ÐÓ°ÉÔ­´´

Three centuries ago, the Dutch philosopher Spinoza made the shocking
claim, published only after his death, that it is possible to devise a scientific
psychology fully consistent with our knowledge of how the body works. He
accused his opponents of underestimating the potentialities of mechanism:
‘They do not know what the body can do, nor what can be deduced from the
consideration of its nature alone’. His ideas contributed to the notion
of man as machine – an idea developed by many, including the eccentric Lady
Ada Lovelace. More than 100 years after Spinoza’s death, she argued that
a machine based on Charles Babbage’s Analytical Engine might carry out other
types of computation besides mathematics: for instance, to ‘compose elaborate
and scientific pieces of music of any degree of complexity or extent’.

Today such ideas are still widely perceived as shocking – perhaps in
some ways even more shocking than in past centuries. Most people are deeply
sceptical about whether computers could ever model the human mind, and many
see the idea as obviously absurd. Moreover, their scepticism and mockery
are now commonly accompanied by fears that if we allow humanity’s image
to be moulded in the likeness of a computer, humane values must take second
place, or even be negated altogether.

So it is not surprising that people mistrust psychological theories
based on computer modelling. Similar fears have greeted the technology of
artificial intelligence, which aims to develop computers that can perform
the same feats as the human mind. The deepest anxiety is that such theories
and technologies will impoverish our image of ourselves, lower our morale
and increase the individual’s sense of helplessness in the face of life’s
challenges. If theoretical psychology tells us we are ‘nothing but machines’,
it would seem to follow that social practices and personal attitudes which
value mankind’s specifically human qualities must be sentimental illusions.
If computational psychology gains ground, and if artificial intelligence
succeeds as a technology, such practices and attitudes will – it seems –
be undermined accordingly.

In asking whether computer models of mind must necessarily have this
pernicious social effect, we should remember that Spinoza warned against
underestimating the potentialities of mechanism. For the dehumanisation
so widely feared is inevitable only if computational theories cannot in
principle explain how it is possible for a mechanism to generate the properties
of the human mind.

Our estimation of what an understanding of mechanism can achieve has
broadened since Spinoza’s time. But the mechanisation of thought is only
possible if other things besides numbers can be computed. For example, we
have to be able to show that the human brain operates as a particular type
of machine, doing things in ways the scientists can explain.

Computer scientists focus on the computational properties of electronic
computers. And psychologists who ground their theories in computer modelling
are using electronic computers to illustrate their claims about the psychological
functions of the brain. It is on this point that many people base their
critique of computer models of mind.

According to such critics, if Gertrude Stein had ever said ‘A machine
is a machine, is a machine . . .’, she would have been wrong. The brain
is a machine of a very special sort. The fact that it can support psychological
functions does not show that computers can. Indeed (so the objection continues)
they almost certainly cannot, because they are machines of a type fundamentally
different from the brain. The computers used for mental modelling are digital,
serial, general-purpose devices. The brain is an analogue, parallel-processing,
special-purpose machine. Whether only brains, or computer models with brain-like
properties, can do what human minds can do is controversial. In the practice
of computer modelling, the pendulum has swung between one answer and the
other.

The clearest example of this pendulum swing concerns computer models
of vision. In the early days, research on vision attempted to mimic the
parallel functioning of the human brain. This research subsequently fell
out of favour, but recently some of the special features of the human brain
are being taken seriously again. This is true not only of work in theoretical
psychology, but of artificial intelligence too. For the technologist, like
the psychologist, needs to know how it is possible for a visual system to
form accurate descriptions of things in the external world, given only the
input of light, or an image. How, for example, can one see something as
a cube or a teddy bear, or as something which is a couple of yards away,
about a foot long, with an undulating spotted surface slanting away from
the ground?

The earliest work in the computer modelling of vision produced parallel-processing
systems analogous to simple neural networks, collections of units broadly
modelled on the cells of the brain. These programs, which could recognise
simple patterns, were made up of several independent information-processing
units, or ‘demons’, and a ‘master demon’. Each low-level demon watched out
for one part of the pattern, and sent a message to the master demon when
it found it. A demon could ‘shout’ or ‘whisper’ according to the probability
and importance of its message, but could not influence the loudness of its
neighbours’ shouting. On the basis of the messages coming in from the various
demons, the master demon would decide what overall pattern was present.
For instance, it would report an ‘F’ if it was told of two horizontals and
one vertical in the appropriate positions.

There is a clear analogy between these single-minded demons and the
remarkable ‘feature-detector’ cells in the visual cortex, which respond,
for instance, only to a line of a certain orientation, or an edge moving
in a particular direction. Indeed, it was the ideas of these early computer
modellers which first suggested that feature detectors might exist, so prompting
neurophysiologists to search for them. This example shows that computational
models of psychological processes can sometimes help brain science by concentrating
on functional units rather than focusing attention on material ones.

It was because the computational functions necessary for vision had
not yet been properly identified that these early models were abandoned.
Despite their apparently brain-like organisation, they were incapable of
‘seeing’ solid objects such as cubes or teddy bears for what they are. While
they could distinguish some patterns, they could not interpret them, because
they embodied no knowledge about how three dimensions can be projected into
two. Vision in general is not mere pattern recognition, but image interpretation.
Consequently, the mere counting and classification of pattern properties
is not enough to enable a system to see.

When this point was realised, the pendulum in the computer modelling
of vision swung away from models of neural networks. Research turned instead
to visual interpretations, whereby images were treated not as patterns but
as representations of the real world.

The interpretations were built up step by step by general-purpose digital
computers and were based on projective geometry, which describes how solid
objects of certain types would appear to an observer from different points
of view. So systematic geometrical knowledge about mapping from two dimensions
to three dimensions was built into computer programs for ‘scene analysis’
(as opposed to ‘pattern recognition’). These programs used their stored
knowledge to build sensible 3-D descriptions of objects, given depictions
of those objects in 2-D line drawings. A drawing of a cube, for instance,
would be recognised as a representation of a cube.

In general, a scene analysis program could interpret line drawings of
those types of object which it knew about already. High-level knowledge
about a given class of objects would be used to guide the visual interpretation:
thus these programs ‘knew what to look for’ in the 2-D image. The idea was
that in perceiving a cube, for instance, we unconsciously use our knowledge
about how the various corners will appear in an image of the cube.

It is no accident that my main example here has been a cube, rather
than the teddy bear mentioned earlier. Teddy bears were, in effect, invisible
to scene analysis computers, which could not even identify them as solid
objects. Still less could they describe a teddy bear as something with a
smooth, furry surface, having bits sticking out here and there, and two
shiny, round bumps near one end.

There were three reasons for the ‘invisibility’ of teddy bears. First,
the simple projective geometry being used allowed straight edges to be described,
but not ‘curvy’ outlines. Secondly, these models had to be preprogrammed
with detailed knowledge of what they were going to see: since they were
not told (and could not be told) what teddy bears look like, they had no
way of finding their salient curves or surfaces. And thirdly, scene analysis
programs could not see localised depth, orientation or surface texture.
So even a fabric-covered cube, such as one might give to a very young baby,
could not be described by a scene analysis program as furry, nor a plastic
one as smooth. These serious deficiencies are now being largely overcome,
as the pendulum swings back towards more ‘brain-like’ models. Once again,
the emphasis is on parallel processing within networks of highly specialised,
elementary units.

But there are three important differences between current ‘connectionist’
systems and neural networks as previously understood. Each individual unit
(or demon) in a connectionist system gives a 3-D interpretation of the 2-D
image point it looks at. The units are not independent, but can influence
the loudness of their neighbours’ shouts by way of feedback rather like
the excitation and inhibition between neurons in the brain. And the units
are specifically designed, and interconnected, by reference to a powerful
general theory of image interpretation.

This theory relies on the physics of image formation, and the associated
principles of mapping two dimensions to three dimensions, which describe
how light can be reflected from physical surfaces of various sorts. They
specify, for instance, how light is reflected from a surface oriented at
a particular angle relative to the viewer, lying at a particular distance
from the eye, or perhaps at differing distances from the two eyes.

The overall surface of any physical object is made up of many tiny areas,
among which the local neighbours are usually similar. Thus a furry area
on the body surface of a teddy bear is usually surrounded by other furry
areas, and a glassy area by other glassy areas. Only round the edge of a
glass eye, or a claw, will this not be true. Consequently, neighbouring
points in the image tend to be similar too, so that boundaries in the image
often correspond to real boundaries between objects in the real world.

Often – but not always. In a world where Dalmatians exist, not every
discernible difference boundary in the image can be correctly interpreted
in this way. The white and black patches on a Dalmatian’s coat are not different
objects: they are part of one and the same physical surface, attributable
to one and the same object.

However, it is a rare Dalmatian whose surface markings coincide exactly
with its bodily contours. Consequently, the outer black and white points
on neighbouring patches usually lie at the same distance from the viewer.
So patch edges can be distinguished from the genuine contours of the object,
provided that some visual units can interpret image points in terms of their
distance from the viewer, and provided that they can influence the interpretations
suggested by the patch detectors. Only if the distance detectors signal
a depth discontinuity at the relevant position in three-dimensional space
will a colour boundary be interpreted by the visual system as a bodily contour,
rather than as a surface marking or patch.

The overall three-dimensional interpretation arrived at by the system
is usually sensible, because the feedback between individual units takes
account of the physical possibilities of images in general. That is, the
excitatory and inhibitory connections enable the units to tell each other
to shout more or less loudly, so resulting in a mutually consistent set
of shouts.

This sort of visual system could even pick out a Dalmatian dog lying
on a black and white background. To do this, one must be able to distinguish
a black image patch caused by a black spot on the dog’s coat from an immediately
contiguous black image patch caused by markings on the surface the dog is
lying on. Suppose some processing units had described this part of the image
as one black patch (for that is how it appears, in the image). If the distance
detectors identified a line of depth disparity running through this area,
they could pass messages to the patch-detecting units concerned, to inhibit
them from describing this as a patch ascribable to one surface. Similarly,
the units signalling an object contour (because of the depth disparity)
would be activated if the texture-detecting units at that position were
signalling different textures on either side of the depth-disparity line.
So this visual system (just like you) would find it easier to see a Dalmatian
if it was lying on black and white tiles than if it was lying on a black
and white fur rug.

The pendulum swing we have discussed suggests that any visual system
capable of reliable mapping from two to three dimensions must be organised
in ways broadly similar to the brain. These recent computer models are ‘brain-like’
in that they are based on parallel processing, using analogue units (which
can shout more or less loudly) dedicated to seeing certain things (the units
which can see depth cannot see lines or colour). At present they are usually
simulated in digital computers, but suitable parallel processing machines
are currently being developed.

Although many questions remain, the psychology of vision has been significantly
advanced by this approach. For example, we now have a much better understanding
of ‘stereopsis’, the sort of depth vision that relies on disparities between
the images presented to the two eyes. The reason is not that these computer
models have provided any new physiological facts, but that we now understand
what it is that a stereoptic system has to do, and how it could be doing
it. However, these ideas might help neurophysiology if they help relevant
brain mechanisms to be identified; this has already happened in the case
of the single-cell feature detectors mentioned above.

Psychology is not about the material nature of the brain, but about
what the brain is doing. What is psychologically important about a machine
that can pick out Dalmatians on piebald rugs is not the material it is made
of, whether protoplasm or silicon chips, but how the material is organised
and what computational functions it can support. These functions, and the
knowledge embodied in it (the general rules for mapping from two to three
dimensions) enable it to construct sensible representations of objects placed
on rugs. These internal representations are what the system uses to make
judgments or form beliefs about the objects such as where they are, and
how big they are; such beliefs might influence decisions about how to walk
round the object without stepping on it. And if the system also has some
specific knowledge about Dalmatians, it will be able to recognise some objects
as Dalmatians.

Because the natural sciences have nothing to say about computational
functions, knowledge, representation and the like, they are unsuitable for
understanding the mind. Psychology is radically different from physics and
neurophysiology. The natural sciences have no way of saying that individual
people see the world differently, that one person’s actions differ from
another’s because they have different political beliefs, personal priorities
or cultural backgrounds. Questions about such human matters therefore cannot
be asked or answered using the vocabulary of natural science. This is why
people committed to a scientific world-view often ignore them, or dismiss
them as woolly minded sentimentalities. At best, they are relegated to the
sphere of literature or poetry: all very well for a wet Sunday afternoon,
but nothing to do with the serious business of the scientific life.

Yet thinking of the mind as a computational system leaves room for the
view that our beliefs and values are psychologically crucial. Whether idiosyncratic
or culturally shared, they inform and mediate our perceptions, actions and
decisions. If one does not believe that Dalmatians are black and white dogs,
then one will not be able to identify the Dalmatian on the fireside rug
– even though one may be able to see it as a solid object, or as a dog.
And if one does not object to stepping on dogs, one may not need information
of its whereabouts in deciding how to walk across the room. In general,
we are subjective creatures. Or, as a computationally inclined colleague
put it, ‘we inhabit our data structures’.

This in no way conflicts with the mechanistic view. Because computer
models show how it is possible for specific machines to make complex visual
discriminations, they help us to appreciate the potentialities of mechanism
in general. People acquainted only with machines such as cars or typewriters
understandably feel denigrated to be classified by science as (nothing but)
a machine. To be compared to a computational system of the kind psychologists
are interested in is a different matter.

This is not to say that all our mental powers are fully understood,
or have been modelled on computers. As Lady Lovelace might have put it,
theoretical psychologists have not expressed the ‘mutual fundamental relations’
of all thought processes by ‘those of the abstract science of operations’.
Even assuming this to be in principle possible, it will not be achieved
soon. For the human mind is far richer than we usually think. Many humanists
have an intuitive sense of the mind’s subtlety, but they have not provided
a detailed account of the complexities involved. Indeed, with the possible
exception of Freud, they have not even attempted to do so. It is only when
psychologists try to express a theory sufficiently clearly and fully to
make a computer do something ‘human-like’ that they begin to realise the
extent of the task.

Even the everyday abilities that we all share, and which to introspection
seem simple and effortless, turn out to be enormously complex. Consequently,
the technology of artificial intelligence is, and may forever remain, far
less powerful than its images in science fiction – or in our fears. People
who feel denigrated because, unlike computer programs, they cannot solve
mathematical equations in a split second should remember that they can do
other things which no computer can do. Their morale and self-image may be
less threatened if they realise that the ‘thinking’ and ‘understanding’
of even the most impressive computer programs of today are paltry compared
to ours.

This is often not realised, because we do badly what programs do well.
Computers are better at mathematics, and at precisely specifiable scientific
reasoning, than many (sometimes, most) human specialists are. This is why
so-called ‘expert systems’ are possible: programs which can solve problems
in narrowly delimited areas of stereochemistry, geological oil prospecting
or medical diagnosis. But what is usually forgotten is that the things that
we all do well, computers can hardly do at all. We can understand our mother
tongue; recognise objects partially hidden by other things; use our common
sense in tackling a problem, or in ‘reading between the lines’ of speeches,
letters or newspapers; and we can use our fingers dextrously to perform
a variety of manual tasks. Current computer technology can do these things
only in highly limited ways.

In sum, and paradoxical though it may seem, computer models of mind
can be positively rehumanising. Thanks to their influence, ‘mind’ and ‘mental’
processes are now respectable concepts in psychology (which in the days
of behaviourism they were not). This is important not only for psychologists,
but for society in general. For, as counselling psychologists rightly remind
us, how people think about themselves matters.

Science will be dehumanising only if it has no room for mental concepts,
and no vocabulary of subjectivity. The natural sciences, including ‘pure’
neurophysiology, do not. But psychology, computer science and neuroscience,
in so far as they focus on the brain’s computational functions, all do.

So the claim that a scientific psychology is possible should not be
so shocking now as it was in Spinoza’s time. Provided they are properly
understood, computer models of mind need not be socially pernicious.

Margaret Boden is professor of philosophy and psychology at the University
of Sussex. Her latest book is The Creative Mind: Myths and Mechanisms, published
by Weidenfeld & Nicolson, 1990. This article is an edited version of
a chapter in Modelling the Mind, edited by K. A. Mohyedlin Said et al, published
by the Clarendon Press, 1990.