Hints from Life to AI, edited by Ugur HALICI, METU, 1994 ã
from
brain to mind;
what
clues from the natural world can tell us
David Davenport
Computer Eng. & Information Sciences Dept.,
Bilkent University, Ankara 06533 - Turkey.
david@bilkent.edu.tr
The
human brain is incredibly complex. Understanding how it functions and how a
conscious self-aware being emerges from its mass of gray matter, is one of the
last great challenges of our age. Evidence for its mechanism has been
accumulating under the banner of cognitive science for many years. From such
diverse disciplines as philosophy, medicine, psychology, computer science and
more recently neurobiology, a wealth of new and exciting discoveries are
providing ever more clues to the inner workings of this most mysterious of
objects. Models based on these observations have already proved valuable in
developing artificially intelligent solutions to practical engineering
problems. Yet, despite these successes and despite the mass of evidence, our
understanding of the relationship between mind and brain remains sketchy at
best. Indeed, the biggest stumbling block is perhaps not the lack of clues, but
the lack of a suitable framework within which to organise them. This paper
reviews some of the available evidence, both to show what we do know and to
illustrate the difficulties involved in knowing more.
1. Introduction
The human brain is the most complex structure in the known
universe. Understanding how it functions, how it makes sense of the world
around it and, above all, how a conscious self-aware being emerges from its
mass of gray matter, is one of the last great challenges of our age. It is a
problem which has exercised and confounded some of the greatest thinkers
through history, from Plato to Decartes, to Russell, Wittgenstein and Quine.
Nor should we forget the psychologists,
the linguists, the computer engineers, the clinicians and the neuroscientists,
who have also brought their very special skills to bare.
The brain is yielding its secrets, albeit reluctantly, as
the scientific method delves ever deeper into its mysteries. We may still be a
long way from establishing a link from brain to mind, but gradually, the
Dualist, Divine and Magical explanations of old, are being banished in favour
of a more materialist understanding. This knowledge may not come for free,
however, for, if we succeed, we risk replacing that intangible human spirit
with a mindless mechanism. So why, then, do we expend such effort? In part, of
course, it is human nature to question and to explore, and there can be no
greater intellectual challenge than that which understanding ourselves
presents. But there are more practical considerations too, for an appreciation
of the workings of our brains may provide valuable insights into the treatment
of various physical and mental ailments. On a more commercial note, such
knowledge can also enable us to build better, more sophisticated, machines.
Artificial Intelligence (AI) research has already demonstrated the utility of
copying and even improving upon nature's designs. A better understanding of the
functioning of, and the relationship between, mind and brain, can only lead to
further improvements.
In fact, we know a great deal about the brain. Our intimate
personal attachment to it means that we have first hand knowledge of many of
its strengths and weaknesses. We are aware of its prodigious memory, capable of
remembering minute details and arranging them in appropriate ways, all without
any apparent effort. We are aware also, that sometimes we are unable to
remember the most obvious or important of facts, however hard we try. Yet at
other times, despite being aware that we do indeed know something, we fail to
recall it when we want to, although we are able to remember it clearly at a later
time or when given an appropriate hint. We are aware of our incredible ability
to recognise people and things that we have seen before, however briefly. We
are aware of our creative and linguistic abilities, and of our ability to learn
new skills and to adapt to new conditions.
To this first hand knowledge, we can add that from the
fields of linguistics, psychology, neurobiology, philosophy and computer
science, where new and exciting discoveries continue to provide ever more clues
to the inner workings of the mind. We now know that the brain is not the
amorphous mass of gray matter it once appeared to be. It has a complex internal
structure composed of cells called neurons. These neural cells have many
inputs, called dendrites, but generally only a single output, called the axon.
The axon of one cell connects to the dendrites of others via a junction known
as a synapse, see Figure 1. Neuroscientists have built up an incredibly
detailed picture of the biological and chemical mechanisms involved in the neuron,
its interconnections, and, above all, in the synaptic junction itself.
Unfortunately, discussion of this work is beyond the scope of this paper, the
interested reader is referred to the many standard works on this topic, e.g.
[1,2].
|
|
Figure
1. Building blocks of
the brain, neurons, axons, dendrites and synapses
(from [6])
Clearly then, we have a mountain of information about the
brain, from the behavioural level down to the molecular level. Yet, for all the
clues at our disposal, there is one thing we apparently still do not know.
Incredible as it may seem, we really don't know how the brain/mind works! How
is it possible to know so much and yet so little? Why can we not piece together
a proper understanding, when we seem to have so much information available? Why
is it so difficult? These are the questions which this paper attempts to
answer. To do so, we will first look at what would constitute an understanding
or explanation, and how we would actually go about achieving this. We then
review some of the clues uncovered thus far, both to illustrate how difficult
and error-prone the process of discovery can be, and to give some hint at what
we do currently know about the brain. We conclude by showing how the current
models of mind apparently fail to provide a suitable framework into which to
organise our knowledge.
2.
Models and Explanations
Why do we still have only a vague understanding of the way
the mind functions, despite an apparent mass of information? Why is it so
difficult to piece together a complete picture of that with which we have so
intimate a relationship? To see the problem, it is necessary to appreciate just
what we mean when we say we understand a complex system. Having got an idea of
the form our understanding will take, we can begin to see why the process by
which it is acquired is so full of pitfalls.
Models are the key to understanding the functioning of
complex systems. We build models as an aid to comprehension. Most often this involves
constructing an abstract view of a system, one which omits much of the
intricate detail. In the extreme, a simplified model would treat a system
simply as a "black box", whose output is some function of its inputs
and related internal states (which may themselves be a function of previous
inputs). This straightforward mapping of inputs to outputs constitutes an
essentially "behavioural" description of the system. In fact, we
frequently retain some of the internal causal structure of the system in our
simplified model. Such intermediate-level models can be viewed as a set of even
simpler black box models, together with appropriate causal connections between
them. Employing such abstract descriptions (models) proves advantageous because
of our limited cognitive abilities. We find it very difficult to remember and
use large numbers of unorganised facts, such as would comprise a fully detailed
description of a complex system. Simplified models have other benefits, too,
especially in terms of allowing faster predictions to be made and of reducing
the amount of time required to learn about a system.
We can obtain more abstract models in two ways, either by
collapsing parts of the system into black box models or by reducing the
precision with which we describe the system. Conceptually, we could collapse
any sub-section of the system to a black box, although it will usually be
convenient to divide it up along causal lines. Reducing the number of
components and interconnections in this way has obvious effects on the amount
of computation needed when using the model and may well result in fewer
internal state being required. Precision, on the other hand, relates to the
amount of detail with which inputs, outputs and internal states are
represented, such as, present/absent, low/normal/high, a real value between 5
and 5000, etc. Reducing precision helps by lowering the number of input/output
mappings to be considered when using the model to make predictions. This is
particularly important if the model must be used "bidirectionally",
that is, when computing outputs given inputs or computing inputs given outputs.
Notice that precision is not the same as accuracy, which relates to the
system's ability to make correct predictions. Hence, it is possible that an
imprecise model may make accurate (right) predictions and a precise model
inaccurate ones. The trick, of course, is to build an accurate model which has
a level of precision appropriate to the purpose at hand. In practice, it is
often convenient to retain a hierarchy of models each with varying degrees of
precision and internal structure. We can then select the most suitable model
depending on the circumstances at the time, paying attention to the mapping
between the various models and the actual system.
Now, consider a situation in which we have a model at a
particular level of description, but we wish to improve it. It may be that the
model is not sufficiently detailed for our purposes, or perhaps, it is not
completely accurate. Obviously, it would be relatively easy to start from the
original system and derive another model, which is more accurate or detailed
than the one we currently possess. But, what if we don't understand or know any
more about the actual system than the information contained in the present model?
How can we possibly create a better model when we have no idea of what it
should be like? This may sound like an extremely unlikely scenario, but it is
not. It is precisely the sort of situation we face when we wish to improve a
model which represents our current knowledge of some natural phenomena. The
scientific quest for understanding is exactly this process of creating ever
better models of the natural world.
Improving a model can be done in three ways, by increasing
precision, by taking into account previously neglected inputs, or by
discovering more of the system's internal structure. As discussed before, our
models are either black boxes or are composed of them. In this case, however,
the input-output behaviour is usually determined as the result of observation
rather than abstraction. In fact, it is very often attempts to confirm a
model's accuracy through closer observations, perhaps in different contexts,
that reveal its shortcomings in the first place. When this occurs it is
necessary to postulate more appropriate state variables and/or structure.
Unfortunately, this is a very difficult and error-prone task, hence, having
once selected a 'new' model, it must be put to the test. Confirming a model
usually involves a search for supporting evidence. It is generally recognised
that, for a model to be accepted, it should not only account for all of the
existing observations, but also predict as yet unobserved effects. The best
confirmation of a model's accuracy thus comes from experiments which reveal the
existence of these hitherto unseen phenomena. Actually, in practice, most
theories are unable to account even for the known evidence. For this reason, a
model is considered more acceptable if evidence of its components can be
confirmed by multiple independent means. For example, if a certain I/O mapping
necessitates a causal pathway being postulated between two subsystems and there
is evidence of an actual physical link, severance of which produces the
expected dysfunctional behaviour. Of course, none of this can ever prove the
new model is correct or even the best or most detailed model possible, but it
does lend credence to it.
While it is impossible to prove a model correct, a single
piece of evidence could prove it wrong or, at least, in need of improvement.
Unfortunately, human nature is such that it often affords more weight to
positive supporting evidence, with the result that potentially significant
negative evidence tends to be overlooked. On occasion, this may actually be a
preferable, for, if we were to try to solve the entire puzzle at once, we would
get totally confused. Thus, scientific understanding must progress in steps,
each aiming to be more accurate and complete than its predecessor. When
agreeing to overlook some evidence, however, we must be careful, lest we choose
to ignore the wrong clues and end up constructing a misleading model, one that
has little or no basis in reality. Overthrowing established but invalid
theories of the world is remarkably difficult. This is partly due to the fact
that investigations are performed within the scope of the prevailing model and
hence tend to be biased towards locating corroborating evidence and partly
because of the natural reluctance to discard prevailing ideas and start all
over again. Rather, when a relatively well established model is found wanting,
it tends to be "patched" by appending "special cases" to
cover the exceptions. Only when the number of special cases becomes
overwhelming, is the original model thrown into question and the search for a
new one begins.
The following section reviews some of the clues which have
been accumulated in the search for a model of mind. It is organised so as to
try to build up a picture of cognitive process starting with basic questions of
how memory functions, through architectural features and development, to the
question of the relation of mind to the brain itself. In the course of this
journey, we will see many illustrations of the sorts of mistakes described
above, indicating just how difficult and error-prone the process of scientific
modeling can be.
3. Some
Clues
It is almost impossible to provide a complete review of the
mountain of work which constitutes the study of the mind, now going back
hundreds or even thousands of years. Indeed, such is the volume of material
that it is impossible to do more than scratch the surface of work done even
very recently. The following selection thus makes no claims to be comprehensive
or even to be right, for, in the light of the preceding discussion, to do so would
be folly.
Cognition, in essence, is the ability of an agent to detect
and store information about itself, its environment, and its interactions with
the environment, and to subsequently use this stored knowledge when deciding
upon future actions. This view is predicated on the assumption that there is a
certain regularity to the world and that knowledge of which actions have proved
successful or unsuccessful in similar situations in the past, can thus help
when selecting the most appropriate course of action. Memory, therefore, is
central to the whole cognitive process and any attempt to explain human
cognition must offer a vision of how the mind represents the world and how it
comes to acquire this representation.
The neural basis of cognition was first uncovered by Cajal
in 1891, yet, it was only in the 1940's and 50's that many of the details of
its functioning really began to emerge. One of the most important questions
which required answering concerned the number of neurons needed to store each 'concept.'
At one extreme the compact or punctuate view suggests that one neuron per
concept is sufficient, while at the other extreme the diffuse view holds that
each concept is stored as a pattern of activity across all of the neurons in
the brain. Arguments against the punctuate model focus on the fact that a large
number of neurons die each day. Accordingly, we should expect to lose at least
some concepts (at random) every day, which we clearly do not. Other evidence
arrayed against the compact view is the observation that a small stimulus gives
rise to a large amount of neural activity and that, theoretically, there are
simply not enough neurons to have one per concept anyway. However, perhaps the
most significant factor arguing against compact models and for diffuse models,
was Lashley's now classic work with lesioned rats [3]. Despite having large
portions of their cortex removed, the rats continued to show good maze running
performance. While the diffuse model has undoubted advantages in terms of fault-tolerance
and generalisation, it also presents extremely serious theoretical flaws. The
basic difficulties concern cross-talk, communication and the inability to
capture structure. If each concept is represented as a pattern of activity
across all neurons, then trying to consider two concepts at the same time will
result in overlap and confusion [4].
In fact, more recent work has established that remarkably
compact cerebella structures do exist, so that there must be another
explanation for Lashley's results [5]. Moreover, individual 'command' neurons
have been found [6], as have neurons in male zebra finches which respond only
to the song of that particular bird's father, not those of other males of the
same species or to any other pure tones [6]. Experiments conducted on patients
undergoing brain surgery (during which the patient can remain conscious, since
there are no pain sensors in the brain) show that stimulating individual
neurons usually produces very specific sensations. From the foregoing arguments
it should be apparent that the brain seems to use a very compact means of
representation afterall. In fact, the arguments forwarded against the compact model can be easily overcome. For
example, the problem due to neuron death is easily avoided if each 'concept' is
represented by several neurons rather than just one, so that the chances of
losing all of them is very small indeed.
While it seems certain that the brain uses compact storage,
there seems to be no clear-cut evidence to show whether the it uses a prototype
or instance model of concept storage, indeed there is a suggestion that it
employs both together. A pure prototype scheme would not store the individual
instances of a concept but, rather, just some 'average' of them. Obviously,
this results in considerably lower storage requirements, but alas, it throws
away information about the variability of categories, information which people
appear to retain. Another difficulty concerns concept acquisition. How, for
example, is the prototype adjusted as new instances are observed, particularly
in the early stages of concept formation?
In fact, once a concept has been established, it seems easier to
remember atypical instances rather than focal ones which should match the
prototype better.
Interestingly, there is good evidence that recognition and
remembering are separate processes. In one experiment, subjects were shown a
sequence of pictures and asked to identify each as being a member of a
specified category or not. A subsequent test in which the subjects were shown
pictures and asked if they had been part of the original test sequence, showed
very poor recall. The conclusion was that, although only a few tens of
milliseconds were required to correctly match the picture with a category,
several hundred milliseconds were needed to actually record (remember) an
individual picture. Of course, there are lots of other factors affecting
memory, in particular, repetition. The fascinating interplay of recognition and
memory is amply illustrated by the common occurrence of noticing a car number
plate. If you see a car having a personalised number plate, e.g. "The
King", you would almost certainly remember it, whereas, if you saw a
normal license number, you would probably remember only a little bit of it, if
any. Yet, if you saw an Arabic number plate, although you would undoubtedly
remember that you had seen such an unusual license number, you would probably
be unable to recall any detail of it!
Cognitive psychology has uncovered many clues as to the
nature of the recognition process. For example, in an effect known as
"word superiority", a briefly presented single letter is recognised
more accurately when it is alone or part of a word, than when it is part of an
unfamiliar letter string [7]. The key to this effect is a 'mask' made up of
simple lines and shapes, which replaces the letter/word after a short period,
say 60ms. The theory is that recognition requires a hierarchy in which low-
level feature-analyzers, pass their results through to letter- analyzers which
in turn pass the results of their analyses on to word and, eventually,
concept-analyzers. Replacing the target letter with the mask stops further
low-level processing immediately, while the higher levels can continue slightly
longer. Using a mask which does not redirect the low-level detectors
obliterates the effect.
Figure
2. An
example of Leeper's degraded pictures
Another indication of the interplay between various 'levels'
can be seen in an experiment by Neely [8]. It demonstrates that a word is
recognised more rapidly if it is preceded by a priming word that is related in
meaning, than if it is preceded by an unrelated word. Moreover, recognition can
actually be hindered if a related but unexpected word is used as the prime.
Again a mask is used, this time so that the word is 'seen' for only 10ms. The
experiment suggests that even a very brief presentation of a word can activate
the meaning of a related word. One final example of top-down bottom-up
processing is demonstrated by the degraded figures used by Leeper [9], see
Figure 2. Identifying the picture usually proves rather difficult, yet, if you
are told what to expect, it suddenly becomes very obvious (for solution see [9]
in the reference section).
In a certain sense, these observations may be open to
question, since they involve the interaction of many complex and, as yet, only
vaguely understood features of the mind. There are, however, some very well
known clues which may relate to the implementation level more directly. The
first such hint is provided by illusions such as those of Figure 3.
Figure
3. Classic illusions
which may provide clues to brain functioning
You know there is no complete triangle or square, but the
expectation remains. This clearly demonstrates the same sort of processes we
saw above, but this time without the additional complication of language.
Notice, that this hints also, at the same basic mechanisms being used in all stages
of the cognitive process. Another hint is provided by pictures which 'flip'
between two possible interpretations. The "Necker cube" and the
"old man, young woman" are prime examples, the rabbit in Figure 4 is
another (turn the page sideways to see a different animal!). What they seem to
indicate is the existence of a very low-level winner-take-all mechanism.
Another hint, although more difficult to decipher, may be
offered by the effect seen when a plain colour is replaced by white. In such
situations, people generally observe a faint trace of the original shape, but
in the complementary colour. Is this perhaps an overshoot from a resettling of
the winner-take-all effect? Whatever may be the case, there is one more
potentially important hint that is generally completely overlooked. As the
title of Boole's classic book 'The Laws of Thought' quite clearly announces, a
major clue is provided by logic itself. Logic is an abstraction of human
reasoning, but one which concerns itself only with right reasoning. As Dennett
observed [10], logical behaviour is the hallmark of an intentional agent,
illogical agents simply will not survive. We do, on the whole, reason
logically, however, we also display some characteristically illogical traits,
such as 'denying-the-antecedent' and
'asserting-the-consequent'. What we need then is a mechanism which is
essentially logical, but which exhibits the sort of lapses of reasoning which
we ourselves so often do. One possibility would be to remain within the
deductive framework and somehow suppress the valid inferences, see [11]. A more
radical option, and one which would appear to offer more hope, would be to
switch from a deductive to an abductive framework [12]. In any case, since
logic is so deeply engrained in our being, one may expect these reasoning
methods to be found at the neural level.
Figure
4. Another clue? Turn the page sideways..
At this point, we might ask whether recent developments in
neurobiology might not shed more light on the subject. As we indicated in the
introduction, there is now a wealth of minute detail regarding the functioning
of the brain and of the synapses in particular. Unfortunately, it is far from
clear how all this knowledge helps and how it fits into the overall picture. The
difficulty arises mainly because we have not yet managed to find an appropriate
mapping from the biological, let alone the molecular level, to the level of
knowledge and thoughts. An analogy may make the problem clearer. Suppose you
were given a piece of equipment constructed from transistors, together with a
detailed description of the workings of the transistors themselves. Would this
be sufficient for you to discover how the equipment works? Probably not, for
one thing understanding how each component works gives no hint as to what the
machine does, it could be a radio, a television, a washing machine or a
computer. Moreover, even knowing what behaviour the equipment is supposed to
exhibit, does not necessarily help understand its inner workings, for it may be
implemented in any number of different ways. You may not even be sure what the
basic components are, for example, are they the soldered joints, the integrated
circuits or the "transistors" inside them? Recent work in
neuroscience has suggested that the basic building blocks of the brain may, in
fact, be the synapses not the neuron [21]. Anyway, even if you can correctly
identify what counts as a basic component, knowing how it functions doesn't
help, for, unless you know what information is being 'processed', it is
impossible to explain the higher-level behaviour. Deciding what constitutes the
information and what its content is, is very difficult. Is the information
encoded in, for example, the voltage, the current, the pulse width, the
frequency, a digital code or the phase. And what is the information, is it
sound, picture, synchronisation, teletext, control or what? The apparent
mountain of molecular and chemical evidence, then, is relatively worthless
until it can be placed in an appropriate higher-level context.
What is of much more import, however, is the knowledge of
the structural characteristics of the brain which have been painstakingly
pieced together. The most detailed evidence relates to the visual system and
results partly from experiments on animals and partly from human patients
suffering from a variety of brain tumors. Interpreting this evidence, however,
is not always as straightforward as it might seem [13]. As an example, a
condition known as prosopagnosia, in which the subject is unable to recognise
faces, would seem to suggest the existence of a functional component dedicated
to face recognition. In fact, recent studies have shown that patient's deficits
are usually more extensive, so that such a conclusion is unwarranted.
Despite the pitfalls, we do have a reasonably good
understanding of how information from the eyes is processed within the visual
cortex. Surprisingly, signals for colour, dynamic form, motion, and forms with
colour, are all processed in parallel and along relatively independent paths
[14]. Figure 5 shows the basic organisation. Input from the retina arrives at
layer V1 and is then relayed via layer V2 to layers V3, V4 and V5. There are
also direct connections from cells in layer 4B of V1 to V3 and V5. Lesions in
each of these areas produce distinct pathologies. For example, patients with
lesions only in V4 view the world in shades of gray, while those with lesions
in V5 can see objects which are stationary, but not ones that are in motion.
Some patients suffering from severe carbon monoxide poisoning can end up having
only the very limited colour information which part of V1 still provides to
them, having lost most of V1, V2, V3, V4 and V5. Without any knowledge of form,
they are forced to guess what an object is, based solely on its colour and may,
for example, misidentify all blue objects as 'ocean'. Layer V1 is obviously
critical for vision and any damage may leave the patient either completely
blind or with an inability to comprehend even simple shapes. If V1 is still
intact, a patient may be able to reproduce a drawing, but, because of other
damage, be unable to comprehend what it is a picture of. Interestingly, in such
cases, the patient may well be able to name an object by touch or smell, but be
unable to identify it by sight.
Clearly, then, vision is a very complex process, only the
first stages of which are explained above. Additional processes are needed to
achieve real spatial understanding, for, when we view the world we see only a
very small part of it at any one instant. When reading, for example, we
obviously do not perceive the entire page at once, but rather move our eyes
from word to word. Dennett [15, p361] recounts an experiment in which a subject
reads text from a computer screen. However, by means of an eye tracking system,
the computer can determine on which word the eye will settle next and can
change that word before the eye actually reaches it. To onlookers the screen is
in a continuous state of change, to the subject, however, it appears completely
normal. Similarly then, when we look at someone's face, for example, we do not
immediately comprehend it in its entirety. Rather, we 'scan' it, seemingly at
random, picking out the location and form of its major components and somehow
piecing them together until we achieve recognition.
|
|
Figure
5. Organisation of the
visual cortex (from [14])
Another area which has received significant attention is that
of language. Language is a distinctly human phenomena, thus, until recently,
research has had to rely almost entirely on nature's experiments, in the form
of lesions, to help unravel its mysteries. The development of non-invasive
techniques has provided much needed confirmation of earlier theories. Magnetic
resonance imaging (MRI), for example, allows the exact location of lesions to
be determined. This technique has been able to show that specific dysfunctions
are always associated with the same specific regions of the brain. Positron
emission tomography( PET), on the other hand, has enabled the mapping, in
healthy individuals, of brain function to location. This enables us to see
normal brain activity while performing a variety of tasks (Figure 6.) These
dual sources of evidence have helped build up a reasonably good picture of the
various subsystems which cooperate to give us our linguistic abilities.
In essence language appears to involve three sets of
structures [16]. The first deals with non-language interaction between the body
and its environment, creating representations for form, colour actions, etc.
The second deals with the components of speech, phonemes, phoneme combinations
and rules of syntax. On the one hand, it allows sentences to be generated and
spoken or written, and, on the other, it performs the initial processing for
speech understanding. The third set of structures mediates between the other
two, either taking a concept and causing an appropriate word or sentence to be
generated, or receiving words and evoking the corresponding concepts.
|
|
Figure
6. PET scans showing
brain activity during various tasks (from [6])
Lesions in each of these structures display specific pathologies.
For example, patients may continue to experience colour normally but be unable
to name them correctly. Alternatively, they may produce phonemically distorted
names, such as 'buh' for 'blue'. In other cases, patients may substitute a more
general word for the missing one ('people' for 'woman') or use a semantically
related word ('headman' for 'president'). Besides these deficits, they speak in
a perfectly normal manner. Lesions in the anterior and midtemporal cortices
(which handle word selection and sentence generation) produce slightly
different symptoms. Sometimes they will result in an inability to recall common
nouns and proper nouns. Such a patient, would usually be unable to name
friends, relatives, celebrities or places. Shown a picture of Marilyn Monroe,
although they could not name her, they would most definitely recognise her and
be able to retrieve additional information, such as, that she was in the
movies, had a affair with the president, etc. While such patients speak
normally, they tend to substitute very general words like 'thing' or 'stuff,
'it' or 'she or 'they', for the missing nouns and pronouns. In contrast,
patients with left frontal damage have far more difficulty recalling verbs and
functors. Since these constitute the core of syntactic structure, it is not
surprising that such patients also have trouble producing grammatically correct
sentences.
These studies show that the brain has distinct regions for
the various components of language generation and understanding. The functional
separation of storage for phonemes, for proper and common nouns, for verbs and
syntax, offer significant clues as to the origin of our language abilities.
Still, it is very important to be cautious in our deductions. For example,
evidence has also shown that in some bilingual individuals, words for different
languages are found in distinct regions of the brain; however, we should
presumably not conclude that the brain has evolved a specialised component to
process each unique language! Another example concerns more general aspects of
memory. Psychology has long distinguished between so called short term, or
working memory, which acts as temporary storage during the recognition process,
and long term memory which retains known facts. The basic assumption is that we
first store 'things' in short term memory and only later, if they have not been
forgotten, are they transferred to long term memory for use on later occasions.
Recent work in neuroscience seems to provide confirmation for this idea, suggesting
that short-term memory may be located in the prefrontal lobes of the cerebral
cortex [17]. We should be wary of jumping to such a conclusion, however, for
both the original conception and the supporting evidence are based on an
information processing view of cognition, a view which may itself be suspect.
It is difficult, for example, to reconcile this conception with
connectionist-like neural storage representations.
Perhaps one of the most intriguing findings in recent times
concerns brain development [18]. It was originally thought that the brain's
wiring was genetically determined, however, new evidence shows that this is
only part of the story. We are born with all the neurons we will ever have,
approximately 100 billion of them, of which about 100 thousand die each day.
Although this may sound a lot it is comparatively small, for example, one would
have lost only about 2.5% of neural cells after 70 years. While the neurons
themselves do not really change, their interconnections most certainly do. Axon
and dendrite growth account for the large increase in brain weight following
birth. Recent research has shown that, while genes determine the general region
in which connections will be made, the final location depends on neural
activity. Lack of such activity can seriously impede development. Thus, infants
who spend most of their first year in cribs develop abnormally slowly, some
cannot sit up after 21 months and less than 15% are able to walk after 3 years.
In fact, axons and dendrites appear to be "plastic", growing and
shrinking continuously in response to neural excitation, throughout a person's
life. These observations seem to be ignored in current computational models of
brain function, but they may yet prove to have some deep significance in the
overall scheme of things.
Finally, the thorny question of mind: Do the sort of brain
features we have considered here provide a link, a stepping stone, to
explaining emotions, feelings, consciousness, awareness, intention, self?
Obviously, it is impossible to give a definite answer to this; yet, one by one
the barriers are falling. Not so long ago, most people would have thought that
language, in particular, was something uniquely human, yet slowly we are
unfolding a rational (non-magical) picture of its functioning. In fact, mind is
not so much a thing as a process or a set of abstract processes. Two brief
observations will serve to illustrate the apparent physical basis of these
processes. First is the well known fact that certain drugs affect one's mind.
They can cause or control depression, can stop pain and can simulate thought.
Secondly, are experiments which investigate intention. Researchers have found
neural activity just prior to and correlated with the initiation of some
action. In one experiment [19], subjects were asked to watch a slide show and
were given a button to push when they wished to view the next slide in the
sequence. The button, however, did not really control the slide projector, but
rather, electrodes attached to the subject detected their 'intention' to change
the slide and initiated the change before they had actually pressed the button.
Subjects reported that, just as they decided to press the button, they saw the
slide change, though, they were unable to stop their finger pressing the button
anyway.
The interplay between consciousness, awareness and intention
is still very much the realm of philosophy. Probably the clearest explanation
to-date has been offered by Dennett [15], who demolishes prevalent views of mind
wherein everything, every thought, must come to some central stage for
conscious consideration (what he calls the Cartesian Theater). According to
Dennett, not only is there no central stage, but there is no sharp dividing
line between conscious and unconscious thought.
4.
Concluding Remarks
The preceding sections both reviewed some of our current
ideas about the functioning of the brain and indicated why developing such
ideas was a particularly risky task. It should be obvious that there is a lot of
very interesting work going on and that there is certainly no lack of clues to
help us in our quest for understanding. Indeed the problem is not so much a
lack of evidence, as lack of a suitable framework into which to place it all
[20]. We seem to be in desperate need of a model, or hierarchy of models, which
can provide us with an overview of the entire cognitive process. Such a model,
if accurate, would act both to interpret the available evidence and to guide
the search for a deeper understanding. Only with such a model can we expect
progress in cognitive science to match that in biology, chemistry and physics.
Are there any models which might provide a suitable
framework? To the author's knowledge, there are no really good candidates. Both
the major computational paradigms fall short of our requirements. The symbolic
paradigm is undoubtedly able to model any aspect of cognition, but it fails to
offer any real insights. In essence, you get out what you put in, mainly
because symbolic models are simply unable to provide an implementation-level
account of cognition. In contrast, the connectionist paradigm, being founded at
just such a level might be thought to offer more hope. Indeed, the artificial
neural network is generally assumed to be a very good approximation to the real
thing, with connectionist models having demonstrated apparent solutions to many
problems. Unfortunately, this success may be illusory for the sorts of networks
used in these exemplar systems are often not as plausible as they may at first sight appear to be. For one thing,
they fail to account for many of the known facts, such as, synaptic outputs as
well as inputs, neuromodulation and second messengers [21]. Artificial neural
networks frequently employ biologically implausible mechanisms, such as back
propagation. Moreover, they have been criticised on theoretical grounds.
According to Fodor and Pylyshyn [22] they lack representational structure and
are based on the discredited Associationalist philosophy. While some of the
more recent work in this field has addressed the former problem (for a review
see [23]), the latter one remains. Indeed, some connectionists [24] have even
claimed that their paradigm can overcome the philosophical difficulties
supposedly inherent to associationalism, but this seems unlikely.
While all this may appear to leave us without
any viable model, there is at least one further possibility which seemingly
both explains the available evidence and is philosophically plausible. In
essence this model simply records "everything". Whenever it
"sees something", however, it first attempts to find a match in its
memory and records the results of this process rather than "raw"
inputs (for a fuller description of how this might work see [25]). The only
problematic aspect of this model concerns its biological implementation, which
is somewhat at odds with current thinking in neurobiology. Rather than storing
knowledge in the interconnection weights (synaptic strength), knowledge is
considered to be inherent in the pattern of connections which are formed either
during development or dynamically as we see and remember things. The fact that
this alternative model fails to coincide with present understanding in this
field, does not necessarily mean that the theory is wrong, of course. As we
observed previously, evidence is sort for and interpreted within the context of
an existing model. Thus, it is just possible that the clues might be open to a
different interpretation, one which actually supports this alternative model.
If we are to make progress in our quest for understanding,
we must remain open to, and indeed specifically search out, alternative
proposals. It is only through trial and error, through the building and testing
of models that we can gain deeper insights. The natural world offers us an
endless supply of clues to the inner secrets of the mind, but it is up to us to
select and interpret them, and to organise them into a meaningful framework.
References
1. Shepherd, G.M. (1988) Neurobiology, Oxford University Press.
2. Churchland, P.S. (1986) Neurophilosophy: Toward a Unified Science of the Mind-Brain, MIT
Press.
3. Lashley, K.S. (1950) "In search of the
engram", Symposia of the Society for
Experimental Biology, No.4, Physiological Mechanisms in Animal Behavior,
Academic Press, pp.454-483. Referenced in [4].
4. Feldman, J.A. (1990) "Computational
Constraints on Higher Neural Representations", in Schwartz E.L. (ed), Computational Neuroscience, MIT Press.
5. Thompson, R.F., (1986) "The
neurobiology of Learning and Memory", Science
233, pp.941-947. Referenced in [4].
6. Fischbach, G.D. (1993) "Mind and
Brain", in Mind and Brain: Readings
from Scientific American, W.H. Freeman, pp.2-14
7. Glass, A.L. & Holyoak, K.J. (1986) Cognition 2nd edition, Random House
p.27.
8. Neely, J.H. (1977) "Semantic priming
and retrieval from lexical memory: Role of inhibitionless spreading activation
and limited capacity attention", Journal
of Experimental Psychology: General, 106, pp.226-254. Referenced in Glass
A.L. & Holyoak K.J. (1986) Cognition
2nd edition, Random House p.63.
9. Leeper, R. (1935) "A Study of a
neglected portion of the field of learning - the development of sensory
organization", Journal of Genetic
Psychology, 46, pp.41-75. Referenced in Glass A.L. & Holyoak K.J. (1986)
Cognition 2nd edition, Random House,
p.127. (Solution to Fig.2: A boy with a dog).
10. Dennett, D.C. (1971) "Intensional
Systems" in Journal of Philosophy, LXVIII, 4, pp.87-106, reprinted in
Dennett, D.C. (1986) Brainstorms:
Philosophical Essays on Mind and Psychology, Harvester Press Ltd. UK.
11. Byrne, R.M.J. (1989) "Suppressing valid
inferences with Conditionals", Cognition
31, pp.61-83. (For possible refutation
see Politzer G. & Braine M.D.S. (1991) "Responses to inconsistent
premisses cannot count as suppression of valid inferences", Cognition 38, pp.103-108 )
12. Davenport, D. (1993) "Intelligent
Systems: The Weakest Link?" in Kaynak O., Honderd G. & Grant E. (eds),
Intelligent Systems: Safety, Reliability
and Maintainability Issues, NATO ASI Series F, Vol.114, Springer-Verlag,
pp.60-73.
13. Grobstein, P. (1990) "Strategies for
Analyzing Complex Organization in the Nervous System: I. Lesion
Experiments", in Schwartz E.L. (ed), Computational
Neuroscience, MIT Press.
14. Zeki, S. (1993) "The Visual Image in Mind
and Brain", in Mind and Brain:
Readings from Scientific American, W.H. Freeman, p.29- 39.
15. Dennett, D. (1991) Consciousness Explained, Little & Brown Pub.
16. Damasio, A.R. & Damasio, H. (1993)
"Brain and Language", in Mind
and Brain: Readings from Scientific American, W.H. Freeman, pp.54-65.
17. Goldman-Rakic, P. (1993) "Working Memory
and the Mind", in Mind and Brain:
Readings from Scientific American, W.H. Freeman, pp.67- 77
18. Shatz, C. (1993) "The Developing
Brain", in Mind and Brain: Readings
from Scientific American, W.H. Freeman, pp.15-26.
19. Grey Walter's Precognitive Carousel (1963).
Referenced in [15] p.167.
20. Kosslyn, S.M. & van Kleeck, M. (1990)
"Broken Brains and Normal Minds: Why Humpty Dumpty Needs a Skeleton",
in Schwartz E.L. (ed), Computational
Neuroscience, MIT Press.
21. Shepherd, G.M. (1990) "The Significance
of Real Neuron Architectures for Neural Network Simulations", in Schwartz
E.L. (ed), Computational Neuroscience,
MIT Press.
22. Fodor J. & Pylyshyn Z. (1989) "Connectionism
and cognitive architectures: A critical analysis", in Pinker S. &
Mehler J. (eds), Connections and Symbols
(A special issue of the journal Cognition), Bradford Books, MIT Press.
23. Bletchley W. (1993) "Currents in
Connectionism", in Minds and Machines
3, pp.125-153.
24. Bechtel, W. & Abrahamsen, A.(1991) Connectionism and the Mind, Basil
Blackwell Pub.
25. Davenport
D. (1993) "Inscriptors: Knowledge Representation for Cognition", in Proc. 8th Int. Symp. on Computer and
Information Sciences, Istanbul.