See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/200033859

Parallel distributed processing: explorations in the microstructure of

cognition. Volume 1. Foundations

Book · January 1986

CITATIONS READS

613 19,694

4 authors, including:

James L Mcclelland
Stanford University
374 PUBLICATIONS 91,964 CITATIONS

SEE PROFILE

All content following this page was uploaded by James L Mcclelland on 16 September 2014.

The user has requested enhancement of the downloaded file.

PARALLEL DISTRIBUTED
PROCESSING
Explorations in the Microstructure
of Cognition
Volume 1: Foundations

David E. Rumelhart James L. McClelland

and the PDP Research Group
Chisato Asanuma Alan H. Kawamoto Paul Smolensky
Francis H. C . Crick Paul W. Munro Gregory 0. Stone
Jeffrey L. Elman Donald A. Norman Ronald J . Williams
Geoffrey E. Hinton Daniel E. Rabin David Zipser
Michael I. Jordan Terrence J. Sejnowski

Institute for Cognitive Science

University of California, San Diego

A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
0 1986 by. The Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced

in any form by any electronic or mechanical means (including
photocopying, recording, or information storage and
retrieval) without permission in writing from the publisher.

Printed and bound in the United States of America

Library of Congress Cataloging-in-Publication Data

Rumelhart, David E.
Parallel distributed processing.

(Computational models of cognition and perception)

Vol. 2 by James L. McClelland, David E. Rumelhart,
and the PDP Research Group.
"A Bradford book."
Includes bibliographies and indexes.
Contents: v. 1. Foundations - v. 2. Psychological
and biological models.
1. Human information processing. 2. Cognition.
I. McClelland, James L. 11. University of California,
San Diego. PDP Research Group. 111. Title. IV. Series.
BF455.R853 1986 153 85-24073
ISBN 0-262-18120-7 (v. 1)
0-262- 13218-4 (v. 2)
0-262- 18123-1 (set)

.. ,
I . . . . . . . . . . . . .. . .. .. . .. ..
 CHAPTER 7

Learning and Relearning in Boltzmann Machines

G. E. HINTON and T . J. SEJNOWSKI

Many of the chapters in this volume make use of the ability of a paral-
lel network to perform cooperative searches for good solutions to prob-
lems. The basic idea is simple: The weights on the connections
between processing units encode knowledge about how things nornlally
fit together in some domain and the initial states or external inputs to a
subset of the units encode some fragments of a structure within the
domain. These fragments constitute a problem: What is the whole
structure from which they probably came? The network computes a
"good solution" to the problem by repeatedly updating the states of
units that represent possible other parts of the structure until the net-
work eventually settles into a stable state of activity that represents the
solution.
One field in which this style of computation seems particularly
appropriate is vision (Ballard, Hinton, & Sejnowski, 1983). A visual
system must be able to solve large constraint-satisfaction problems
rapidly in order to interpret a two-dimensional intensity image in terms
of the depths and orientations of the three-dimensional surfaces in the
world that gave rise to that image. In general, the information in the
image is not sufficient to specify the three-dimensional surfaces unless
the interpretive process makes use of additional plausible constraints
about the kinds of structures that typically appear. Neighboring pieces
of an image, for example, usually depict fragments of surface that have
similar depths, similar surface orientations, and the same reflectance.
The most plausible interpretation of an image is the one that satisfies
1
I
I
constrai~itsof this kind as well as possible, and the human visual sys-
ten1 stores enough plausible constraints and is good enough at applying
them that i r can arrive at the correct interpretation of most normal
images.
The computation may be perl'orrnetl by a n iterative search which
starts with a poor interprelation and progressively improves i t by reduc-
ing a cost function that measures the extent to which the current
interpretation violates the plausible constraints. Suppose, for example,
that each unit stands for a small three-dimensional surface fragment,
and the state of the unit indicates the current bet about whether that
surface fragment is part of the best three-dimensional interpretation.
Plausible constraints about the nature of surfaces can then be encoded
by the pairwise interactions between processing elements. For
example, iwo units that stand Sor neighboring surface fragments of
similar depth and surface orientation can be mutually excitatory to
encode the constraints that each of these hypotheses tends to support
the other (because objects tend to have continuous surfaces).

RELAXATION SEARCHES

The general idea of using parallel networks to perform relaxation

searches that simultaneously satisf'y multiple constr~~ints
is appealing. I t
might even provide a successor to telephone exchanges, holograms, or
con~n~unities of agents as a metaphor f'or the style o f computation in
cerebral cortex. But some tough technical questions have to be
answered before this style of computation can be accepted as either
efficient or plausible:

Will the network settle down or will it oscillate or wander aim-

lessly?

What does the network compute by settling down? We need

some characterization of the computation that the network per-
forms other than the network itself. Ideally we would like to
be able to say what ought to be computed (Marr, 1982) and
then to show that a network can be made to compute i t .

How long does the network take to settle on a solution? If

thousands of iterations are required the method becomes
implausible as a model of how the cortex solves constraint-
satisfaction problems.
 How much information does each unit need to convey to its
'neighbors? In many relaxation schemes the units communicate
accurate real values to one another on each iteration. Again
this is implausible if the units are intended to be like cortical
neurons which comn~unicateusing all-or-none spikes. To send
a real-value, accurate to within 5%, using firing rates requires
about 100 ms which is about the time allowed for the whole
iterative process to settle down.

How are the weights that encode the knowledge acquired? For
models of low-level vision it is possible for a programmer to
decide on the weights, and evolution might do the same for the
earliest stages of biological visual systems. But if the same kind
of constraint-satisfaction searches are to be used for higher
level functions like shape recognition or content-addressable
memory, there must be some learning procedure that autornati-
cally encodes properties of the domain into the weights.

This chapter is mainly concerned with the last of these questions, b u ~

the learning procedure we present is an unexpected consequence of our
attempt to answer the other questions, so we shall start with them.

Relaxation, Optimization, and Weak Constraints

One way of ensuring that a relaxation search is computing something

sensible (and will eventually settle down) is to show that it is solving an
optimization problem by progressively reducing the value of a cost
function. Each possible state of activity of the network has an associ-
ated cost, and the rule used for updating activity levels is chosen so
that this cost keeps falling. The cost function must be chosen so that
low-cost states represent good solutions to problems in the domain.
Many optimization problems can be cast in a framework known as
linear programming. There are some variables which take on real
values and there are linear equality and inequality constraints bet ween
variables. Each combination of values for the variables has an associ-
ated cost which is the sum over all the variables of the current value
times a cost-coefficient. The aim is to find a combination of values
that satisfies all the constraints and minimizes the cost function. If the
variables are further constrained to take on only the values 1 or 0 the
problem is called zero-one programming. Hinton (1977) has shown
that certain zero-one programming problems can be implemented as
relaxation searches in parallel networks. This allows networks to find
 7 L E A R N I N G IN I 3 0 L T Z M A N N MACf1INES 285

good solutions to problems in which there are discrete hypotheses that

are true or false. Even though the allowable solut~onsall assign values
of 1 or 0 to the hypotheses, the relaxation process works by passing
through intermediate states in which hypothesis units have real-valued
act~vitylevels lying between 1 and 0. Each constraint is enforced by a
feedback loop that measures the amount by which the current values
violate the constraint and tries to alter the values of the variables to
reduce this violation.
Linear programming and its variants make a sharp distinction
between constraints (which must be satisfied) and costs. A solution
which achieves a very low cost by violating one or two of the con-
straints is simply not allowed. In many domains, the distinction
between constraints and costs is not so clear-cut. In vision, for
example, i t is usually helpful to use the constraint that neighboring
pieces of surface are at similar depths because surfaces are mostly con-
t ~ n u o u sand are rarely parallel to the line of sight. But this is not an
absolute constraint. It doesn't apply at the edge of an object. So a
visual system needs to be able to generate interpretations that violate
this constraint if it can satisfy many other constraints by doing so.
Constraints like these have been called "weak" constraints (Blake, 1983)
and ~t is possible to formulate optimization problems In which all the
constraints are weak and there is no distinction between constraints and
costs. The optimal solution is then the one w h ~ c hm i n ~ m ~ z ethe
s total
constraint violation where different constraints are glven different
strengths depending on how reliable they are. Another way of saying
this is that all the constraints have associated plaus~bilities, and the
most plausible solution is the one which fits these plausible constraints
as well as possible.
Some relaxation schemes dispense with separate feedback loops for
the constraints and implement weak constraints directly in the excita-
tory and inhibitory interactions between units. We would like these
networks to settle into states in which a few units are fully active and
the rest are inactive. Such states constitute clean "digital" interpreta-
tions. T o prevent the network from hedging its bets by settling into a
state where many units are slightly active, it is usually necessary to use
a strongly nonlinear decision rule, and this also speeds convergence.
However, the strong nonlinearities that are needed to force the network
to make a decision also cause it to converge on different states on dif-
ferent occasions: Even with the same external inputs, the final state
depends on the initial state of the net. This has led many people (Hop-
field, 1982; Rosenfeld, Hummel, & Zucker, 1976) to assume that the
particular problem to be solved should be encoded by the initial state of
the network rather than by sustained external input to some of its
units.
 Hummel and Zucker (1983) and Hopfield (1982) have shown that
s o m e relaxation Schemes have an associated "potential" or cost function
and that the states to which the network converges are local minima of
this function. This means that the networks are performing optiniiza-
tion of a well-defined function. Unfortunately, there is no guarantee
that the network will find the best minimum. O n e possibility is to
redefine the problem as finding the local minimum which is closest to
the initial state. This is useful if the minima are used to represent
"items" in a memory, and the initial states are queries to memory
which may contain missing or erroneous information. T h e network
simply finds the minimum that best fits the query. This idea was used
by Hopfield (1982) who introduced an interesting kind of network in
which the units were always in o n e of two states.' Hopfield showed that
if the units are syn~metricallyconnected (i.e., the weight from unit i to
unit j exactly equals the weight from unit j to unit i) and if they are
updated o n e at a time, each update reduces (or at worst does not
increase) the value of a cost function which h e called "energy" because
of the analogy with physical systems. Consequently, repeated iterations
are guaranteed to find an energy minimum. T h e global energy of the
system is defined as

where w,,is the strength of connection (synaptic weight) from the ,jth
to the ith unit, s, is the state of the i t h unit (0 or l ) , and 0 , is a
threshold.
T h e updating rule is to switch each unit into whichever of its two
states yields the lower total energy given the current states of the other
units. Because the connections are symmetrical, the difference between
the energy of the whole system with the k t h hypothesis false and its
energy with the k th hypothesis true can b e determined locally by the
k t h unit, and is just

Therefore, the rule for minimizing the energy contributed by a unit is

to adopt the true state if its total input from the other units exceeds its
threshold. This is the familiar rule for binary threshold units.

I Hopfield used the states 1 and - 1 because his model was derived from physical sys-
tems called spin glasses in which spins are either "up" or "down." Provided the units
have thresholds, models that use 1 and - 1 can be translated into models that use 1 and 0
and have different thresholds.
 7 L E A R N I N G IN BOLTZMANN MACHINES 287

) U s i n g Probabilistic D e c i s i o n s t o E s c a p e F r o m Local M i n i m a

1 At about the same time that Hopfield showed how parallel networks
of this kind could be used to access memories that were stored as local
minima, Kirkpatrick, working at IBM, introduced an interesting new
i search technique for solving hard optimization problems on conven-
tional computers.
O n e standard technique is to use gradient descent: The values of the
variables in the problem are modified in whatever direction reduces the
cost function (energy). For hard problems, gradient descent gets stuck
at local minima that are not globally optimal. This is an inevitable
consequence of only allowing downhill moves. If jumps to higher
energy states occasionally occur, it is possible to break out of local
minima, but i t is not obvious how the system will then behave and it is
far from clear when uphill steps should be allowed.
Kirkpatrick, Gelatt, and Vecchi (1983) used another physical analogy
to guide the use of occasional uphill steps. T o find a very low energy
state of a metal, t h e best strategy is to melt it and then to slowly reduce
its temperature. This process is called annealing, and so they named
their search method "simulated annealing." Chapter 6 contains a dis-
cussion of why annealing works. W e give a simple intuitive account
here.
O n e way of seeing why thermal noise is helpful is to consider the
energy landscape shown in Fig~lre1. Let us suppose thar a ball-bearing
starts at a randomly chosen point on the landscape. I f i t always goes
downhill (and has n o inertia), i t will have an even chance of ending up
at A or B because both minima have the same width and so the initial

FIGURE 1. A simple energy landscape containing two local nimrna separated by a n

energy barrier. Shaking can be used to allow the state of the network (represented here
by a ball-bearing) to escape from local minima.
random point is equally likely to lie in either minimum. If we shake
the .whole system, we are more likely to shake the ball-bearing from A
to B than vice versa because the energy barrier is lower from the A
side. If the shaking is gentle, a transition from A to B will be many
times as probable as a transition from B to A, but both transitions will
be very rare. So although gentle shaking will ultimately lead to a very
high probability of being in B rather than A, it will take a very long
time before this happens. On the other hand, if the shaking is violent,
the ball-bearing will cross the barrier frequently and s o the ultimate
probability ratio will be approached rapidly, but this ratio will not be
very good: With violent shaking it is almost as easy to cross the barrier
in the wrong direction (from B to A) as in the right direction. A good
compromise is to start by shaking hard and gradually shake more and
more gently. This ensures that at some stage the noise level passes
through the best possible compromise between the absolute probability
of a transition and the ratio of the probabilities of good and bad transi-
tions. I t also means that at the end, the ball-bearing stays right at the
bottom of the chosen minimum.
This view of why annealing helps is not the whole story. Figure 1 is
misleading because all the states have been laid out in one dimension.
Complex systems have high-dimensional state spaces, and so the barrier
between two low-lying states is typically massively degenerate: The
number of ways of getting from one low-lying state to another is an
exponential function of the height of the barrier one is willing to cross.
This means that a rise in the level of thermal noise opens u p an enor-
mous variety of paths for escaping from a local minimum and even
though each path by itself is unlikely, i t is highly probable that the sys-
tem will cross the barrier. We conjecture that simulated annealing will
only work well in domains where the energy barriers are highly
degenerate.

Applying Simulated Annealing to Hopfield Nets

There is a simple modification of Hopfield's updating rule that allows

parallel networks to implement simulated annealing. If the energy gap
between the 1 and 0 states of the k t h unit is AEk then, regardless of
the previous state set, sk = 1 with probability
where T is a parameter which acts like the temperature of a physical
system. This local decision rule ensures that in thermal equilibrium the
relative probability of two global states is deterniined solely by their
energy difference, and follows a Boltzmann distribution:

where P, is the probability of being in the a t h global state, and E,, is

the energy of that state.
At low temperatures there is a strong bias in favor of states with low
energy, but t h e time required to reach equilibrium may be long. At
higher temperatures the bias is not s o favorable, but equilibrium is
reached faster. T h e fastest way to reach equilibrium at a given tem-
perature is generally to use simhlated annealing: Start with a higher
temperature and gradually reduce it.
T h e idea of implementing constraints as interactions between sto-
chastic processing elements was proposed by Moussouris (1 974) who
discussed the identity between Boltzniann d~srributionsand Markov
random fields. T h e idea of using simulated annealing to find low
energy states in parallel networks has been investigated independently
by several different groups. S. Genian and D. Geman (1984) esta-
blished limits on the allowable speed of the anneal~ngschedule and
showed that siniulated annealing can be very efl'ect~ve for removing
noise from images. Hinton and Sejnowski (1983b) showed how the use
of binary stochastic elements could solve some problems that plague
other relaxation techniques, in particular the problem of learning the
weights. Smolensky ( 1 983) has been investigating a similar scheme
which h e calls "harmony theory." This scheme is discussed in detail in
Chapter 6. Smolensky's harmony is equivalent to our energy (with a
sign reversal).

Pattern Completion

O n e way of using a parallel network is to treat it as a pattern comple-

tion device. A subset of the units are "clamped" into their on or off
states and the weights in the network then complete the pattern by
determining t h e states of the remaining units. There are strong limita-
tions o n the sets of binary vectors that can be learned if the network
has o n e unit for each component of the vector. These limits can be
transcended by using extra units whose states d o not correspond to
components in t h e vectors to b e learned. T h e weights of connections
t o these extra units can b e used to represent complex interactions that
cannot be expressed as pairwise correlations between the con~ponents
of the' vectors. We call these extra units hiddcn uuits (by analogy with
hidden Markov processes) and we call the units that are used to specify
the patterns to be learned the visible unils. T h e visible units are the
interface between the network and the environment that specifies vec-
tors for it to learn or asks i t to complete a partial vector. T h e hidden
units are where the network can build its own internal representations.
Sometimes, we would like to be able to complete a paltern from any
sufficiently large part of i t without knowing in advance which part will
be given and which part must be completed. Other times we know in
advance which parts will be given as input and which parts will have to
be completed as output. So there are two different completion para-
digms. In the first, any of the visible units might be part of the
required output. In the second, there is a distinguished subset of the
visible units, called the input units, which are always clamped by the
environment, s o the network never needs to determine the states of
these units.

EASY A N D HARD LEARNING

Consider a network which is allowed to run freely, using the proba-

bilistic decision rule in Equation 3, without having any of its units
clamped by the environment. When the network reaches thermal
equilibrium, the probability of finding i t in any particular global state
depends only on the energy of that state (Equation 4 ) . We can there-
fore control the probabilities of global states by controlling their ener-
gies. If each weight only contributed to the energy of a single global
state, this would be straightforward, but changing a weight will actually
change the energies of many different states s o i t is not immediately
obvious how a weight-change will affect the probability of a particular
global state. Fortunately, if we run the network, until it reaches thermal
equilibrium, Equations 3 and 4 allow us to derive the way in which the
probability of each global state changes as a weight is changed:

where SF is the binary state of the i t h unit in the a t h global state and
P i is the probability, at thermal equilibrium, of global state a of the
network when none of the visible units are clamped (the lack of clamp-
ing is denoted by the superscript - ) . Equation 5 shows that the effect
 7

of a weight on the log probabil ity of a global state can be computed

from purely local information because it only involves the behavior of
the two units that the weight connects (the second term is just the
probability of finding the i t h and j t h units on together). This makes i t
easy to manipulate the probabilities of global states provided the desired
probabilities are known (see Hinton & Sejnowski, 1983a, for details).
Unfortunately, it is normally unreasonable to expect the environment
o r a teacher to specify the required probabilities of entire global states
of the network. T h e task that the network must perform is defined in
terms of the states of the visible units, and s o the environment or
teacher only has direct access to the states of these units. The difficult
learning problem is to decide how to use the hidden units to help
achieve the required behavior of the visible units. A learning rule
which assumes that the network is instructed from outside on how to
use all of its units is of limited interest because it evades the main
problem which is to discover appropriate representations for a given
task among the hidden units.
In statistical terms, there are many kinds of statistical structure irnpli-
cit in a large ensemble of environmental vectors. The separate proba-
bility of each visible unit being active is the first-order structure and
can be captured by the thresholds of the visible units. The v2/ 2 pair-
wise correlations between the 11 visible units constitute the second-
order structure and this can be captured by the weights between p t l r j of'
units.* All structure higher than second-order cannot be captured by
pairwise weights be tweet^ tlic vis~hlr~it1ir.s. A simple example may help to
clarify this crucial point.
Suppose that the ensemble consists of the vectors: ( 1 1 01, ( 1 0 l ) ,
(0 1 l ) , and (0 0 01, each with a probability of 0.25. There I S clearly
s o m e structure here because four of the eight possible 3-bit vectors
never occur. However, the structure is entirely third-order. The f ~ r s t -
order probabilities are all 0.5, and the second-order correlations are all
0, s o if we consider only these statistics, this ensemble is indistinguish-
able from the ensemble in which all eight vectors occur equiprobably.
T h e Widrow-Hoff rule or perceptron convergence procedure (Rosen-
blatt, 1962) is a learning rule which is designed to capture second-order
structure and it therefore fails miserably o n t h e example just given. If
t h e first two bits are treated as an input and the last bit is treated as the
required output, the ensemble corresponds to the function "exclusive-
or" which is o n e of the examples used by Minsky and Papert (1969) to
show t h e strong limitations of one-layer perceptrons. T h e Widrow-Hoff

2 Factor analysis confines itself to capturing as much of the second-order structure as

possible in a few underlying "factors." It ignores all higher order structure which is where
much of the interesting information lies for all but the most simple ensembles of vectors.
rule can d o easy learning, but i t cannot d o the kind of hard learning
that involves deciding how to use extra units whose behavior is not
directly specified by the task.
It is tempting to think that networks with pairwise connections can
never capture higher than second-order statistics. There is o n e sense in
which this is true and another in which it is false. By introducing extra
units which are not part of the definition of the original ensemble, it is
possible to express the third-order structure of the original ensemble in
the second-order structure of the larger set of units. In the example
given above, we can add a fourth component to get the ensemble
( ( 1 101), ( l o l o ) , (01 l o ) , (0000)). It is now possible to use the thresh-
olds and weights between all four units to express the third-order struc-
ture in the Sirst three components. A more familiar way of saying this
is that we introduce an extra "feature detector" which in this example
detects the case when the first two units are both on. W e can then
make each of the first two units excite the third unit, and use strong
inhibition from the feature detector to overrule this excitation when
botli of the first two units are on. T h e difficult problem in introducing
[he extra unit was deciding when i t should be on and when i t should be
off-deciding what feature it should detect.3
One way of thinking about the higher order structure of an ensemble
of environnlental vectors is that i t implicitly specifies good sets of
underlying features that can be used to model the structure of the
environment. In common-sense terms, the weights in the network
should be chosen so that the hidden units represent significant underly-
ing features that bear strong, regular relationships to each other and to
the states of the visible units. T h e hard learning problem is to figure
out what these features are, i.e., to find a set of weights which turn the
hidden units into useful feature detectors that explicitly represent
properties of the environment which are only implicitly present as
higher order statistics in the ensemble of environmental vectors.

Maximum Likelihood Models

Another view of learning is that the weights in the network consti-

tute a generative model of the environment-we would like to find a
set of weights s o that when the network is running freely, the patterns
of activity that occur over the visible units are the same as they would
be if the environment was clamping them. T h e number of units in the

3 I n this example there are six different ways o f using rhe extra unit to solve the task
 network and their interconnectivity define a space of possible models of'
the environment, and any particular set of weights defines a particular
model within this space. T h e learning problem is to I'ind a combination
of weights that gives a good model given the limitations imposed by the
architecture of the network and the way i t runs.
More forn~ally,we would like a way of finding the combination ol'
weights that is most likely to have produced the observed ensemble of
environmental vectors. This is called a ~ ? ~ a s i t ?l ~~~uctl ~
i h o omodel
ti and
there is a large literature within statistics on maximum likelihood esti-
mation. T h e learning procedure we describe actually has a close rela-
tionship to a r n e ~ h o d called Expectation and Maximiza~ion ( E M )
(Dempster, Laird, & Rubin, 1976). EM is used by statisticians for
estimating missing pararneters. I t represents probability distributions by
using pararneters like our weights that are exponentially related to
probabilities, rather than using probabilities themselves. The EM algo-
rithm is closely related to an earlier algorithm invented by Baum that
manipulates probabilities directly. Baum's algorithm has been used suc-
cessfully for speech recognition (Bahl, Jelinek, & Mercer, 1983). I t
estimates the parameters of a hidden Markov chain-a transition net-
work which has a fixed s t r u c t ~ ~ rbut
e variable probabilities on the arcs
and variable probabilities of eniirting a particular output symbol ;is i t
arrives at each internal node. Given an ensemble of strings of' synibols
and a fixed-topology transition network, the algor-ith m finds the combi-
nation of transition probabilities and output probabilities that is most
likely to have produced these strings (actually i t only finds a local max-
imum).
Maxirnuni likelihood methods work by adjusting the p a r m e t e r s to
increase the probability that the generative model will produce the
observed data. Baum's algorithm and EM are able to estimate new
values for the probabilities (or weights) that are guaranteed to be better
than the previous values. Our algorithm simply estimates the gradient
of the log likelihood with respect to a weight, and so the magnitude of
the weight change must be decided using additional criteria. Our algo-
rithm, however, has the advantage that i t is easy to implement in a
parallel network of neuron-like units.
T h e idea of a stochastic generative model is attractive because it pro-
vides a clean quantitative way of comparing alternative representational
schemes. T h e problem of saying which of two representational schemes
is best appears to be intractable. Many sensible rules of thumb are
available, but these are generally pulled out of thin air and justified by
commonsense and practical experience. They lack a firm mathematical
foundation. If we confine ourselves to a space of allowable stochastic
models, we can then get a simple Bayesian measure of the quality of a
representational scheme: How likely is the observed ensemble of
environmental vectors given the representational scheme? In our net-
works, representations are patterns of activity in the units, and the
representational scheme therefore corresponds to the set of weights that
determines when those patterns are active.

T H E BOLTZMANN MACHINE LEARNING ALGORITHM

If we make certain assumptions it is possible to derive a measure of

how effectively the weights in the network are being used for modeling
the structure of the environment, and it is also possible to show how
the weights should be changed to progressively improve this measure.
W e assume that the environment clamps a particular vector over the
visible units and it keeps it there long enough for the network to reach
thermal equilibrium with this vector as a boundary condition (i.e., to
"interpret" i t ) . W e also assume (unrealistically) that the there is no
structure in the sequential order of the environmentally clamped vec-
tors. This means that the complete structure of the ensemble of
environmental vectors can be specified by giving the probability,
P+(<,), of each of the 2' vectors over the v visible units. Notice that
the P + ( V , ) d o not depend on the weights in the network because the
environment clamps the visible units.
A particular set of weights can be said to constitute a perfect model
of the structure of the environment if it leads to exactly the same
probability distribution of visible vectors when the network is running
freely with no units being clamped by the environment. Because of the sto-
chastic behavior of the units, the network will wander through a variety
of states even with n o environmental input and it will therefore gen-
erate a probability distribution, P - ( V,), over all 2' visible vectors.
This distribution can be compared with the environmental distribution,
P + ( V , ) . In general, it will not b e possible t o exactly match the 2"
environmental probabilities using the weights among t h e v visible and
h hidden units because there are at most (v+ h- 1 ) (v+ h ) / 2
'

symmetrical weights and (v+ h ) thresholds. However, it may be possi-

ble t o d o very well if the environment contains regularities that can be
expressed in the weights. An information theoretic measure (Kullback,
1959) of the distance between the environmental and free-running
probability distributions is given by:

where P + ( V,) is t h e probability of t h e a t h state of the visible units in

 7 LI.iAII NINCi I N BOLTZMANN MACIIINI~S

phasef when their states are determined by the environrnen t , and

P - ( V,) is t h e corresponding probabil~tyIn phase- when the network I \
running freely with n o environmental input.
G is never negative and is only zero if the distribut~onsare ident~cal
G is actually t h e distance in bits j i o m the free running d i s t r ~ b u t ~ otno
the environmental d i ~ t r i b u t i o n .It~ is sometimes called the asymnieti-lc
divergence or information gain. T h e measure is not symmetric with
respect t o the two distributions. This seems odd but is actually very
reasonable. W h e n trying to approximate a probabiltty d i s t r ~ b u t ~ o ni t , I S
more important to get the probabilities correct for events that happen
frequently than for rare events. So the match between the actual and
predicted probabilities of an event should be weighted by the actual
probability as in Equation 6.
It is possible to improve the network's model of the structure of its
environment by changing the weights s o as to reduce G To perform '
gradient descent in G , we need to know how G wrll change when a
weight is changed. But changing a single weight changes the energles
of o n e quarter of all the global states of the network, and i t change5 the
probabilities of all the states In ways that depend on 011 the other
weights in t h e network. Consider, for example, the very s ~ m p l enet-
work shown in Figure 2. If we want the two units at the end5 of the
chain to be either both on or both off, how should we change thc
weight w 3 4 ? I t clearly depends on the slgns of remote ue1gh15l ~ k e
because we need to have an even number of i n h ~ b ~ t o r\reightb ! In rhc
chain. S o the partial derivat~ve of G w ~ t h re5pecL to onc wo~ght
depends on all the other weights and minimizing G nppears I O be 'i

input output
unit unit

FIGURE 2 A very s ~ r n p l enetwork w ~ t ho n e Input unlt, o n e output unlt dnd Iiro h ~ d d e n

unlts T h e task is t o m a k e the output unlt adopt t h e s a m e state 'is the Input unlt T h e
d~fficultyIS that t h e correct value for weight w 3 depends o n remote ~nfornintionItke the
value of w e ~ g h tw

4 If we use base 2 logarithms.

5 Peter Brown (personal communication) has pointed out that minimizing G is

equivalent t o maximizing t h e log of t h e likelihood of generating the environmental
probability distribution when the network is running freely at equilibrium.

6 T h e thresholds must also be adjusted appropriately

difficult computational problem that requires nonlocal information.
Fortunately, all the information that is required about the other
weights in order to change w,, appropriately shows up in the behavior
of the i t h and j t h units at thermal equilibrium. In addition to perform-
ing a search for low energy states of the network, the process of reach-
ing thermal equilibrium ensures that the joint activity of any two units
contains all the information required for changing the weight between
them in order to give the network a better model of its environment.
T h e joint activity implicitly encodes information about all the other
weights in the network. T h e Appendix shows that

where p,: is the probability, averaged over all environmental inputs and
measured at equilibrium, that the i t h and j t h units are both on when
the network is being driven by the environment, and p,; is the
corresponding probability when the network is free running. O n e
surprising feature of Equation 7 is that it does not matter whether the
weight is between two visible units, two hidden units, or one of each.
T h e same rule applies for the gradient of G .

Unlearning

Crick and Mitchison (1983) have suggested that a form of reverse

learning might occur during REM sleep in mammals. Their proposal
was based on the assumption that parasitic modes develop in large net-
works that hinder the distributed storage and retrieval of information.
T h e mechanism that Crick and Mitchison propose is based on

More o r less random stimulation of the forebrain by the brain

stem that will tend to stimulate the inappropriate modes of
brain activity . . . and especially those which are too prone to be
set off by random noise rather than by highly structured
specific signals. (p. 1 12)

During this state of random excitation and free running they postulate
that changes occur at synapses to decrease the probability of the
spurious states.
A simulation of reverse learning was performed by Hopfield, Fein-
stein, and Palmer (1983) who independently had been studying ways to
improve the associative storage capacity of simple networks of binary
processors (Hopfield, 1982). In their algorithm an input is presented to
the network as an initial condition, and the system evolves by falling
into a nearby local energy minimum. However, not all local energy
minima represent stored information. In creating the desired minima,
they accidentally create other spurious minima, and ro eliminate these
they use "unlearning": T h e learning procedure is applied with reverse
sign to the states found after starting from random initial conditions.
Following this procedure, the performance of the system in accessing
stored states was found to be improved.
There is an interesting relationship between the reverse learning pro-
posed by Crick and Mitchison and Hopfield et al. and the form of rhe
learning algorithm which we derived by considering how to minimize
an information theory measure of the discrepancy between the environ-
mental structure and the network's internal model (1-Iinton &
Sejnowski, 1983b). T h e two phases of our learning algorithm resemble
the learning and unlearning procedures: Positive Hebbian learning
occurs in phase.+ during which information in the environ~iientis cap-
tured by the weights; during phase- the system randomly samples states
according to their Boltzmann distribution and Hebbian learning occurs
with a negative coefficient.
However, these two phases need not be implemented in the manner
suggested by Crick and Mitchison. For example. during pI1a.s.c the
average co-occurrences could be computed without rnaking any changes
to the weights. These averages could then be used as a baseline for
making changes during phasef; that is, the co-occurrences during
p/ia~o' could be computed and the baseline subtracred before each per-
manent weight change. Thus, an alternative but equivalenr proposal for-
the function of dream sleep is to recalibrate the baseline for-
plasticity-the break-even point which determines whether a synaptic
weight is incremented or decremented. This would be safer than mak-
ing permanent weight decrements to synaptic weights during sleep and
solves the problem of deciding how much "unlearning" to do.
Our learning algorithm refines Crick and Mitchison's interpretation
of why two phases are needed. Consider a hidden unit deep within the
network: How should its connections with other units be changed to
best capture regularity present in the environment? If i t 'does not
receive direct input from the environment, the hidden unit has n o way
to determine whether t h e information it receives from neighboring
units is ultimately caused by structure in the environment or is entirely
a result of the other weights. This can lead to a "folie a deux" where
two parts of t h e network each construct a model of the other and
ignore the external environment. T h e contribution of internal and
external sources can b e separated by comparing the co-occurrences in
phasef with similar information that is collected in the absence of
environmental input. phase- thus acts as a control condition. Because
of the special properties of equilibrium it is possible to subtract off this
purely internal contribution and use the difference to update the
weights. 'Thus, the role of the two phases is to make the system maxi- --
mally responsive to regularities present in the environment and to
prevent the system from using its capacity to model internally-
generated regularities.

Ways in W h i c h t h e Learning Algorithm C a n Fail

T h e ability to discover the partial derivative of 6 by observing /I,;

and p; does not completely determine the learning algorithm. It is still
necessary to decide how much to change each weight, how long to col-
lect co-occurrence statistics before changing the weight, how many
weights to change at a time, and what temperature schedule to use dur-
ing the annealing searches. For very simple networks in very simple
environments, it is possible to discover reasonable values for these
parameters by trial and error. For more complex and interesting cases,
serious difficulties arise because it is very easy to violate the assump-
tions on which the mathematical results are based (Derthick, 1984).
T h e first difficulty is that there is nothing to prevent the learning
algorithm from generating very large weights which create such high
energy barriers that the network cannot reach equilibrium in the allot-
ted time. Once this happens, the statistics that are collected will not be
the equilibrium statistics required for Equation 7 to hold and so all bets
are off. We have observed this happening for a number of different
networks. They start off learning quite well and then the weights
become too large and the network "goes souru-its performance
deteriorates dramatically.
One way to ensure that the network gets close to equilibrium is to
keep the weights small. Pearlmutter (personal communication) has
shown that the learning works much better if, in addition to the weight
changes caused by the learning, every weight continually decays towards
a value of zero, with the speed of the decay being proportional to the
absolute magnitude of the weight. This keeps the weights small and
eventually leads to a relatively stable situation in which the decay rate
of a weight is balanced by the partial derivative of G with respect to the
weight. This has the satisfactory property that the absolute magnitude
of a weight shows how important it is for modeling the environmental
structure.
T h e use of weight-decay has several other consequences which are
not s o desirable. Because the weights stay small, the network cannot
construct very deep minima in the energy landscape and s o it cannot
make the probability ratios for similar global states be very different.
This means that it is bound to give a significant number of errors in
modeling environments where very similar vectors have very different
probabilities. Better pe~forrnance can be achieved by annealing the net-
work t o a lower final temperature (which is equivalent to making all the
weights larger), but this will make the learning worse for two separate
reasons. First, with less errors there is less to drive the learning
because it relies o n the difference between the phase' and phase-
statistics. Second, i t will be harder to reach thermal equilibrium at this
lower temperature and s o the co-occurrence statistics will be unreliable.
O n e way of getting good statistics to drive the learning and also getting
very few overt errors is to measure the co-occurrence statistics at a
temperature higher than the final one.
Another way of ensuring that the network approaches equilibrium is
to eliminate deep, narrow minima that are often not found by the
annealing process. Derthick (1984) has shown that this can be done
using a longer gentler annealing schedule in phase-. This means that
the network is more likely to occupy the hard-to-find minima in phase-
than in phase', and s o these minima will get filled in because the learn-
ing rule raises the energies of states that are occupied more in phase-
than in phase+.

A N EXAMPLE OF H A R D LEARNING

A simple example which can only be solved by capturing the higher

order statistical structure in the ensemble of input vectors is the
"shifter" problem. T h e visible units are divided into three groups.
G r o u p V 1 is a one-dimensional array of 8 units, each of which is
clamped o n or off at random with a probability of 0.3 of being o n .
G r o u p V, also contains 8 units and their states are determined by shift-
ing and copying t h e states of the units in group V,. T h e only shifts
allowed are o n e to the left, o n e to the right, or n o shift. Wrap-around
is used s o that when there is a right shift, the state of the right-most
unit in V , determines the state of the left-most unit in V2. T h e three
possible shifts are chosen at random with equal probabilities. Group V 3
contains three units to represent the three possible shifts, s o at any o n e
time o n e of t h e m is clamped on and the others are clamped off.
T h e problem is t o learn the structure that relates the states of the
three groups. O n e facet of this problem is to "recognize" the shift-
i.e., to complete a partial input vector in which the states of V, and V2
are clamped but t h e units in V 3 are left free. It is fairly easy to see why
this problem cannot possibly be solved by just adding together a lot of
pairwise interactions between units in V,, V,, and V 3 . If you know
that a particular unit in V1 is on, it tells you nothing whatsoever about
what t h e s h i f t is. It is only by finding combinations of active units in V1
and V2 that it is possible to predict the shift, s o the information
required is of at least third-order. This means that extra hidden units
are required to perform the task.
T h e obvious way to recognize the shift is to have extra units which
detect informative features such as an active unit in Vl and an active
unit o n e place to the right in V2 and then support the unit V3 that
represents a right shift. T h e empirical question is whether the learning
algorithm is capable of turning s o m e hidden units into feature detectors
of this kind, and whether it will generate a set of detectors that work
well together rather than duplicating the same detector. T h e set of
weights that minimizes G defines t h e optimal set of detectors but it is
not at all obvious what these detectors are, nor is i t obvious that the
learning algorithm is capable of finding a good set.
Figure 3 shows the result of running a version of the Boltzniann
machine learning procedure. Of t h e 24 hidden units, 5 seem to be
doing very little but the remainder are sensible looking detectors and
most of them have become spatially localized. One type of detector
which occurs several times consists of two large negative weights, one
above the other, flanked by smaller excitatory weights on each side.
This is a more discriminating detector of no-shift than simply having
two positive weights, o n e above the other. It interesting to note that
the various instances of this feature type all have different locations in
V, and V2, even though the hidden units are not connected to each
other. T h e pressure for the feature detectors to be different from each
other comes from the gradient of G , rather than from the kind of
lateral inhibition among the feature detectors that is used in "competi-
tive learning" paradigms (Fukushima, 1980; Rumelhart & Zipser,
1985).

The Training Procedure

T h e training procedure alternated between two phases. In phase', all

the units in V,, V2, and V3 were clamped into states representing a pair
of 8-bit vectors and their relative shift. T h e hidden units were then
allowed to change their states until the system approached thermal
equilibrium at a temperature of 10. T h e annealing schedule is
described below. After annealing, t h e network was assumed to be close
to thermal equilibrium and it was then run for a further 10 iterations
during which time the frequency with which each pair of connected
units were both o n was measured. This was repeated 20 times with
 F I G U R E 3. 'The w e i g h ~ol'thc
~ 24 hidden units In the jh~l'tcrni.l\r.c~rl\ 1:ach large region
corresponds to a unit. Within this region the black rectangle, ri.prc\ent ncgatiw weight\
iind the white rectangles represent positive ones. The .;i/e 01' ;I sc~,tanglerepresents the
magnitude of the weight. The two rows of weights at the t>ottonl of' e;~cliunit are its con-
nections to the two groups 01' input units, I", a n d C'2. T-he\c wc~ght\t h e r c l ' o ~represe~it
~~
the "receptive lield" of the hidden unit. The three we~ghlsIn the middle 01' the top row
01' each unit are its connections to the three output units that ~reprcsent shift-left. n o -
shil't, k~nd shift-right. The solitary weight at the top Icl't 01' ~ ~ ucn i ~hij its threshold.
Each hidden unit is directly connected to all 16 input units and all 3 output units. In this
example, the hidden units are not connected to each other. The top-left unit has weights
that are easy to understand: Its optimal stimulus is activity in tho I'ourth unit of I/, and
the fifih unit of V 2 , and it votes for shift-right. It has negative u ~ l g h t sto rnake i t less
likely to come on when there is an alternative explanation for wh! its two favorite input
units are active.

different clamped vectors and the co-occurrence stat~sticswere averaged

over all 20 runs to yield an estimate, for each connection, of p,f in
Equation 7. In phase-, none of the units were clamped and the net-
work was annealed in the same way. T h e network was then run for a
further 10 iterations and the co-occurrence statistics were collected for
all connected pairs of units. This was repeated 20 times and the co-
occurrence statistics were averaged to yield an estimate of p,;.
302 BASIC MECfIANlSMS

T h e entire set of 40 annealings that were used to estimate p,; and p,;
was called a sweep. After each sweep, every weight was incremented
by 5(p,: - p,;). In addition, every weight had its absolute magnitude
decreased by 0.0005 times its absolute magnitude. This weight decay
prevented the weights from becoming too large and it also helped to
resuscitate hidden units which had predominantly negative or predom-
inantly positive weights. Such units spend all their time in the same
state and therefore convey n o information. T h e phase' and phase-
statistics are identical for these units, and s o the weight decay gradually
erodes their weights until they come back to life (units with all zero
weights come on half the time).

The Annealing Schedule

T h e annealing schedule spent the following number of iterations at

the following temperatures: 2 at 40, 2 at 35, 2 at 30, 2 at 25, 2 at 20, 2
at 15, 2 at 12, 2 at 10. One iteration is defined as the number of ran-
d o m probes required s o that each unit is probed one time on average.
When it is probed, a unit uses its energy gap to decide which of its two
states to adopt using the stochastic decision rule in Equation 3. Since
each unit gets to see the most recent states of all the other units. an
iteration cannot be regarded as a single parallel step. An truly parallel
asynchronous system must tolerate time delays. Units must decide on
their new states without being aware of very recent changes in the
states of other units. It can be shown (Sejnowski, Hinton, Kienker, &
Schumacher, 1985) that first-order time delays act like added tempera-
ture and can therefore be tolerated by networks of this kind.

The Performance of the Shifter Network

T h e shifter network is encouraging because it is a clear example of

the kind of learning of higher order structure that was beyond the capa-
bility of perceptrons, but it also illustrates several weaknesses in the
current approach.

T h e learning was very slow. 1t required 9000 learning sweeps,

each of which involved reaching equilibrium 20 times in phasef
with vectors clamped o n V1, V Z , and V3, and 20 times in
phase- with n o units clamped. Even for low-level perceptual
learning, this seems excessively slow.
 T h e weights are fairly clearly not optimal because of the 5 hid-
d e n units that app'ear to d o nothing useful. Also, the
performance is far from perfect. When the states of the units
in V , and V 2 are clamped and the network is annealed gently to
half the final temperature used during learning, the units in V 3
quite frequently adopt the wrong states. If the number of on
units in V , is 1,2,3,4,5,6,7, the percentage of correctly recog-
nized shifts is 50°/o, 719'0, 81°h, 86%, 89%, 82%, anti 66%
respectively. T h e wide variation in the number of active units
in V , naturally makes the task harder to learn than if a constant
proportion of the units were active. Also, some of the input
patterns are ambiguous. When all the units in V , and V 2 are
off, the network can d o n o better than chance.

ACHIEVING RELIABLE COMPUTATION WITH

UNRELIABLE HARDWARE

Conventional computers only work if all their individual components

work perfectly, s o as systems become larger they become more and
more unreliable. Current computer technology uses extremely reliable
components and error-correcting memories to achieve overall reliability.
T h e brain appears to have much less reliable components, and so i t
must use much more error-correction. It is conceivable that the brain
uses the kinds of representations that would be appropriate given reli-
able hardware and then superimposes redundancy to compensate for its
unreliable hardware.
T h e reliability issue is typically treated as a tedious residual problem
to be dealt with after the main decisions about the form of the compu-
tation have been made. A more direct approach is to treat reliability as
a serious design constraint from the outset and to choose a basic style
of computation that does not require reliable components. Ideally, we
want a system in which none of the individual components are critical
to the ability of the whole system to meet its requirements. In other
words, we want s o m e high-level description of the behavior of the sys-
tem to remain valid even when the low-level descriptions of the
behavior of s o m e of the individual components change. This is only
possible if the high-level description is related to the low level descrip-
tions in a particular way: Every robust high-level property must be
implemented by the combined effect of many local components, and n o
single component must be crucial for the realization of the high-level
property. This makes distributed representations (see Chapter 3) a
natural choice when designing a damage-resistant system.
 Distributed representations tend to behave robustly because they
have an internal coherence which leads to an automatic "clean-up"
effect. This effect can be seen in the patterns of activity that occur
within a group of units and also in the interactions between groups. If
a group of units, A, has a number of distinct and well-defined energy
minima then these minima will remain even if a few units are removed
or a little noise is added to many of the connections within A. T h e
damage may distort the minima slightly and it may also change their
relative probabilities, but minor damage will not alter the gross topogra-
phy of the energy landscape, s o it will not affect higher level descrip-
tions that depend only on this gross topography.
Even if the patterns of activity in A are slightly changed, this will
often have no effect on the patterns caused in other groups of units. If
the weights between groups of units have been fixed s o that a particular
pattern in A regularly causes a particular pattern in B, a small variation
in the input coming from A will typically make no difference to the pat-
tern that gets selected in B, because this pattern has its own internal
coherence, and if the input from A is sufficiently accurate to select
approximately the right pattern, the interactions among the elements in
B will ensure that the details are right.
Damage resistance can be achieved by using a simple kind of
representation in which there are many identical copies of each type of
unit and each macroscopic item is encoded by activity in all the units of
o n e type, In the undamaged system all these copies behave identically
and a lot of capacity is therefore wasted. If we use distributed
representations in which each unit may be used for representing many
different items we can achieve comparable resistance to damage without
wasting capacity. Because all the units behave differently from each
other, t h e undamaged system can implement many fine distinctions in
the fine detail of the energy landscape. At the macroscopic level, these
fine distinctions will appear as somewhat unreliable probabilistic ten-
dencies and will be very sensitive to minor damage.
T h e fine details in the current energy landscape may contain the
seeds of future changes in the gross topography. If learning novel dis-
tinctions involves the progressive strengthening of regularities that are
initially tentative and unreliable, then it follows that learning may well
suffer considerably when physical damage washes out these minor
regularities. However, the simulations described below d o not bear on
this interesting issue.

AN EXAMPLE OF THE EFFECTS OF DAMAGE

T o show t h e effects of damage o n a network, it is necessary to
choose a task for the network to perform. Since we are mainly
concerned with properties that are fairly domain-independent, the
details of the task are n@ especially relevant here. For reasons
described in Chapter 3, we were interested in networks that can learn
a n arbitrary mapping between items in two different domains, and we
use that network to investigate the effects of damage. As we shall see,
t h e fact that the task involves purely arbitrary associations makes i t
easier to interpret s o m e of the interesting transfer effects that occur
when a network relearns after sustaining major damage.

T h e Network

T h e network consisted of three groups or layers of units. T h e gra-

pheme group was used to represent the letters in a three-letter word. It
contained 30 units and was subdivided into three groups of 10 units
each. Each subgroup was dedicated to o n e of the three letter positions
within a word, and it represented one of the 10 possible letters in that
position by having a single active unit for that letter. T h e three-letter
grapheme strings were not English words. They were chosen randomly,
subject to t h e constraint that each of the 10 possible graphemes in each
position had to be used at least once. T h e sen~emegroup was used to
encode the semantic features of the "word."' I t contained 30 units, o n e
for each possible semantic feature T h e semantic features to be associ-
ated with a word were chosen randomly, with each feature having a
probability of 0.2 of being chosen for each word. There were connec-
tions between all pairs of units in the s e m e m e group to allow the net-
work to learn familiar combinations of semantic features. There were
n o direct connections between the grapheme and s e m e m e groups.
Instead, there was an intermediate layer of 20 units, each of which was
connected to all the units in both the grapheme and t h e semenie
groups. Figure 4 is an artist's impression of the network. It uses
English letters and words to convey the functions of the units in t h e
various layers. Most of the connections are missing.

T h e Training Procedure

T h e network was trained to associate each of 20 patterns of activity

in t h e grapheme units with an arbitrarily related pattern in the s e m e m e

7 T h e representation of meaning 1s clearly more complicated than just a set of features,

so the use of the word "semantic" here should not be taken too literally.
FIGURE 4. Part of the network used for associating three-letter words with sets of
semantic features. English words are used in this figure to help convey the functional
roles of the units. In the actual simulation, the letter-strings and semantic features were
chosen randomly.

units. As before, the training procedure alternated between two phases.

In phase+ all the grapheme and s e m e m e units were clamped in states
that represented the physical form and the meaning of a single word,
and the intermediate units were allowed to change their states until the
system approached thermal equilibrium at a temperature of 10. T h e
annealing schedule was: 2 at 30, 2 at 26, 2 at 22, 2 at 20, 2 at 18, 2 at
16, 2 at 15, 2 at 14, 2 at 13, 4 at 12, 4 at 11, 8 at 10. After annealing,
the network was assumed to be close to thermal equilibrium and it was
then run for a further 5 iterations during which time the frequency with
which each pair of connected units were both o n was measured. This
was repeated twice for each of the 20 possible graphemelsememe asso-
ciations and the co-occurrence statistics were averaged over all 40
annealings to yield an estimate, for each connection, of p;. In phase-,
only the grapheme units were clamped and the network settled to
equilibrium (using the s a m e schedule as before) and thus decided for
itself what s e m e m e units should be active. T h e network was then run
for a further 5 iterations and the co-occurrence statistics were collected
for all connected pairs of units. This was repeated twice for each of the
20 grapheme strings and the co-occurrence statistics were averaged to
yield an estimate of p i . Each learning sweep thus involved a total of
80 annealings.
After each sweep, every weight was either incremented or decre-
mented by 1, with the sign of the change being determined by the sign
of p: - p . In addition, s o m e of the weights had their absolute

8 See Hinton, Sejnowski, and Ackley (1984) for a discussion of the advantages of
discrete weight increments over the more obvious steepest descent technique in which
the weight increment is proportional to p g - pi,-.
magnitude decreased by 1. For each weight, the probability of this hap-
pening was 0.0005 times the absolute magnitude of the weight.
We found that the network performed better if there was a
preliminary learning stage which just involved the sememe units. In
this stage, the intermediate units were not yet connected. During
phasef the required patterns were clamped on the senieme units and p,:
was measured (annealing was not required because all the units
involved were clamped). During phase- no units were clamped and the
network was allowed to reach equilibrium 20 times using the annealing
schedule given above. After annealing, p,; was estimated from the co-
occurrences as before, except that only 20 phase- annealings were used
instead of 40. There were 300 sweeps of this learning stage and they
resulted in weights between pairs of sememe units that were sufficient
to give the sememe group an energy landscape with 20 strong minima
corresponding to the 20 possible "word meanings." This helped subse-
quent learning considerably, because it reduced the tendency for the
intermediate units to be recruited for the job of modeling the structure
among the sememe units. They were therefore free to model the struc-
ture between the grapheme units and the sememe units.9 The results
described here were obtained using the preliminary learning stage and
so they correspond to learning to associate grapheme strings with
"meanings" that are already familiar.

The Performance of the Network

Using the same annealing schedule as was used during learning, the
network can be tested by clamping a grapheme string and looking at the
resulting activities of the sememe units. After 5000 learning sweeps, it
gets the semantic features exactly correct 99.3% of the time. A
performance level of 99.9% can be achieved by using a "careful"
annealing schedule that spends twice as long at each temperature and
goes down to half the final temperature.

The Effect of Local Damage

T h e learning procedure generates weights which cause each of the

units in the intermediate layer to be used for many different words.

9 There was no need to have a similar stage for learning the structure among the gra-
pheme units because in the main stage of learning the grapheme units are always clamped
and so there is no tendency for the network to try to model the structure among them.
This kind of distributed representation should be more tolerant of local
damage than the more obvious method of using one intermediate unit
per word. W e were particularly interested in the pattern of errors pro-
duced by local damage. If the connections between sememe units are
left intact, they should be able to "clean up" patterns of activity that are
close to familiar ones. So the network should still produce perfect out-
put even if the input to the sememe units is slightly disrupted. If the
disruption is more severe, the clean-up effect may actually produce a
d@erent familiar meaning that happens to share the few semantic
features that were correctly activated by the intermediate layer.
T o test these predictions we removed each of the intermediate units
in turn, leaving the other 19 intact. We tested the network 25 times on
each of the 20 words with each of the 20 units removed. In all 10,000
tests, using the careful annealing schedule, it made 140 errors (98.6%
correct). Many errors consisted of the correct set of semantic features
with o n e or two extra or missing features, but 83 of the errors consisted
of the precise meaning of some other grapheme string. An analysis of
these 83 errors showed that the hamming distance between the correct
meanings and the erroneous ones had a mean of 9.34 and a standard
deviation of 1.27 which is significantly lower ( p < .01) than the com-
plete set of hamming distances which had a mean of 10.30 and a stan-
dard deviation of 2.41. We also looked at the hamming distances
between the grapheme strings that the network was given as input and
the grapheme strings that corresponded to the erroneous familiar mean-
ings. T h e mean was 3.95 and the standard deviation was 0.62 which is
significantly lower ( p < .01) than the complete set which had mean
5.53 and standard deviation 0.87. (A hamming distance of 4 means
that the strings have one letter in common.)
In summary, when a single unit is removed from the intermediate
layer, the network still performs well. The majority of its errors consist
of producing exactly the meaning of some other grapheme string, and
the erroneous meanings tend to be similar to the correct one and to be
associated with a grapheme string that has one letter in common with
the string used as input.

T h e Speed of Relearning

T h e original learning was very slow. Each item had to be presented

5000 times to eliminate almost all the errors. One reason for the slow-
ness is the shape of the G-surface in weight-space. It tends to have
long diagonal ravines which can be characterized in the following way:
In the direction of steepest descent, the surface slopes steeply down for
 7. L E A R N I N G IN BOLTZMANN MAC'I1INES 309

a short distance and then steeply up again (like the cross-section of a

ravine).I0 l n most other directions the surface slopes gently upwards. In
a relatively narrow cone of directions, the surface slopes gently down
with very low curvature. This narrow cone corresponds to the floor of
the ravine and to get a low value of G (which is the definition of good
performance) the learning must follow the floor of the ravine without
going u p the sides. This is particularly hard in a high-dimensional
space. Unless t h e gradient of the surface is measured very accurately, a
step in t h e direction of the estimated gradient will have a component
along t h e floor of the ravine and a component up o n e of the many
sides of t h e ravine. Because the sides are much steeper than the floor,
t h e result of the step will be to raise the value of G which makes
performance worse. Once out of the bottom of the ravine, almost all
the measurable gradient will be down towards the floor of the ravine
instead of along the ravine. As a result, the path followed in weight
space tends to consist of an irregular sloshing across the ravine with
only a small amount of forward progress. W e are investigating ways of
ameliorating this difficulty, but it is a well-known problem of gradient
descent techniques in high-dimensional spaces, and it niay be
unavoidable.
T h e ravine problem leads to a very interesting prediction about
relearning when random noise is added to the weights. T h e original
learning takes the weights a considerable distance along a ravine which
is slow and difficult because most directions in weight space are up the
sides of t h e ravine. When a lot of random noise is added, there will
typically be a small component along the ravine and a large component
up the sides. Performance will therefore get much worse (because
height in this space means poor performance), but relearning will be
fast because the network can get back most of its performance by sim-
ply descending to the floor of the ravine (which is easy) without niak-
ing progress along the ravine (which is hard).
T h e s a m e phenomenon can be understood by considering the energy
. landscape rather than the weight-space (recall that o n e point in weight-
space constitutes a whole energy landscape). Good performance
requires a rather precise balance between the relative depths of the 20
energy minima and it also requires that all the 20 minima have consid-
erably lower energy than other parts of the energy landscape. T h e bal-
ance between t h e minima in energy-space is the cross-section of t h e
ravine in weight-space (see Figure 5 ) and the depth of all t h e minima
compared with t h e rest of the energy landscape corresponds to the
direction along the ravine. Random noise upsets the precise balance

l o The surface is never very steep. Its gradient parallel to any weight axis must always
lie between 1 and - 1 because it is the difference of two probabilities.
3 10 BASIC MECHANISMS

increase weights that help B -b

FIGURE 5 . O n e cross-section of a ravine in weight-space. Each point in weight space

corresponds to a whole energy landscape. T o indicale (his, we show how a very simple
landscape changes as the weights are changed. Movement to the right along the x-axis
corresponds to increasing the weights between pairs of units that are both on in state B
and not both o n in state A. This increases the depth of A. If the task requires that A
and B have about the same depth, an imbalance between them will lower the
performance and thus raise G.

between the various minima without significantly affecting the gross

topography of the energy landscape. Relearning can then restore most
of the performance by restoring the balance between the existing
minima.
T h e simulation behaved as predicted. The mean absolute value of
the weights connecting the intermediate units to the other two groups
was 21.5. These weights were first perturbed by adding uniform ran-
dom noise in the range -2 to +2. This had surprisingly little effect,
reducing the performance using the normal annealing schedule from
99.3% to 98.0%. This shows that the network is robust against slight
noise in the weights. T o cause significant deterioration, uniform ran-
dom noise between -22 and +22 was added. On average, this perturbs
each weight by about half its magnitude which was enough to reduce
normal performance to 64.3%correct. Figure 6 shows the course of the
relearning and compares it with the speed of the original learning when
performance was at this level. It also shows that other kinds of damage
produce very similar relearning curves.
 I

0 5 10 15 20 25 30
Learning sweeps

FIGURE 6. The recovery of performance after various types of damage. Each data-
point represents 500 tests (25 with each word). The heavy line is a section of the original
learning curve after a considerable number of learning sweeps. I t shows that in the origi-
nal learning, performance increases by less than 10°/o in 30 learning sweeps. All the other
lines show recovery after damaging a net that had very good performance (99.3%
correct). The lines with open circles show the rapid recovery after 20% or 50% of the
weights to the hidden units have been set to zero (but allowed to relearn). The dashed
line shows recovery after 5 of the 20 hidden units have been permanently ablated. The
remaining line is the case when uniform random noise between -22 and f 2 2 is added to
all the connections to the hidden units. In all cases, a successful trial was defined as one
in which the network produced exactly the correct semantic features when given the gra-
phemic input.

Spontaneous Recovery of Unrehearsed Items

When it learns the associations, the network uses distributed

representations among the intermediate units. This means that many
of the weights are involved in encoding several different associations,
and each association is encoded in many weights. If a weight is
changed, it will affect several different energy minima and all of them
will require the same change in the weight to restore them to their
3 12 BASIC MECHANISMS

previous depths. So, in relearning any one of the associations, there

should be a positive transfer effect which tends to restore the others.
This effect is actually rather weak and is easily masked s o it can only be
seen clearly if we retrain the network on most of the original associa-
tions and watch what happens to the remaining few. As predicted,
these showed a marked improvement even though they were only ran-
domly related to the associations on which the network was retrained.
W e took exactly the same perturbed network as before (uniform ran-
d o m noise between +22 and -22 added to the connections to and from
the intermediate units) and retrained it on 18 of the associations for 30
learning sweeps. T h e two associations that were not retrained were
selected to be ones where the network made frequent minor errors
even when the careful annealing schedule was used. As a result of the
retraining, the performance on these two items rose from 301100
correct to 90/100 correct with the careful schedule, but the few errors
that remained tended to be completely wrong answers rather than
minor perturbations of the correct answer. We repeated the experiment
selecting two associations for which the error rate was high and the
errors were typically large. Retraining on the other 18 associations
caused a n improvement from 17/100 correct to 981100 correct.
Despite these impressive improvements, the effect disappeared when
we retrained o n only 15 of the associations. The remaining 5 actually
got slightly worse. It is clear that the fraction of the associations which
needs to be retrained to cause improvement in the remainder depends
o n how distributed the representations are, but more analys~s is
required to characterize this relationship properly.
T h e spontaneous recovery of unrehearsed items seems paradoxical
because the set of 20 associations was randomly generated and so there
is n o way of generalizing from the 18 associations on which the net-
work is retrained to the remaining two. During the original learning,
however, t h e weights capture regularities in the whole set of associa-
tions. In this example, t h e regularities are spurious but the network
doesn't know that-it just finds whatever regularities it can and
expresses the associations in terms of them. Now, consider two dif-
ferent regularities that are equally strong among 18 of the associations.
If o n e regularity also holds for the remaining two associations and the
other doesn't, t h e first regularity is more likely to be captured by the
weights. During retraining, the learning procedure restores the weights
to the values needed t o express the -regularities it originally chose to
capture and it therefore tends to restore the remaining associations.
It would b e interesting to see if any of the neuro-psychological data
on the effects of brain damage could be interpreted in terms of the
kinds of qualitative effects exhibited by the simulation when it is
damaged and relearns. However, we have not made any serious
attempt t o fit the simulatiofl to particular data.

CONCLUSION

W e have presented three ideas:

Networks of symmetrically connected, binary units can escape

from local minima during a relaxation search by using a sto- ..
chastic decision rule.

T h e process of reaching thermal equilibrium in a network of

stochastic units propagates exactly the information needed to d o
credit assignment. This makes possible a local learning rule
which can modify the weights s o as to create new and useful
feature detectors. The learning rule only needs to observe how
often two units are both active (at thermal equilibrium) in two
different phases. It can then change the weight between the
units to make the spontaneous behavior of the network in one
phase mimic the behavior that is forced on it in the other
phase.

T h e learning rule tends to construct distributed representations

which are resistant to minor damage and exhibit rapid relearn-
ing after major damage. The relearning process can bring back
associations that are not practiced during the relearning and are
only randomly related to the associations that are practiced.

These three ideas can be assessed separately. In particular, resistance

to damage, rapid relearning, and spontaneous recovery of unrehearsed
items can be exhibited by other kinds of parallel network that use
distributed representations. The use of stochastic units, annealing
search, and the two-phase learning algorithm are not crucial for these
properties, though they are a convenient testbed in which to investigate
them. Hogg and Huberman (1984) have demonstrated self-repair
effects in nonstochastic, layered networks similar to those used by
Fukushima (1980).
W e have left many loose ends, some of which are discussed else-
where. Sejnowski and Hinton (in press) give a detailed example of a
search problem where annealing helps, and they also discuss the rela-
tionship between between these networks and the mammalian cortex.
Ackley, Hinton, and Sejnowski (1985) give a different example of
learning in which the network constructs efficient internal codes for
communicating information across narrow bandwidth channels. At
present, the learning algorithm is too slow to be tested properly on
large networks and future progress hinges on being able to speed it up.

ACKNOWLEDGMENTS

This research was supported by grants from the System Development

Foundation. W e thank David Ackley, Peter Brown, Francis Crick,
Mark Derthick, Scott Fahlman, Stuart Geman, John Hopfield, Paul
Kienker, Jay McClelland, Barak Pearlmutter, David Rumelhart, Tim
Shallice, and Paul Smolensky for helpful discussions.
 7 L E A R N I N G IN BOLTZMANN MACIiINES 315

APPENDIX:
DERIVATION OF THE, LEARNING ALGORITHM

When a network is free-running at equilibrium the probability distrib-

ution over the visible units is given by
xp- 1 Emil (8)
P
P-( v ) = C P - (V,,A H @ )=
LY
@ x$/
A l*

where V,, is a vector of the states of the visible units, Ha is a vector of

states of the hidden units, and E,, is the energy of the system in state
V,Y Hp
E,,, = - x
[<I
w, s,"hs;'@.

Hence,

Differentiating ( 8 ) then yields

This derivative is used to compute the gradient of the G-measure

where p + ( V a ) is the clamped probability distribution over the visible

units and is independent of w U . So
Now,
P + ( V , A H,,) = p+(HpI
P-(v,AH,,) = p-(HpI
and

Equation 9 holds because the probability of a hidden state given some

visible state must be the same in equilibrium whether the visible units
were clamped in that state or arrived there by free-running. Hence,

Also,

Therefore,

where

and

The Boltzmann machine learning algorithm can also be formulated as

an input-output model. The visible units are divided into an input set I
and an output set 0, and an environment specifies a set of conditional
probabilities of the form P (Og I I,). During phasef the environment
+
 7. LEARNING I N BOLTZMANN MACI-IINES 3 17

clamps both the input and output units, and the p,Ts are estimated.
During phase- the inputsunits are clamped and the output units and
hidden units free-run, and the p,;s are estimated. T h e appropriate G
measure in this case is
P+(O, I I,)
G = ~ P + ( I , A O,)ln
cx P P - ( O p 1 IJ -
Similar mathematics apply in this formulation and a G / a wi, is the same
as before.

View publication stats