Spring and Fall Fashions in
Cognitive Science[1]
Zenon Pylyshyn
Rutgers Center For Cognitive Science
Rutgers University
This is indeed an auspicious
time for Cognitive Science. I stand here before you this evening as the first Chair
to give a presidential address to this austere body, to place on record before
you what you are to accept as the Society's official view on the new science of
the mind. This is a particularly
important period in the history of the Cognitive Science Society. It is the
eighth anniversary of the founding of this society. Many of you may know that
the society was founded eight years ago in a historic hotel room at the Kansas
City Airport, which I believe has now been converted into a national historical
monument. It was there that we founded
this society and announced to the world the dawning of a new science, a science
that would soon unravel the mysteries of the mind and reveal the secrets of
human nature. A science that might one day even be turned upon itself to answer
the question; Why did 12 reasonably intelligent people, backed by the boundless
resources of the Alfred P. Sloan Foundation, choose to travel thousands of
miles to meet at the Kansas City Airport?
This eighth anniversary is a particularly important one. Although other
societies frequently pay more attention to the tenth, twentieth, and so on
anniversaries, we -- of all people -- understand that this is a chauvinistic
view that arises from the mere contingent fact that humans happen to have ten
fingers -- more or less.
Dedicated as we are to transcending the merely anthropocentric viewpoint
in our search for Cognitive universals, natural constraints and timeless
invariants, we are well aware that in an absolute notation -- such as one based
on the prime factor decomposition -- eight is a much more important
number, since it decomposes into three multiples of the lowest prime know to
Cognitive Science. This, then, is the day we celebrate our first triple-prime
anniversary. At that first meeting
in, as I said before -- the Kansas City Airport, an attempt was made to
represent all potentially relevant constituencies. Like Noah's are we packed
the airport hotel room with pairs of linguists, psychologists, computer
scientists, philosophers, and even a lone anthropologist (I leave it to you to
speculate on the deeper implications of there being only one anthropologist in
this arc). This ecumenism, I believe, was one of our first fateful mistakes,
since it started the field off in a direction of permissiveness and
libertarianism from which it may never recover. One consequence is that an unusually large proportion the
membership of the society is dedicated to promoting the view that the field
which the society represents does not really exist. As I find myself very often
rubbing shoulders at conferences with many of these detractors, I sometimes
feel like one of the ministers in this New Yorker cartoon.
Now, on its eighth anniversary, the Cognitive Science Society -- if not
the field has matured to the point where it can offer its chairman a serious
recompense for using his name on its letterhead. The reward is that the
chairman is offered a captive and appreciative audience which is committed to
sit still for an hour or so while the chairman holds forth on matters that
would normally be considered too ill mannered to present in public. Needless to
say, this talk has not been refereed, although I have checked my facts over
several times for sincerity.
Until a few years ago the way to earn your credentials in cognitive
science -- the way to publish in one of the mainstream cognition journals --
was simply to run a reaction time study that yielded a straight line -- any
straight line would do, it didn't matter what the independent variable was. Now
I must confess that I have not been notably successful in meeting this
challenge, although I have generated a great deal of very reliable data -- a
fragment of which I will now show you to prove that I am no mere armchair
cognitive scientist.
As you can see this slide shows that I was able to measure reaction
time for a group of subjects while manipulating various independent variables
(it doesn't matter which, the results remain pretty much the same). I know that
it might seem to some of you who are not experienced in such matters that these
data are not clean and precise. However, much depends on how you look at it.
For example, in the next slide I show you that if you look closely at the data,
they turn out to be extremely clean and precise.
Of course it's only when you do serious statistical analyses of the raw
data on a powerful computer that you really see the significance of this work.
We have over the years replicated these results in our laboratory over and over
again, and we have been able to check them against the unpublished results of
other investigators, so that now we are morally certain of the conclusion to
which they lead us. And that conclusion is this: for normal populations of
college undergraduates, when you carefully control for experimental artifacts
the null hypothesis is always empirically true.
In case you are one of those of people who managed to get a straight
line reaction- time result, let me assure you that I am not questioning your
honesty. There have literally been
thousands of reaction time experiments with linear outcomes. For example, there
have been hundreds of variations on the Sternberg rapid memory scan experiment,
hundreds more on mentally rotating images, mentally comparing magnitudes, and
scanning mental images, as well as visually scanning displays for various
features. All have been done with the utmost integrity. Indeed, notwithstanding
the results I showed you earlier, I have to admit that once even I obtained a
reliable linear reaction-time function that I will tell you about. This result
illustrates what I suspect is behind many of the linear results obtained in the
literature -- it's certainly what's behind all the mental image scanning
results.
What I did was this. Following some work carried out in France about
thirty years ago by Fraise, I asked subjects to generate a reaction time
proportional to various magnitudes I showed them or told them about -- such as
length of lines, size of numbers, and so on.
I found that subjects could do this very well. They readily generated reaction-times that were a linear
function of some magnitude or other that they had in mind. This demonstrates clearly that subjects can
produce beautifully linear reaction time functions if you just explain to them
clearly that this is what you want and if you ask them nicely and politely to
do so. And conversely, as many of you who have heard me talk about certain
mental imagery experiments know, we can also make many linear reaction-time
functions go away by using this same clever procedure of persuading subjects
that there should be no linear relation.
It seems that in their zeal to be laboratory experimentalists,
psychologists have forgotten that, within fairly wide margins, subjects can do
whatever they like. This, by the way, is not unrelated to an even deeper
message about cognition I shall deliver later in this talk.
Incidentally, psychologists are not the only cognitive scientists who
fall victim to the cleverness of their subjects. Consider the plight of the
poor anthropologists:
Methodological problems of anthropologists
I don't
want to leave you with the impression that finding linear reaction-time
functions has been the only way of discovering truths about the mind. To
illustrate the wide range of methods that have been developed for investigating
cognition, I will mention a few randomly chosen examples of some classical
methods that have been popular in certain quarters of cognitive science. For
instance, there was the form of investigation shown in the next slide, which
begins with a set of sentences, some of which have asterisks in front of them,
and which ends with the conclusion that language is innate. One might call this
method Chomsking to conclusions.
Linguistic argument for nativism
As another example, in the area of visual perception the old method of
becoming famous by discovering an illusion still continues. But real fame and
fortune is gained by those who produce a picture that people can use in order
to argue their favorite theory. For example, take the old top-down versus
bottom-up argument. This cute teddy bear: Marr’s Teddy bear has inspired-many people to view vision as being a
bottom up process which begins with the detection of zero-crossings in the
second derivative of the intensity function, whereas this equally attractive Dalmatian is held out by others as showing that vision
must be top-down. The debate over the direction of perceptual processing is
another of those arguments that goes back and forth in a cyclical manner. No
sooner does one side demonstrate a strong top down effect (in which the
conceptual "big picture" determines what is seen, as metaphorically
shown in this slide top-down) than the other
group shows that this process can be undermined by bottom-up data-driven
processes, as metaphorically shown in this next slide bottom-up.
Yet another example comes from the study of problem solving, an area that
has made great strides in the past 15 years. The growth of this area is due in
part to the work in Artificial Intelligence, which has provided much of the
theoretical framework. But in part it is also due to the invention of a
devilishly clever methodology for discovering how a subject solves a problem.
This technique consists of asking the subject. It is best if you do it while
the subject is solving the problem, because then he tends to make a lot of
mistakes which require a cognitive scientist to explain (that's because, you
see, when the subject gets it right this is the province of Al). This method is called protocol analysis and
is very important now that the business world has discovered "expert
systems", which are computer programs that make some of the same mistakes
as people, only faster. This obviously requires a careful analysis of the
expert's mistakes.
+++++++++++++++
Over the past decade Cognitive Science has made considerable progress
not only in methodology, but also in its general approach to research and in
the way it selects worthwhile research projects. We have all come a long way
from the old days when we spent our time searching for straight-line reaction
time functions or trying to demonstrate that perception was top-down or bottom-up.
Today we are beginning to realize the deeper significance and far-reaching
implications of our research. We see that research such as that which I
discussed earlier can be used not only to argue for nativism or modularity, but
also to argue against American foreign policy or against the Unix operating
system. We have gone beyond the laboratory and are reaching out towards some of
that loose Al venture capital by recognizing that our research is contributing
to an understanding of the "user interface" and what makes it
"user-friendly", or to improving the lot of
"knowledge-engineers" by helping them enhance their
"knowledge-acquisition" procedures. No longer is the ambition of new
graduate students limited to the hope that they will successfully complete 3
experiments and a 5 page LISP program -- that well-known formula for getting
published in Cognitive Psychology.
Today they are out there interviewing experts in every walk of life to
discover their "knowledge-structures", or are trying to teach 3 year
olds to program in LISP, because they know that at the end of that road
lies user-friendly private enterprise: Not only Teknowledge, but
Tekmoney!
But I'm not being fair. Of course, not everyone is tempted by such
pecuniary motives. Many people just want to be invited to conferences in the
south of France. Others have purely intellectual motives. For example, some would like to show that
cognition can be studied in the same esoteric macho manner as solid state
physics. That's why we are seeing such a flurry of exotic mathematics recently.
Math is back in as a symbol of virility -- as I believe Gary Larsen was first
to recognize.
Technical tools have always been important in the development of
science, and I do not wish to minimize their role. But sometimes the
application is premature: either the tools are not ready, or the field is not
ready to use them, as was the case, I understand, with the early microscope, as
shown in this slide:
It is also true in cognitive science, as it has been in all other areas
of human endeavour, that a very important reason for doing brilliant work is to
show up the great people who have gone before you -- the Chomskys or Minskys or
Newells and Simons. Of course the real big-fish target these days is John Von
Neuman. Sociologically-speaking, the
great break-through of the past decade has been the ganging up on Von
Neuman -- or at least on the Von Neuman computer. Nobody knows exactly what makes a computer Von Neumanish, but
there is general agreement that whatever it is, it is a bad thing. What is a good thing nowadays is
to be parallel. Not just plain ordinary multi-processor parallel, but massively
parallel. Once again nobody quite
knows what that is, and whether it is to be contrasted with, say, being massively
serial, but we do know that it is a good thing. Just why it is generally thought to be a
good thing is a topic to which I will return shortly. But first I need
to back away from such specifics in order to exercise my chairman's prerogative
to be profound. I begin, therefore,
with some background profundities.
++++++++++++++
In its formative years a discipline is very sensitive to the implicit
foundational and philosophical assumptions of its practitioners. Moreover, it
is also very susceptible to the whims of fashion and to the ever-changing
styles of theory and methodology. Despite the proliferation of fashions I
believe there are only a small number of fundamentally different options
available. What I would like to do in the remainder of this brief time I have
been awarded, is to tell you what I think are the main options and to try to
set out where some of us stand on these options and why.
Distinctions
are very important in an evolving science. Whether or not you start off with
the assumption that natural versus violent motion is a fundamental distinction
makes a big difference to your research program, as does the precise way you
distinguish among weight, mass, and inertia. The reason that such distinctions
are important is that you can only have a uniform scientific discipline of X if
X is a natural domain -- if it includes phenomena that fall under some
reasonably uniform set of principles.
The distinctions that are of fundamental importance to Cognitive
Science are not, in my opinion, those between serial and parallel processing,
or between continuous and discrete operations, Von Neuman, Non-Von Neuman or
production system architectures. These are all real enough distinctions, and
which side wins out will make a difference, inasmuch as one side will be right
and the other will be wrong. But these are all local differences: it will
surely turn out that some processes are serial and some parallel, some
continuous and some discrete, some Von Neumanish and some not, etc. Each
particular case will be decided empirically on its merit.
But what really is fundamental is a distinction with which scholars
have had a love- hate relation ever since they first thought about cognition.
It is a distinction that has been studied and debated by nearly a century of
philosophers, has been made scientifically respectable by computer science, and
yet continues to arouse strong emotional reactions in almost everyone.
Concerted opposition to it has spawned several influential schools of
psychology, including behaviorism and Gibson's "direct realism". More
recently it has been responsible for the growth in popularity of the so-called
"new connectionism". It
is a distinction which is denied daily
by a growing number of otherwise reasonable people, and yet a distinction which
-- in their heart -- everyone accepts. If this distinction turns out to
be defensible -- and indeed if it turns out to be indispensable, as I believe
it is -- it will mean that large parts of cognition will require explanations
that differ in a fundamental way from explanations in other natural sciences.
Having set myself up in this way, I now offer my official diagnosis of
"what ails cognitive science" in terms of this distinction.
The
distinction I have in mind is between two ways of explaining an observed
regularity in the behavior of some system. One way is by appealing to
properties or mechanisms that are intrinsic to the system, in other words, by appealing
to certain functional properties or capacities of the system (what many of us
have called its functional architecture).
The other is by appealing to the existence of particular representations
which the system manipulates according to certain rules (for example, rules
of inference).
I will not discuss the
distinction in any detail today, partly because I want to get to the point
about why it's a problem for recent approaches to Cognitive Science, and partly
because then you won't feel any need to buy my book which is on sale out in the
hall at the MIT table. It's only $9.95 for the paperback version, with a
substantial discount if you promise to believe what it says. I just want to point out, using a few
simple examples, that the idea of understanding certain behaviors in terms of
goals and beliefs, and the semantics of stimulus events, is so widely accepted
that even the most doctrinaire behaviorist believes it. Why else would he
solicit subjects by writing sentences on notices -- sentences which subjects
have never before seen and which assert such propositions as that volunteers
will be paid for participating. And of course the behaviorist implicitly
assumes that there will be a rational connection between the content of the
experimental instructions, the subject's desire to do as he is told, and the
subject's behavior in the experiment.
It would also be absurd to
deny, say, that part of the explanation for why I am here making these
particular noises has something to do with the fact that Chuck Clifton invited
me and promised that my food and lodgings would be provided, and has something
to do with my desire to promulgate a certain view that I hold about Cognitive
Science. Even if you don't believe the
view I hold about Cognitive Science, you doubtlessly believe that it is correct
to say that I believe it and that I am at the moment trying to persuade you of
it! Moreover, not only is there a rational connection between such things as
invitations, my beliefs, etc and my behavior, but the particular behavior I
exhibited could have been quite different if I had interpreted the antecedent
events differently. For example, my behavior would have been different -- yet
still rationally connected to my beliefs -- if Jim Moyer, the local organizer,
had phoned me after the invitation had been issued and had said that Chuck had
flipped out from overwork and was inviting everybody he could get hold of to
give the after dinner speech, and moreover the entire budget of the society had
already been spent. The regularity
connecting stimulus events with behavior is subject to systematic change by
collateral information. The proprietary
term for this sort of plasticity is Cognitive Penetrability and you may hear me
use that phrase a lot.
Although I have been talking
very roughly about goals and beliefs, it still takes a lot of nerve to deny the
relevance of such factors in determining behavioral regularities. Usually people only deny such obvious
truisms when they are asked to give professional philosophical opinions, so it
is surprising to find this kind of talk in a conference like this one.
Let
me add just another example of when such an explanation is called for so that I
can have a less facetious example to refer back to later. The question is, How
do people know what the underlined pronoun refers to in each of the following
sentences (based on an example due to Terry Winograd).
People readily understand
these sentences, which entails that they assign a reference to the pronoun in
each case. Surely, the explanation for how the reference is assigned must
mention what the listener knows about city councilors, workers, demonstrations,
communists, and so on. Only factors like this would explain why in particular
cases the pronouns are assigned different referents in the two sentences and
why the reference assignment could be easily changed by altering the
information available to the listener. For example, a few years ago I
inadvertently provided an example of the effect of collateral knowledge when I
used these sentences in a talk I gave in Florence, Italy. I had forgotten that
the city council of Florence was in fact communist. Because of this my audience
had assigned the same referent to the pronoun in the first two sentences (which
were the only ones I used in that talk) and the point of the example had been
lost on them!
Notice that the point here is not merely that the process of
pronomial reference- assignment is context-dependent. The point is that it is
dependent in a rational way on the semantic information provided by the
context. Moreover, what constitutes a relevant context can reach back into the
indefinite past and become relevant through arbitrarily long chains of
inference.[2] It does not matter how you come to
believe that the city councilors in question are communists -- if you are
paranoid enough you might infer it from the most tenuous evidence, as did
Senator McCarthy. Yet no matter how you reached that belief, your
pronoun-assignment process would systematically and rationally take it into
account.
+++++++++++++++++
OK now here comes one of the most important morals of this story. This
is the hard part, so make sure you are listening carefully. Because of the sorts of considerations sketched
here, cognitive science will not be able to get away solely with
theories that tell us how the brain is wired or how activation spreads or how
states of equilibrium are reached. That's because all such theories -- of
necessity -- treat the cognitive system as being in causal contact with a
environment. They cannot treat
environments as the organism interprets them because interpreting is a
relation that requires both environmental stimulation and reasoning from
beliefs, goals, and expectations. This, in turn, requires inferential processes
which such systems cannot support -- for reasons to which I will return in a
moment. Because of this confinement to causal relations, such systems respond
to physical properties of the environment (since these are the only ones that
enter into causal laws). And that has
been the Achilles' heal of every naturalistically-motivated school of
psychology since Pavlov -- or perhaps Socrates (everything seems to go back to Socrates).
The problem is that the relevant relation between a cognizing organism
and its environment is not a physical-causal one, but an informational, or
to be more precise, an interpretive (or what some philosophers call an
intentional) one. If you try to relate intelligent behavior to physically
described environments you will come to the conclusion that, except for
tripping, failing, bouncing off walls, and so on, behavior is largely
independent of the environment -- which is clearly false. That was the elegant
message of Chomsky's critique of Skinner way back in 1959. The plain fact is
that people do things for such reasons as that they wish to find the holy
grail, they wish to win someone's love, or they wish to get tenure. Yet neither
the grail nor the presumably nonexistent love or tenure cause the
behavior, only a representation of them can enter into the determination
of the behavior. We reach for a cup of coffee not because of the coffee in the
cup, for there
may be no coffee in it after all, but because we believe there to be coffee,
or because we see the black stuff there as coffee.
Once you accept this (surely obvious) story you have to face a further
problem which, until recently, was thought to be insurmountable by a
materialistic theory. That is the question of how a physical system like a mind
can behave in ways that are coherently described in terms of goals and beliefs
-- i.e. in terms of things that refer to aspects of the world that are not
physical inasmuch as they are the world-as-believed (indeed, as in the previous
examples, the relevant aspects may not even exist). Now, within the past half
century, we have had the first serious proposal for how this might be possible.
That proposal started with the formalist movement in mathematics and logic
early this century, but the version that is of particular relevance to
cognitive science is what Al Newell calls the Physical Symbol System
Hypothesis. The proposal is that meanings are encoded as symbolic
expressions and inferences are carried out over them by symbol manipulation.
The critical fact is that nobody has the slightest idea how inference
can be done without such a process.
You might say that nobody has a notion of a nonsymbolic reasoning
process. Moreover, there are some good arguments why you will never have a
reasoning system that fails to meet certain conditions. Among those conditions
is the requirement that the states of such a reasoning system must have a
componential and combinatorial structure. In other words, the states must have
functionally distinguishable parts that can combine in novel ways, just the way
symbolic expressions do. Furthermore, the component parts must be replicable --
that is, the system must be able to have and to functionally distinguish
repeated tokens of the same substate type. That's because it is constitutive of
reasoning that you be able to think different things about the same objects of
thought. If an intelligent system can think that apples are red it must also
have the capacity to think other things about apples (minimally that there are
such things as apples). For example, if it can think that apples are red and
that sugar is sweet, it must have the capability of thinking that apples are
sweet (whether or not it does, of course, is another matter).
Another reason you can't get away with a system that has one brain
state for apples- are-red, another f or apples-grow-on-trees, a
third f or apples-are-a-type-of-fruit, a fourth for fruits are sweet and
so on, is that such a system lacks the intrinsic capacity to draw
inferences to new beliefs. It could not, for example, infer that there are such
things as apples, let alone that apples are sweet. No matter how much
additional complexity is built into such a system it cannot infer new beliefs
of this sort because it does not have access to the component parts of the
fused brain-states that would be required to trigger the appropriate
state-transition. It's the apple part of the apples-are-a-type-of-fruit
brain state, together with the recognition that the fruit part also
occurs as part of the fruits-are- sweet brain-state, that allows the
inference process to take place. Furthermore, without the componential
structure you could not add new information about apples, such as that apples
are edible, in such a way that it would lead to rationally connected behaviors
-- for example, so that it would lead the system to cat those red fruity things
that grow on trees, or to look in trees when it got hungry, or to answer such
questions as "What are those red fruity things that grow in trees?"
And finally, as I argued many years ago in my critique of mental
imagery theories, one of the nice properties that symbolic systems have is the
ability to be indefinite or vague in certain well-defined ways. Unlike the representational states of
fused-state systems (or systems whose "parts" do not correspond to
the semantically interpreted aspects that occur in the knowledge-level
description), the symbolic systems can represent such indefinite states of
affairs as John is married to either Mary or Helen in such a way that
the indefiniteness will interact properly with additional information (e.g. the
new information that John is not married to Mary).
Notice that all these conditions are trivially met by a system that can
write and read symbolic expressions. So far as I can tell, if a system has the
properties of componential structure and replicability of substates, it will be
essentially a physical symbol system -- though it's a bit early to tell if
that's all it needs. So if you accept that people's relation to their
environment is an informational one involving the sort of (typically
unconscious) reasoning that was illustrated in the pronoun reference example,
you are pretty much stuck with symbol processing.[3]
It's the only game in town -- or, as Fodor puts it "the only straw
afloat".
So finally,... if you have been paying close attention, you will find
yourself drawing the inevitable conclusion that in order to understand
cognitive processes one needs to understand at least two different kinds of system
organizations.
1. One needs theories of the mechanisms that support knowledge-based or
rule-based reasoning. This includes
theories of those parts of the perceptual and motor systems that are
themselves noninferential and
are therefore cognitively impenetrable. It also includes theories of the
functional architecture -- that part of the system which allows rules and
representations to be encoded and accessed. There is no doubt that this level
of organization has a lot of parallelism -- maybe even enough to merit the
phrase "massively parallel" -- after all, at the level at which the
architecture is implemented (i.e. at the register transfer level) the VAX is
"massively parallel"!
2. One needs theories of the reasoning competence of intelligent
systems. These theories could turn
out to be similar to the normative logical systems around in philosophic logic,
although the evidence to date (mostly from Al) suggests that current logics are not expressive enough and
current inference systems are not powerful enough to deal with common-sense
reasoning.
Before letting you off the hook so you can get to the reception I want
to add a few caveats and remarks so you will see how completely reasonable this
view really is.
> What about early vision,
etc?
First thing is to recognize that although a great deal of mental
activity is of this knowledge-based sort, not all of it is. Moreover, as we are
discovering, quite a lot of what was thought to be reasoning turns out not to be.
For example, contrary to the teachings of the "new look" movement in
perception, a lot of perception is not knowledge-based (e.g. what is called
early vision, which appears to include everything up to the construction
of the depth map or 2.5-d sketch). Moreover, although the evidence is not all
in, I would not be the least surprised to find a lot of perceptual and motor
learning to be in the non- representational category, as well as such aspects
of language comprehension as lexical lookup and maybe even morphology -- as Jay
McClelland proposed in his discussion this morning. Memory retrieval is surely
another process that has a significant architectural component in humans as it
does in electronic computers, and it almost certainly involves a quite different
mechanism from that of retrieval by address. And finally, I suspect that the
explanation of much of concept-learning and such phenomena as the effect of
moods, emotions and many psychopathologies will involve a major component that
is not knowledge-based. Consequently, I have high hopes that the PDP, Boltzman,
and new connection people can make significant inroads in these areas. But not,
as I have already asserted, where inference is clearly involved, unless these
systems are overlayed with another level of organization which supports symbol
manipulation functions.
> Can't everything be
explained this way?
Belief -desire explanations can be given for almost any phonenomenon:
rivers flow to the sea because they like the salt water and believe they can
get there by following a downward path. Thinking in such belief-desire terms
can even get one into unnecessary difficulties. Consider the thermos bottle. It
keeps warm liquids warm and cold liquids cold. A remarkable feat because how
does it know? The answer to this challenge is that (1) There is no
guarantee that we are not making a mistake in a particular case, (2) De
Morgan's Canon tells us that we should never postulate a higher function when a
lower one will do, (3) The sina qua non of a truly knowledge-based process is
that it is logical labile -- i.e., it is in principle cognitively penetrable®,
and (4) if you think -- as behaviorists did when they accused Tolman and other
early cognitivists of giving vacuous explanations -- that it’s an easy matter
to throw together a set of beliefs and goals to account for any piece of
behavior, then you should try writing an intelligent expert system in even a
narrow domain.
It's true that almost any intelligent behavior can in principle be
explained in terms of goals and beliefs. But almost any chemical reactions can
also be explained by the hypothesis that there are many different kinds of
molecules that interact in certain ways. Neither of these explanations need be
vacuous because there are a large number of constraints that must be
simultaneously satisfied for the explanation to work. In both cases the various
details have to be independently motivated and validated.
> What about chicken-sexing?
One of the things that bothers people about what I'll call the cognitivist
line is that it postulates a whole lot of inferences, reasoning, and problem
solving and other sorts of ratiocination going on where introspection reveals
nothing. When we introspect, for anything but slow and deliberate
problem-solving we are convinced that the answer just emerges from some general
disequilibrium followed by a leap of intuition. Even when we do have some
subjective experiences during problem solving these often do not reveal the
sort of activity that would qualify as "reasoning." Thus, Rudolph Arnheim attacks the Evans
geometry- analogies system on the grounds that when people do these problems
they go through a sequence of states involving a "rich and dazzling
experience" of "instability" of "fleeting, elusive
resemblances" and so on, whereas Evans' program doggedly pursues a search
involving constructing pattern descriptions and description-differences.
Similarly Burt Dreyfus accuses cognitive science of ignoring the difference
between fringe and focal consciousness, and Steve Kosslyn accuses me and others
of not taking seriously subjects' introspective reports of what goes on during
episodes of reasoning using imagery. But one can no more build a theory of
cognitive processing on the evidence of introspection than one can build a
theory of solid state matter on the evidence of the appearance of hard objects.
A theory of visual perception built on such evidence would be like this graphic
Kliban theory of what goes on in the Cat's mind during cat perception (which
may, incidentally, even be true in the case of the cat).
The trouble is that our introspections are not observations of
cognitive processes. They are reports of the content of our thoughts, or
of what our thoughts are about. They can no more be taken as reports of
mental structures or processes than they can be taken as reports of what
neurons are doing. Despite
this rather obvious criticism of introspective reports, I'm sure that the
discrepancy between our experience during certain cognitive episodes and what
cognitivist theories posit to be going on, is behind a lot of the discomfort
that people feel, and is at least part of the attraction of analog and other
nonsymbolic theories. But in this case it is the discomfort that will
have to yield in the face of empirically unavoidable conclusions, such as that
the process of assigning a referent to a pronoun involves a considerable amount
of unconscious inference. Eventually, people will feel at home with such ideas,
the way they got used to equally unintuitive physical theories.
By the way it turns out over and over that when you look closely at
such skills as chess or chicken-sexing (both of which Dreyfus sites as examples
of intuitive, as opposed to reasoning, processes) you find that there is a
large component of reasoning involved (as Irving Beiderman has shown in the
case of chicken-sexing).
> Isn't this way of talking
just a rough approximation?
Sure human behavior is not all rational or even coherent. When you
examine it closely you find that people consistently make certain kinds of
logical errors (such as those investigated by Tversky and Kahneman) and there
are exceptions to any rule- system. In that sense any particular symbol-system
model is an idealization. And. as with all idealizations, there is always the
question of whether the boundaries have been properly drawn.
Now it seems to me that we have two options in dealing with the
"approximate" nature of rule governed theories. The first is to say
that the divergence from pure rule- following that we observe in human behavior
arises out of the interaction of the rules and representations with properties
of the mechanism or architecture. These
impose resource constraints and unreliability on both the representations (e.g.
some of the beliefs may be false) and the processes (e.g. some valid inferences
may not be drawn). This is the traditional way of dealing with idealization in
science; empirical observations are
assumed to be the result of an interaction of the ideal with other factors.
The second option is to deny the idealization entirely. But then you
are faced not only with the problem of providing an explanation of the cases
that deviate from the ideal, but also of providing an alternative explanation
for the cases that would have been clearly covered by the idealized theory,
such as in the examples I gave earlier (e.g. the explanation of why I am here
or of how the referent to the pronoun is assigned). In such cases there seems
to be no way of escaping the view
that interpretation, inference, and decision-making are all relevant operations. And as I said earlier, there is
at present no way to deal with such processes other than by a symbol processing theory. Of course, in the
future someone might come up with an entirely new conceptualization of
the whole process that unifies the clear and the deviant cases, but at the
moment this is no more than an off-the-wall pious hope, bred of frustration with
the admittedly limited success of building grand theories of intelligent
behavior. But when you are thinking about where to place your bets you should
took at what it is that draws you to the various options and ask whether there
is any basis for optimism in the nonsymbolic camp. My own view is that,
although some of these approaches are promising as ways of suggesting new
mechanisms, their record in dealing with reasoning is at best nonexistent.
Moreover, judging by the nonsymbolic models that have been proposed for such
processes as mental imagery, the trend is in the direction of Rube Goldberg
style ad hocery. There comes a time when one should reconsider the
hairy-network theory of the mind -- as this spider has so wisely done with his
network:
> Might not the "explicit reasoning" form of processing be confined to a negligibly small part of the system's activity?
The question of how much of what we now think of as cognition
involves reasoning is not a fundamental question that threatens the whole
representational approach. As I said earlier, it may well be that a lot of what
we pretheoretically think of as cognition will be attributed to fancy memory
storage and retrieval mechanisms, elaborate sensory systems, or pararneter-setting
learning mechanisms, rather than symbolic reasoning. These are empirical
questions and we shall have to wait and see how it turns out. But to those who
think that perhaps the only thing that will be left for representations to do
is deliberate linguistically-mediated puzzle-solving I have two warnings.
First, the evidence from information-processing studies is against you. Careful studies regularly reveal
problem-solving processes taking place where people thought there was only
intuition (e.g. studies of subitizing). Second, if you allow yourself a residue
of reasoning somewhere in the system, there is always the danger that this is
the part that is doing all the work. Extreme examples of this are the cognitive
map theories (and some current holographic theories) that need a homunculus to
make them act. My favorite contemporary example of this is the model of imagery
that posits an internal display which is examined by the "mind's
eye." I have argued that you can account for any possible result using
such a system, providing you allow the "mind's eye" to act as an
executive which examines the display and sets such parameters as the decay
time, the rotation rate, the strategy for scanning the display and so on. But
in that case it turns out that you don't need the display anymore since all the
empirical phenomena are being accounted for by the executive. The point is that
once you allow some reasoning into your nonsymbolic system you have to
watch that the nonsymbolic part is not redundant. The paradigm example of this
trap is the following story I first heard from Hilary Putnam.
The story is about an
engineer who claimed for years that he had invented a perpetual-motion machine.
Of course nobody believed him, but because he was a nice person they continued
to humour him. Finally his bragging became too much and they decided to take
him up on his offer to demonstrate the machine. The engineer led a group of his
friends down to his basement, which was filled with a wondrous array of gears,
levers, cogs, belts, and other visual marvels that glittered under the lights.
His friends were much impressed by his workmanship. But one of them said to the
engineer, "This is all very ingenious and wonderful, but there is one
thing that bothers me. I notice that nothing is moving." Oh that, replied
the engineer. That's nothing serious. It's just that I'm missing one part which
is back-ordered. It’s a little lever
that fits here and goes back and forth like this for ever."
There is the ever-present danger that the little bit of remaining
reasoning capability we will need to allow in our nonsymbolic system is just
this hook.
> What do I think of Cognitive Science?
Finally, my conclusion. What do I think of Cognitive Science, I heard
you ask (didn't you?). I have always found psychology depressing because I came
into it from physics and engineering thinking that, since it experimentally
studied the human mind it was a science. I soon realized that it was not a
science but a catalog, and a methodology for adding to the catalog. I don't
doubt that it is a useful catalog: it's certainly important to know such things
as how to help people who are depressed or to understand how people's memory or
opinions can be changed in emotional contexts or by clever questioning (say in
eyewitness testimony). But many of us had hoped that there was a theoretical
science like physics or chemistry there somewhere and we were disappointed. I
now believe that the problem is simply that there is no unitary subject matter
for psychology -- it is not a natural scientific domain. But I find renewed
hope now that within psychology lies one or more natural scientific domains,
and that cognition, suitably circumscribed to include those aspects that
are explainable in terms of symbol processing operations (together with the
nonsymbolic mechanisms required to support symbol processing) may be one of
those natural scientific domains.
Thus my answer to the question, What do I think of Cognitive
Science? is exactly the answer that Mahatma Gandhi is alleged to have given
to a western reporter who asked him what he thought of Western Civilization.
Gandhi is said to have replied without hesitation, "Yes, that would be a
good idea".
[1] This is the unexpurgated text of the presidential address, presented to the Cognitive Science Society's Eighth Annual Conference, Amherst, Massachusetts, August 16, 1986. It was an after-dinner talk and should be read in that spirit, even though there is a serious message hidden in there somewhere.
[2] When I say inference here I don't just mean logical deduction. I also include such meaning-dependent steps as default assumptions, heuristic reasoning, and even pure guesses. Psychological inference may have its own brand of quasi-logic.
[3] That is not to say, by the way, that connectionist or PDP or some other highly parallel and nonsymbolic system will not do inference. That's clearly not the case since, for example, a network of neurons such as the brain does it and a complex electronic network such as the VAX computer does it. The point is simply that in order to do reasoning or inference, such systems must have an additional level of organization overlaid on top of the network level of organization -- just as does any modern computer. The underlying implementation of the symbol-processing level of organization may well consist of a connectionist-like architecture.
