox for MATLAB.

NNDesign PDF

Uploaded by

Copyright:

Available Formats

NNDesign PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

NNDesign PDF

Uploaded by

Copyright:

Available Formats

Neural

Mark Hudson Beale

Neural Network Design, 2nd Edition, eBook

OVERHEADS and DEMONSTRATION PROGRAMS can be found at the following website:

2 Neuron Model and Network Architectures

4 Perceptron Learning Rule

6 Linear Transformations for Neural Networks

8 Performance Surfaces and Optimum Points

22 Practical Training Issues

24 Case Study 2:Probability Estimation

25 Case Study 3:Pattern Recognition

27 Case Study 5: Prediction

Signal and 5 Variations on 12 19

workstations, which rapidly grew in capability, became widely available. In

Oil and Gas

Figure 1.1 Schematic Drawing of Biological Neurons

[Gros80] S. Grossberg, How does the brain build a cognitive code?

erage. This study demonstrated that the posterior hippoc-

This work describes an experiment in which infants from

2 Neuron Model and Network

Theory and Examples

Inputs General Neuron

Figure 2.1 Single-Input Neuron

If, for instance, w = 3 , p = 2 and b = 1.5 , then

Figure 2.2 Hard Limit Transfer Function

as illustrated in Figure 2.3.

Figure 2.3 Linear Transfer Function

Figure 2.4 Log-Sigmoid Transfer Function

The log-sigmoid transfer function is commonly used in multilayer networks

Table 2.1 Transfer Functions

Inputs Multiple-Input Neuron

Figure 2.5 Multiple-Input Neuron

This expression can be written in matrix form:

Fortunately, neural networks can often be described with matrices. This

Input Multiple-Input Neuron

Figure 2.6 Neuron with R Inputs, Abbreviated Notation

Inputs Layer of S Neurons

Figure 2.7 Layer of S Neurons

Input Layer of S Neurons

Figure 2.8 Layer of S Neurons, Abbreviated Notation

Multiple Layers of Neurons

these layers. We will use superscripts to identify the layers. Specifically, we

Inputs First Layer Second Layer Third Layer

w 11,1 n11 a11 w 21,1 n21 a21 w 31,1 n31 a31

a1 = f 1 (W1p + b1) a2 = f 2 (W2a1 + b2) a3 = f 3 (W3a2 + b3)

Figure 2.9 Three-Layer Network

Input First Layer Second Layer Third Layer

a1 = f 1 (W1p + b1) a2 = f 2 (W2a1 + b2) a3 = f 3 (W3a2 + b3)

Figure 2.10 Three-Layer Network, Abbreviated Notation

Figure 2.11 Delay Block

Figure 2.12 Integrator Block

a(0) = p a(t + 1) = satlins (Wa(t) + b)

Figure 2.13 Recurrent Network

Input Multiple-Input Neuron

Three Layers of Neurons

a1 = f 1 (W1p + b1) a2 = f 2 (W2a1 + b2) a3 = f 3 (W3a2 + b3)

a(0) = p a(t + 1) = satlins (Wa(t) + b)

How to Pick an Architecture

a = hardlim 1.6 = 1.0