Knowledge Component 3: Acquiring Data and Knowledge

3.1.1 Introduction to
Machine Learning
2nd Edition

Ian F. C. Smith
EPFL, Switzerland

1
 Module Information
• Intended audience
– Novice
• Key words
– Machine learning
– Supervised learning
– Unsupervised learning
• Reviewer (1st Edition)
– Ian Flood, U of Florida,
Gainesville, USA

2
 What there is to learn

At the end of this module, there will be answers to the

following questions (see the quiz):

•What are the different ways in which computers can

learn?
•Are there learning tasks that humans can do much
better than computers?

3
 Outline
Machine Learning

Area of Influence

Successful Applications

Forms of Machine Learning

Types of Learning Algorithms

4
 Machine Learning
Humans learn from experience and adapt their actions
for future tasks.

Can machines adapt their behavior using

experience?

Since the 1950s, researchers have been trying to

develop techniques that enable machines to learn.
There have been much success in areas such as
automatic control, recognition systems and natural
language processing. Other successes are emerging.

5
 What is Machine Learning ?

An algorithm is said to learn from experience E with

respect some class of tasks T and performance
measure P …

… if its performance at tasks in T, as measured by

P, improves as it does task in T (experience E).

6
 Machine Learning - Examples

Example 1 Learning to recognize faces

– T: recognize faces
– P: % of correct recognitions
– E: opportunity to makes guesses and
being told what the truth is

Example 2 Learning to find clusters in data

– T: finding clusters
– P: compactness of groups detected
– E: analyses of a growing set of data

7
 Current Status

Existing machine learning techniques are applicable

only when the learning task is well-defined.

In many engineering applications, it is possible to

formalize the learning task of specific “sub-
problems”.

8
 Outline
Machine Learning

Area of Influence

Successful Applications

Forms of Machine Learning

Types of Learning Algorithms

9
 Areas of influence
Machine learning research is often
interdisciplinary. There are synergies in the
following fields:

 Statistics
 Brain models
 Adaptive Control Theory
 Psychology
 Artificial Intelligence
 Evolutionary models
 Information theory
 Philosophy

10
 Outline
Machine Learning

Area of Influence

Successful Applications

Forms of Machine Learning

Types of Learning Algorithms

11
 Successful Applications
 Learning to predict risk of failures for components
and systems of New York city power grid.
(Rudin et al. 2012 )

 Learning to analyze and predict the response of

wind turbine structures to varying wind field
characteristics. (Park et al. 2013)

 Learning to assess chlorine concentration in WDS.

(Cuesta Cordoba et al., 2014)

12
 Outline
Machine Learning

Area of Influence

Successful Applications

Forms of Machine Learning

Types of Learning Algorithms

13
 Forms of Machine Learning
 Supervised learning
A series of examples are used for feedback

 Unsupervised learning
No feedback

 Reinforcement learning
Indirect feedback after experience

14
 Supervised Learning
A learning task involves a set of input variables and a
set of output variables.

The set of possible relationships (hypotheses)

between input and output variables is known as the
hypothesis space.

Hypothesis have representations such as numerical

functions, symbolic rules, decision trees and artificial
neural nets.

Most learning is performed in a closed world where

the hypothesis space is predefined and finite.

15
 Supervised Learning (cont'd.)

The learning algorithm attempts to find the best

hypothesis that maps input to output using
“feedback”.

Feedback consists of a set of points (training data)

for which values of input and output variables are
known.

Since training data are used, this is supervised

learning.

16
 Unsupervised Learning

In unsupervised learning, output variables are not

known.

Unsupervised learning algorithms identify trends in

data and make inferences without knowledge of
correct answers.

17
 Reinforcement Learning

Reinforcement learning is concerned with how

software ought to initiate actions in an environment
so as to maximize some notion of long-term reward.

Reinforcement learning algorithms identify ways to

maps states of the world to the actions the software
ought to take in those states.

Reinforcement learning may involve learning from

mistakes.

18
 Outline

Machine Learning

Area of Influence

Successful Applications

Forms of Machine Learning

Types of Learning Algorithms

19
 Types of Learning Algorithms
There are four types of machine learning algorithms:

 Rote

 Statistical

 Deductive

 "Exploration and discovery“

These types are another way to classify machine

learning and are mostly independent of the forms of
machine learning defined earlier. The next module
provides more detail.
20
 Review Quiz I

 Give an example of a learning task that is easy for a

human being but hard for a computer.

 Name the different forms of machine learning.

 What is the difference between supervised and

unsupervised learning?
 Answers to Review Quiz I

 Give an example of a learning task that is easy for a

human being but hard for a computer.

An example is image recognition. A 5-year old child is

able to distinguish between a car and a tree in a
picture.

This task is difficult for a computer; it is hard to write a

program that distinguishes the two. A machine learning
algorithm could be used successfully to perform image
recognition
 Answers to Review Quiz - I
 Name the different forms of machine learning.
Supervised, unsupervised, reinforcement

 What is the difference between supervised and

unsupervised learning?
In case of supervised learning, there is a training set
which contains input and output for a number of
examples. The output is used as a feedback to learn from
the data.
In unsupervised learning, there is no training set. This kind
of learning algorithms make inference from trends in data.
 Further Reading
 Mitchell, T. Machine Learning. New York: McGraw-
Hill, 1997

 Kromanis et al.(2013). “Support vector regression

for anomaly detection from measurement
histories”.

 Dópido et al. (2013). ”Semisupervised self learning

for hyperspectral image classification”.

 Raphael, B. and Smith, I.F.C. “Engineering

informatics - fundamentals of computer-aided
engineering”, Wiley, 2013.

24
 Further Reading
 Cuesta Cordoba et al. (2014). “Using artificial neural
network models to assess water quality in water
distribution networks”.

 Park et al. (2013). “Multivariate analysis and

prediction of wind turbine response to varying wind
field characteristics based on machine learning”.

 Rudin et al.(1998). “Machine learning for the New

York city power grid”.

25