John R. Searle
There are different ways to present a Presidential Address to the APA; the one I have chosen is simply to report on work that I am doing right now, on work in progress. I am going to present some of my further explorations into the computational model of the mind.\**
The basic idea of the computer model of the mind is
that the mind is the program and the brain the hardware of a computational
system. A slogan one often sees is: "the mind is to the brain as the
program is to the hardware." \**
Let us begin our investigation of this claim by
distinquishing three questions:
I will be addressing 1 and
not 2 or 3. I think 2 can be decisively answered in the negative. Since
programs are defined purely formally or syntactically and since minds have an
intrinsic mental content, it follows immediately that the program by itself
cannot constitute the mind. The formal syntax of the program does not by itself
guarantee the presence of mental contents. I showed this a decade ago in the
Chinese Room Argument (Searle,1980). A computer, me for example, could run the
steps in the program for some mental capacity, such as understanding Chinese,
without understanding a word of Chinese. The argument rests on the simple
logical truth that syntax is not the same as, nor is it by itself sufficient for,
semantics. So the answer to the second question is obviously "No".
The answer to 3. seems to me equally obviously
"Yes", at least on a natural interpretation. That is, naturally
interpreted, the question means: Is there some description of the brain such
that under that description you could do a computational simulation of the
operations of the brain. But since according to Church's thesis, anything that
can be given a precise enough characterization as a set of steps can be
simulated on a digital computer, it follows trivially that the question has an
affirmative answer. The operations of the brain can be simulated on a digital
computer in the same sense in which weather systems, the behavior of the New
York stock market or the pattern of airline flights over Latin America can. So
our question is not, "Is the mind a program?" The answer to that is,
"No". Nor is it, "Can the brain be simulated?" The answer
to that is, "Yes". The question is, "Is the brain a digital
computer?" And for purposes of this discussion I am taking that question
as equivalent to: "Are brain processes computational?"
One might think that this question would lose much of
its interest if question 2 receives a negative answer. That is, one might
suppose that unless the mind is a program, there is no interest to the question
whether the brain is a computer. But that is not really the case. Even for
those who agree that programs by themselves are not constitutive of mental
phenomena, there is still an important question: Granted that there is more to
the mind than the syntactical operations of the digital computer; nonetheless,
it might be the case that mental states are at least computational
states and mental processes are computational processes operating over the
formal structure of these mental states. This, in fact, seems to me the
position taken by a fairly large number of people.
I am not saying that the view is fully clear, but the
idea is something like this: At some level of description brain processes are
syntactical; there are so to speak, "sentences in the head". These
need not be sentences in English or Chinese, but perhaps in the "Language
of Thought" (Fodor, 1975). Now, like any sentences, they have a
syntactical structure and a semantics or meaning, and the problem of syntax can
be separated from the problem of semantics. The problem of semantics is: How do
these sentences in the head get their meanings? But that question can be
discussed independently of the question: How does the brain work in processing
these sentences? A typical answer to that latter question is: The brain works
as a digital computer performing computational operations over the syntactical
structure of sentences in the head.
Just to keep the terminology straight, I call the view
that all there is to having a mind is having a program, Strong AI, the view
that brain processes (and mental processes) can be simulated computationally ,
Weak AI. and the view that the brain is a digital computer, Cognitivism.
This paper is about Cognitivism, and I had better say
at the beginning what motivates it. If you read books about the brain (say
Shepherd (1983) or Kuffler and Nicholls (1976)) you get a certain picture of
what is going on in the brain. If you then turn to books about computation (say
Boolos and Jeffrey, 1989) you get a picture of the logical structure of the
theory of computation. If you then turn to books about cognitive science, (say
Pylyshyn, 1985) they tell you that what the brain books describe is really the
same as what the computability books were describing. Philosophically speaking,
this does not smell right to me and I have learned, at least at the beginning
of an investigation, to follow my sense of smell.
I want to begin the
discussion by trying to state as strongly as I can why cognitivism has seemed
intuitively appealing. There is a story about the relation of human
intelligence to computation that goes back at least to Turing's classic paper
(1950), and I believe it is the foundation of the Cognitivist view. I will call
it the Primal Story:
We begin with
two results in mathematical logic, the Church-Turing thesis (or equivalently,
the Churchs's thesis) and Turing's theorem. For our purposes, the Church-Turing
thesis states that for any algorithm there is some Turing machine that can
implement that algorithm. Turing's thesis says that there is a Universal Turing
Machine which can simulate any Turing Machine. Now if we put these two together
we have the result that a Universal Turing Machine can implement any algorithm
whatever.
But now, what made this
result so exciting? What made it send shivers up and down the spines of a whole
generation of young workers in artificial intelligence is the following
thought: Suppose the brain is a Universal Turing Machine.
Well, are there any good reasons for supposing the
brain might be a Universal Turing Machine? Let us continue with the Primal
Story
It is clear that
at least some human mental abilities are algorithmic. For example, I can
consciously do long division by going through the steps of an algorithm for
solving long division problems. It is furthermore a consequence of the Church -
Turing thesis and Turing's theorem that anything a human can do algorithmically
can be done on a Universal Turing Machine. I can implement, for example, the
very same algorithm that I use for long division on a digital computer. In such
a case, as described by Turing (l950), both I, the human computer, and the
mechanical computer are implementing the same algorithm, I am doing it
consciously, the mechanical computer nonconsciously. Now it seems reasonable to
suppose there might also be a whole lot of mental processes going on in my
brain nonconsciously which are also computational. And if so, we could find out
how the brain works by simulating these very processes on a digital computer.
Just as we got a computer simulation of the processes for doing long division,
so we could get a computer simulation of the processs for understanding
language, visual perception, categorization, etc.
"But what about the
semantics? After all, programs are purely syntactical." Here another set
of logico-mathematical results comes into play in the Primal Story.
The development
of proof theory showed that within certain well known limits the semantic
relations between propositions can be entirely mirrored by the syntactic
relations between the sentences that express those propositions. Now suppose
that mental contents in the head are expressed syntactically in the head, then
all we would need to account for mental processes would be computational
processes between the syntactical elements in the head. If we get the proof
theory right the semantics will take care of itself; and that is what computers
do: they implement the proof theory.
We thus have a well defined
research program. We try to discover the programs being implemented in the
brain by programming computers to implement the same programs. We do this in
turn by getting the mechanical computer to match the performance of the human
computer (i.e. to pass the Turing Test) and then getting the psychologists to
look for evidence that the internal processes are the same in the two types of
computer.
Now in what follows I would like the reader to keep
this Primal Story in mind - notice especially Turing's contrast between the
conscious implementation of the program by the human computer and the
nonconscious implementation of programs, whether by the brain or by the
mechanical computer; notice furthermore the idea that we might just discover
programs running in nature, the very same programs that we put into our
mechanical computers.
If one looks at the books and articles supporting
Cognitivism one finds certain common assumptions, often unstated, but
nonetheless pervasive.
First, It is often assumed that the only alternative to the
view that the brain is a digital computer is some form of dualism. The idea is
that unless you believe in the existence of immortal Cartesian souls, you must
believe that the brain is a computer. Indeed, it often seems to be assumed that
the question whether the brain is a physical mechanism determining our mental
states and whether the brain is a digital computer are the same question.
Rhetorically speaking, the idea is to bully the reader into thinking that
unless he accepts the idea that the brain is some kind of computer, he is
committed to some weird antiscientific views. Recently the field has opened up
a bit to allow that the brain might not be an old fashioned von Neumann style
digital computer, but rather a more sophisticated kind of parallel processing
computational equipment. Still, to deny that the brain is computational is to
risk losing your membership in the scientific community.
Second, it is also assumed that the question whether brain
processes are computational is just a plain empirical question. It is to be
settled by factual investigation in the same way that such questions as whether
the heart is a pump or whether green leaves do photosynthesis were settled as
matters of fact. There is no room for logic chopping or conceptual analysis,
since we are talking about matters of hard scientific fact. Indeed I think many
people who work in this field would doubt that the title of this paper poses an
appropriate philosophic question at all. "Is the brain really a digital
computer?" is no more a philosophical question than "Is the
neurotransmitter at neuro-muscular junctions really acetylcholene?"
Even people who are unsympathetic to Cognitivism, such
as Penrose and Dreyfus, seem to treat it as a straightforward factual issue.
They do not seem to be worried about the question what sort of claim it might
be that they are doubting. But I am puzzled by the question: What sort of fact
about the brain could constitute its being a computer?
Third, another stylistic feature of this literature is the
haste and sometimes even carelessness with which the foundational questions are
glossed over. What exactly are the anatomical and physiological features of
brains that are being discussed? What exactly is a digital computer? And how
are the answers to these two questions supposed to connect? The usual procedure
in these books and articles is to make a few remarks about 0's and 1's, give a
popular summary of the Church-Turing thesis, and then get on with the more
exciting things such as computer achievements and failures. To my surprise in
reading this literature I have found that there seems to be a peculiar
philosophical hiatus. On the one hand, we have a very elegant set of
mathematical results ranging from Turing's theorem to Church's thesis to recursive
function theory. On the other hand, we have an impressive set of electronic
devices which we use every day. Since we have such advanced mathematics and
such good electronics, we assume that somehow somebody must have done the basic
philosophical work of connecting the mathematics to the electronics. But as far
as I can tell that is not the case. On the contrary, we are in a peculiar
situation where there is little theoretical agreement among the practitioners
on such absolutely fundamental questions as, What exactly is a digital
computer? What exactly is a symbol? What exactly is a computational process?
Under what physical conditions exactly are two systems implementing the same
program?
Since there is no universal
agreement on the fundamental questions, I believe it is best to go back to the
sources, back to the original definitions given by Alan Turing.
According to Turing, a Turing machine can carry out
certain elementary operations: It can rewrite a 0 on its tape as a 1, it can
rewrite a 1 on its tape as a 0, it can shift the tape 1 square to the left, or
it can shift the tape 1 square to the right. It is controlled by a program of
instruction and each instruction specifies a condition and an action to be
carried out if the condition is satisfied.
That is the standard definition of computation, but,
taken literally, it is at least a bit misleading. If you open up your home
computer you are most unlikely to find any 0's and 1's or even a tape. But this
does not really matter for the definition. To find out if an object is really a
digital computer, it turns out that we do not actually have to look for 0's and
1's, etc.; rather we just have to look for something that we could treat as
or count as or could be used to function as a 0's and 1's.
Furthermore, to make the matter more puzzling, it turns out that this machine
could be made out of just about anything. As Johnson-Laird says, "It could
be made out of cogs and levers like an old fashioned mechanical calculator; it
could be made out of a hydraulic system through which water flows; it could be
made out of transistors etched into a silicon chip through which electric
current flows; it could even be carried out by the brain. Each of these
machines uses a different medium to represent binary symbols. The positions of
cogs, the presence or absence of water, the level of the voltage and perhaps
nerve impulses" (Johnson Laird, 1988, p. 39).
Similar remarks are made by most of the people who
write on this topic. For example, Ned Block (Block, 1990), shows how we can
have electrical gates where the 1's and 0's are assigned to voltage levels of 4
volts and 7 volts respectively. So we might think that we should go and look
for voltage levels. But Block tells us that 1 is only
"conventionally" assigned to a certain voltage level. The situation
grows more puzzling when he informs us further that we did not need to use
electricity at all but we could have used an elaborate system of cats and mice
and cheese and make our gates in such as way that the cat will strain at the
leash and pull open a gate which we can also treat as if it were a 0 or 1. The
point, as Block is anxious to insist, is "the irrelevance of hardware
realization to computational description. These gates work in different ways
but they are nonetheless computationally equivalent" (p. 260). In the same
vein, Pylyshyn says that a computational sequence could be realized by "a
group of pigeons trained to peck as a Turing machine!"( Pylyshn,1985,p.57)
But now if we are trying to take seriously the idea
that the brain is a digital computer, we get the uncomfortable result that we
could make a system that does just what the brain does out of pretty much
anything. Computationally speaking, on this view, you can make a "brain"
that functions just like yours and mine out of cats and mice and cheese or
levers or water pipes or pigeons or anything else provided the two systems are,
in Block's sense, "computationally equivalent" . You would just need
an awful lot of cats, or pigeons or waterpipes, or whatever it might be. The
proponents of Cognitivism report this result with sheer and unconcealed
delight. But I think they ought to be worried about it, and I am going to try
to show that it is just the tip of a whole iceberg of problems.
Why are the defenders of
computationalism not worried by the implications of multiple realizability? The
answer is that they think it is typical of functional accounts that the same
function admits of multiple realizations. In this respect, computers are just
like carburettors and thermostats. Just as carburettors can be made of brass or
steel, so computers can be made of an indefinite range of hardware materials.
But there is a difference: The classes of carburettors
and thermostats are defined in terms of the production of certain physical
effects. That is why, for example, nobody says you can make carburettors out of
pigeons. But the class of computers is defined syntactically in terms of the assignment
of 0's and 1's. The multiple realizability is a consequence not of the fact
that the same physical effect can be achieved in different physical substances,
but that the relevant properties are purely syntactical. The physics is
irrelevant except in so far as it admits of the assignments of 0's and 1's and
of state transitions between them.
But this has two consequences which might be
disastrous:
Now why exactly would these
consequences be disastrous?
Well, we wanted to know how the brain works,
specifically how it produces mental phenomena. And it would not answer that
question to be told that the brain is a digital computer in the sense in which
stomach , liver, heart, solar system , and the state of Kansas are all digital
computers. The model we had was that we might discover some fact about the
operation of the brain which would show that it is a computer. We wanted to
know if there was not some sense in which brains were intrinsically
digital computers in a way that green leaves intrinsically perform
photosynthesis or hearts intrinsically pump blood. It is not a matter of us
arbitrarily or "conventionally" assigning the word "pump"
to hearts or "photosynthesis" to leaves. There is an actual fact of
the matter. And what we were asking is, "Is there in that way a fact of
the matter about brains that would make them digital computers?" It does
not answer that question to be told, yes, brains are digital computers because
everything is a digital computer.
On the standard textbook definition of computation,
I think the main reason
that the proponents do not see that multiple or universal realizability is a
problem is that they do not see it as a consequence of a much deeper point,
namely that the "syntax" is not the name of a physical feature, like
mass or gravity. On the contrary they talk of "syntactical engines"
and even "semantic engines" as if such talk were like that of
gasoline engines or diesel engines, as if it could be just a plain matter of
fact that the brain or anything else is a syntactical engine.
I think it is probably possible to block the result of
universal realizability by tightening up our definition of computation.
Certainly we ought to respect the fact that programmers and engineers regard it
as a quirk of Turing's original definitions and not as a real feature of
computation. Unpublished works by Brian Smith , Vinod Goel, and John Batali all
suggest that a more realistic definition of computation will emphasize such
features as the causal relations among program states, programmability and
controllability of the mechanism, and situatedness in the real world. But these
further restrictions on the definition of computation are no help in the
present discussion because the really deep problem is that syntax is
essentially an observer relative notion. The multiple realizability of
computationally equivalent processes in different physical media was not just a
sign that the processes were abstract, but that they were not intrinsic to the
system at all. They depended on an interpretation from outside. We were looking
for some facts of the matter which would make brain processes computational;
but given the way we have defined computation, there never could be any such
facts of the matter. We can't, on the one hand, say that anything is a digital
computer if we can assign a syntax to it and then suppose there is a factual
question intrinsic to its physical operation whether or not a natural system
such as the brain is a digital computer.
And if the word "syntax" seems puzzling, the
same point can be stated without it. That is, someone might claim that the
notion of "syntax" and "symbols" are just a manner of
speaking and that what we are really interested in is the existence of systems
with discrete physical phenomena and state transitions between them. On this
view we don't really need O's and 1's; they are just a convenient shorthand.
But, I believe, this move is no help. A physical state of a system is a
computational state only relative to the assignment to that state of some
computational role, function, or interpretation. The same problem arises
without 0's and 1's because notions such as computation, algorithm and program
do not name intrinsic physical features of systems. Computational states are
not discovered within the physics, they are assigned to the
physics.
This is a different argument from the Chinese Room
Argument and I should have seen it ten years ago but I did not. The Chinese
Room Argument showed that semantics is not intrinsic to syntax. I am now making
the separate and different point that syntax is not intrinsic to physics. For
the purposes of the original argument I was simply assuming that the
syntactical characterization of the computer was unproblematic. But that is a
mistake. There is no way you could discover that something is intrinsically a
digital computer because the characterization of it as a digital computer is
always relative to an observer who assigns a syntactical interpretation to the
purely physical features of the system. As applied to the Language of Thought
hypothesis, this has the consequence that the thesis is incoherent. There is no
way you could discover that there are, intrinsically, unknown sentences in your
head because something is a sentence only relative to some agent or user who
uses it as a sentence. As applied to the computational model generally, the
characterization of a process as computational is a characterization of a
physical system from outside; and the identification of the process as
computational does not identify an intrinsic feature of the physics, it is
essentially an observer relative characterization.
This point has to be understood precisely. I am not
saying there are a priori limits on the patterns we could discover in
nature. We could no doubt discover a pattern of events in my brain that was
isomorphic to the implementation of the vi program on this computer. But to say
that something is functioning as a computational process is to say
something more than that a pattern of physical events is occuring. It requires
the assignment of a computational interpretation by some agent. Analogously, we
might discover in nature objects which had the same sort of shape as chairs and
which could therefore be used as chairs; but we could not discover objects in
nature which were functioning as chairs, except relative to some agents who
regarded them or used them as chairs.
So far, we seem to have
arrived at a problem. Syntax is not part of physics. This has the consequence
that if computation is defined syntactically then nothing is intrinsically a
digital computer solely in virtue of its physical properties. Is there any way
out of this problem? Yes, there is, and it is a way standardly taken in
cognitive science, but it is out of the frying pan and into the fire. Most of
the works I have seen in the computational theory of the mind commit some
variation on the homunculus fallacy. The idea always is to treat the brain as
if there were some agent inside it using it to compute with. A typical case is
David Marr(1982) who describes the task of vision as proceeding from a
two-dimensional visual array on the retina to a three-dimensional description
of the external world as output of the visual system. The difficulty is: Who is
reading the description? Indeed, it looks throughout Marr's book, and in other
standard works on the subject, as if we have to invoke a homunculus inside the
system in order to treat its operations as genuinely computational.
Many writers feel that the homunculus fallacy is not
really a problem, because, with Dennett (1978), they feel that the homunculus
can be "discharged". The idea is this: Since the computational
operations of the computer can be analyzed into progressively simpler units,
until eventually we reach simple flip-flop, "yes-no", "1-0"
patterns, it seems that the higher-level homunculi can be discharged with
progressively stupider homunculi, until finally we reach the bottom level of a
simple flip-flop that involves no real homunculus at all. The idea, in short,
is that recursive decomposition will eliminate the homunculi.
It took me a long time to figure out what these people
were driving at, so in case someone else is similarly puzzled I will explain an
example in detail: Suppose that we have a computer that multiplies six times
eight to get forty-eight. Now we ask "How does it do it?" Well, the
answer might be that it adds six to itself seven times.\** But if you ask
"How does it add six to itself seven times?", the answer might be
that, first, it converts all of the numerals into binary notation, and second,
it applies a simple algorithm for operating on binary notation until finally we
reach the bottom level at which the only instructions are of the form,
"Print a zero, erase a one." So, for example, at the top level our
intelligent homunculus says "I know how to multiply six times eight to get
forty-eight". But at the next lower-level he is replaced by a stupider
homunculus who says "I do not actually know how to do multiplication, but
I can do addition." Below him are some stupider ones who say "We do
not actually know how to do addition or multiplication, but we know how to
convert decimal to binary." Below these are stupider ones who say "We
do not know anything about any of this stuff, but we know how to operate on
binary symbols." At the bottom level are a whole bunch of a homunculi who
just say "Zero one, zero one". All of the higher levels reduce to
this bottom level. Only the bottom level really exists; the top levels are all
just as-if.
Various authors (e.g. Haugeland (1981), Block (1990))
describe this feature when they say that the system is a syntactical engine
driving a semantic engine. But we still must face the question we had before:
What facts intrinsic to the system make it syntactical? What facts about the
bottom level or any other level make these operations into zeros and ones? Without
a homunculus that stands outside the recursive decomposition, we do not even
have a syntax to operate with The attempt to eliminate the homunculus
fallacy through recursive decomposition fails, because the only way to get the
syntax intrinsic to the physics is to put a homunculus in the physics.
There is a fascinating feature to all of this.
Cognitivists cheerfully concede that the higher levels of computation , e.g.
"multiply 6 times 8" are observer relative; there is nothing really
there that corresponds directly to multiplication; it is all in the eye of the
homunculus/beholder. But they want to stop this concession at the the lower
levels. The electronic circuit , they admit, does not really multiply 6X8 as
such, but it really does manipulate 0's and 1's and these manipulations, so to
speak, add up to multiplication. But to concede that the higher levels of
computation are not intrinsic to the physics is already to concede that the
lower levels are not intrinsic either. So the homunculus fallacy is still with
us.
For real computers of the kind you buy in the store,
there is no homunculus problem, each user is the homunculus in question. But if
we are to suppose that the brain is a digital computer, we are still faced with
the question "And who is the user?" Typical homunculus questions in
cognitive science are such as the following: "How does the visual system
compute shape from shading; how does it compute object distance from size of
retinal image?" A parallel question would be, "How do nails compute
the distance they are to travel in the board from the impact of the hammer and
the density of the wood?" And the answer is the same in both sorts of
case: If we are talking about how the system works intrinsically neither nails
nor visual systems compute anything. We as outside homunculi might describe
them computationally, and it is often useful to do so. But you do not
understand hammering by supposing that nails are somehow intrinsically
implementing hammering algorithms and you do not understand vision by supposing
the system is implementing, e.g, the shape from shading alogorithm.
Certain sorts of
explanations in the natural sciences specify mechanisms which function causally
in the production of the phenomena to be explained. This is especially common
in the biological sciences. Think of the germ theory of disease, the account of
photosynthesis, the DNA theory of inherited traits, and even the Darwinian
theory of natural selection. In each case a causal mechanism is specified, and
in each case the specification gives an explanation of the output of the
mechanism. Now if you go back and look at the Primal Story it seems clear that
this is the sort of explanation promised by Cognitivism. The mechanisms by
which brain processes produce cognition are supposed to be computational, and
by specifying the programs we will have specified the causes of cognition. One
beauty of this research program, often remarked, is that we do not need to know
the details of brain functioning in order to explain cognition. Brain processes
provide only the hardware implementation of the cognitive programs, but the
program level is where the real cognitive explanations are given. On the
standard account, as stated by Newell for example, there are three levels of
explanation, hardware, program, and intentionality( Newell calls this last
level, the knowledge level), and the special contribution of cognitive science
is made at the program level.
But if what I have said so far is correct, then there
is something fishy about this whole project. I used to believe that as a causal
account the cognitivist's theory was at least false; but I now am having
difficulty formulating a version of it that is coherent even to the point where
it could be an empirical thesis at all. The thesis is that there are a whole
lot of symbols being manipulated in the brain, 0's and 1's flashing through the
brain at lightning speed and invisible not only to the naked eye but even to
the most powerful electron microscope, and it is these which cause cognition.
But the difficulty is that the 0's and 1's as such have no causal powers at all
because they do not even exist except in the eyes of the beholder. The
implemented program has no causal powers other than those of the implementing
medium because the program has no real existence, no ontology, beyond that of
the implementing medium. Physically speaking there is no such thing as a
separate "program level".
You can see this if you go back to the Primal Story and
remind yourself of the difference between the mechanical computer and Turing's
human computer. In Turing's human computer there really is a program level
intrinsic to the system and it is functioning causally at that level to convert
input to output. This is because the human is consciously following the rules
for doing a certain computation, and this causally explains his performance.
But when we program the mechanical computer to peform the same computation, the
assignment of a computational interpretation is now relative to us, the outside
homunculi. And there is no longer a level of intentional causation intrinsic to
the system. The human computer is consciously following rules, and this fact
explains his behavior, but the mechanical computer is not literally following
any rules at all. It is designed to behave exactly as if it were following
rules, and so for practical, commercial purposes it does not matter. Now
Cognitivism tells us that the brain functions like the commercial computer and
this causes cognition. But without a homunculus, both commercial computer and
brain have only patterns and the patterns have no causal powers in addition to
those of the implementing media. So it seems there is no way Cognitivism could
give a causal account of cognition.
However there is a puzzle for my view. Anyone who
works with computers even casually knows that we often do in fact give causal
explanations that appeal to the program. For example, we can say that when I
hit this key I got such and such results because the machine is implementing
the vi program and not the emacs program; and this looks like an ordinary
causal explanation. So the puzzle is, how do we reconcile the fact that syntax,
as such, has no causal powers with the fact that we do give causal explanations
that appeal to programs? And, more pressingly, would these sorts of
explanations provide an appropriate model for Cognitivism, will they rescue
Cognitivism? Could we for example rescue the analogy with thermostats by
pointing out that just as the notion "thermostat" figures in causal
explanations independently of any reference to the physics of its
implementation, so the notion "program", might be explanatory while
equally independent of the physics.
To explore this puzzle let us try to make the case for
Cognitivism by extending the Primal Story to show how the Cognitivist
investigative procedures work in actual research practice. The idea, typically,
is to program a commercial computer so that it simulates some cognitive
capacity, such as vision or language. Then, if we get a good simulation, one
that gives us at least Turing equivalence, we hypothesize that the brain
computer is running the same program as the commercial computer, and to test
the hypothesis we look for indirect psychological evidence, such as reaction
times. So it seems that we can causally explain the behavior of the brain
computer by citing the program in exactly the same sense in which we can
explain the behavior of the commerical computer. Now what is wrong with that?
Doesn't it sound like a perfectly legitimate scientific research program? We
know that the commercial computer's conversion of input to output is explained
by a program, and in the brain we discover the same program, hence we have a
causal explanation.
Two things ought to worry us immediately about this
project. First, we would never accept this mode of explanation for any function
of the brain where we actually understood how it worked at the neurobiological
level. Second we would not accept it for other sorts of system that we can
simulate computationally. To illustrate the first point, consider for example
the famous account of "What the Frog's Eye Tells the Frogs Brain"
(Lettvin, et al. 1959 in McCulloch, 1965) The account is given entirely in terms
of the anatomy and physiology of the frog's nervous system. A typical passage,
chosen at random goes like this:
"1.
Sustained Contrast Detectors.
An unmyelinated
axon of this group does not respond when the general illumination is turned on
or off. If the sharp edge of an object either lighter or darker than the
background moves into its field and stops, it discharges promptly and continues
discharging, no matter what the shape of the edge or whether the object is
smaller or larger than the receptive field."(p. 239).
I have never heard anyone
say that all this is just the hardware implementation, and that they should
have figured out which program the frog was implementing. I do not doubt that
you could do a computer simulation of the frog's "bug detectors".
Perhaps someone has done it. But we all know that once you understand how the
frog's visual system actually works, the "computational level" is
just irrelevant.
To illustrate the second point, consider simulations
of other sorts of systems. I am for example typing these words on a machine
that simulates the behavior of an old fashioned mechanical typewriter.\** As
simulations go, the word processing program simulates a typewriter better than
any AI program I know of simulates the brain. But no sane person thinks:
"At long last we understand how typewriters work, they are implementations
of word processing programs." It is simply not the case in general that
computational simulations provide causal explanations of the phenomena
simulated.
So what is going on? We do not in general suppose that
computational simulations of brain processes give us any explanations in place
of or in addition to neurobiological accounts of how the brain actually works.
And we do not in general take "X is a computational simulation of Y"
to name a symmetrical relation. That is, we do not suppose that because the
computer simulates a typewriter that therefore the typewriter simulates a
computer. We do not suppose that because a weather program simulates a
hurricane, that the causal explanation of the behavior of the hurricane is
provided by the program. So why should we make an exception to these principles
where unknown brain processes are concerned? Are there any good grounds for
making the exception? And what kind of a causal explanation is an explanation
that cites a formal program?
Here, I believe, is the solution to our puzzle. Once
you remove the homunculus from the system, you are left only with a pattern of
events to which someone from outside could attach a computational interpretation.
Now the only sense in which the specification of the pattern by itself provides
a causal explanation is that if you know that a certain pattern exists in a
system you know that some cause or other is responsible for the pattern. So you
can, for example, predict later stages from earlier stages. Furthermore, if you
already know that the system has been programmed by an outside homunculus you
can give explanations that make reference to the intentionality of the
homunculus. You can say, e.g. this machine behaves the way it does because it
is running vi. That is like explaining that this book begins with a bit about
happy families and does not contain any long passages about a bunch of
brothers, because it is Tolstoy's Anna Karenina and not Dostoevsky's The
Brothers Karamozov. But you cannot explain a physical system such as a
typewriter or a brain by identifying a pattern which it shares with its
computational simulation, because the existence of the pattern does not explain
how the system actually works as a physical system. In the case of
cognition the pattern is at much too high a level of abstraction to explain
such concrete mental (and therefore physical) events as the occurrence of a
visual perception or the understanding of a sentence.
Now, I think it is obvious that we cannot explain how
typewriters and hurricanes work by pointing to formal patterns they share with
their computational simulations. Why is it not obvious in the case of the
brain?
Here we come to the second part of our solution to the
puzzle. In making the case for Cognitivism we were tacitly supposing that the
brain might be implementing algorithms for cognition, in the same sense that
Turing's human computer and his mechanical computer implement algorithms. But
it is precisely that assumption which we have seen to be mistaken. To see this
ask yourself what happens when a system implements an algorithm. In the human
computer the system consciously goes through the steps of the algorithm, so the
process is both causal and logical; logical, because the algorithm provides a
set of rules for deriving the output symbols from the input symbols; causal,
because the agent is making a conscious effort to go through the steps.
Similarly in the case of the mechanical computer the whole system includes an
outside homunculus, and with the homunculus the system is both causal and
logical; logical because the homunculus provides an interpretation to the the
processes of the machine; and causal because the hardware of the machine causes
it to go through the processes. But these conditions cannot be met by the
brute, blind, nonconscious neurophysiological operations of the brain. In the
brain computer there is no conscious intentional implementation of the
algorithm as there is in the human computer, but there can't be any
nonconscious implementation as there is in the mechanical computer either,
because that requires an outside homunculus to attach a computational
interpretation to the physical events. The most we could find in the brain is a
pattern of events which is formally similar to the implemented program in the
mechanical computer, but that pattern, as such, has no causal powers to call
its own and hence explains nothing.
In sum, the fact that the attribution of syntax
identifies no further causal powers is fatal to the claim that programs provide
causal explanations of cognition. To explore the consequences of this let us
remind ourselves of what Cognitivist explanations actually look like.
Explanations such as Chomsky's account of the syntax of natural languages or
Marr's account of vision proceed by stating a set of rules according to which a
symbolic input is converted into a symbolic output. In Chomsky's case, for
example, a single input symbol, S, is converted into any one of a potentially infinite
number of sentences by the repeated application of a set of syntactical rules.
In Marr's case, representations of a two dimensional visual array are converted
into three dimensional "descriptions" of the world in accordance with
certain algorithms. Marr's tripartite distinction between the computational
task, the algorithmic solution of the task and the hardware implementation of
the algorithm, has (like Newell's distinctions) become famous as a statement of
the general pattern of the explanation.
If you take these explanations naively, as I do, it is
best to think of them as saying that it is just as if a man alone in a room
were going through a set of steps of following rules to generate English
sentences or 3D descriptions, as the case might be. But now, let us ask what
facts in the real world are supposed to correspond to these explanations as
applied to the brain. In Chomsky's case, for example we are not supposed to
think that the agent consciously goes through a set of repeated applications of
rules; nor are we supposed to think that he is unconsciouly thinking his way
through the set of rules. Rather the rules are "computational" and
the brain is carrying out the computations. But what does that mean? Well, we
are supposed to think that it is just like a commercial computer. The sort of
thing that corresponds to the ascription of the same set of rules to a
commercial computer is supposed to correspond to the ascription of those rules
to the brain. But we have seen that in the commercial computer the ascription
is always observer relative, the ascription is made relative to a homunculus
who assigns computational interpretations to the hardware states. Without the
homunculus there is no computation, just an electronic circuit. So how do we
get computation into the brain without a homunculus? As far as I know neither
Chomsky nor Marr ever addressed the question or even thought there was such a
question. But without a homunculus there is no explanatory power to the
postulation of the program states. There is just a physical mechanism, the
brain, with its various real physical and physical/mental causal levels of
description.
In
this section I turn finally to what I think is, in some ways, the central issue
in all of this, the issue of information processing. Many people in the
"cognitive science" scientific paradigm will feel that much of my
discussion is simply irrelevant and they will argue against it as follows:
"There is a difference between the brain and
all of these other systems you have been describing and this difference
explains why a computational simulation in the case of the other systems is a
mere simulation whereas in the case of the brain a computational simulation is
actually duplicating and not merely modeling the functional properties of the
brain. The reason is that the brain, unlike these other systems, is an information
processing system. And this fact about the brain is, in your words,
"intrinsic". It is just a fact about biology that the brain functions
to process information, and since we can also process the same information
computationally, computational models of brain processes have a different role
altogether from computational models of, for example, the weather.
So there is a well defined research question:
"Are the computational procedures by which the brain processes information
the same as the procedures by which computers process the same
information?"
What
I just imagined an opponent saying embodies one of the worst mistakes in
cognitive science. The mistake is to suppose that in the sense in which
computers are used to process information, brains also process information. To
see that that is a mistake contrast what goes on in the computer with what goes
on in the brain. In the case of the computer, an outside agent encodes some
information in a form that can be processed by the circuitry of the computer.
That is, he or she provides a syntactical realization of the information that
the computer can implement in, for example, different voltage levels. The
computer then goes through a series of electrical stages that the outside agent
can interpret both syntactically and semantically even though, of course, the
hardware has no intrinsic syntax or semantics: It is all in the eye of the
beholder. And the physics does not matter provided only that you can get it to
implement the algorithm. Finally, an output is produced in the form of physical
phenomena which an observer can interpret as symbols with a syntax and a
semantics.
But now contrast that with the brain. In the case of
the brain, none of the relevant neurobiological processes are observer relative
(though of course, like anything they can be described from an observer
relative point of view) and the specificity of the neurophysiology matters
desperately. To make this difference clear, let us go through an example.
Suppose I see a car coming toward me. A standard computational model of vision
will take in information about the visual array on my retina and eventually
print out the sentence, "There is a car coming toward me". But that
is not what happens in the actual biology. In the biology a concrete and
specific series of electro-chemical reactions are set up by the assault of the
photons on the photo receptor cells of my retina, and this entire process
eventually results in a concrete visual experience. The biological reality is
not that of a bunch of words or symbols being produced by the visual system,
rather it is a matter of a concrete specific conscious visual event; this very
visual experience. Now that concrete visual event is as specific and as
concrete as a hurricane or the digestion of a meal. We can, with the computer,
do an information processing model of that event or of its production, as we
can do an information model of the weather, digestion or any other phenomenon,
but the phenomena themselves are not thereby information processing systems.
In short, the sense of information processing that is
used in cognitive science, is at much too high a level of abstraction to
capture the concrete biological reality of intrinsic intentionality. The
"information" in the brain is always specific to some modality or
other. It is specific to thought, or vision, or hearing, or touch, for example.
The level of information processing which is described in the cognitive science
computational models of cognition , on the other hand, is simply a matter of
getting a set of symbols as output in response to a set of symbols as input.
We are blinded to this difference by the fact that the
same sentence, "I see a car coming toward me", can be used to record
both the visual intentionality and the output of the computational model of
vision. But this should not obscure from us the fact that the visual experience
is a concrete event and is produced in the brain by specific electro-chemical
biological processes. To confuse these events and processes with formal symbol
manipulation is to confuse the reality with the model. The upshot of this part
of the discussion is that in the sense of "information" used in
cognitive science it is simply false to say that the brain is an information
processing device.
This brief argument has a simple logical structure and
I will lay it out:
Block, Ned (1990,
forthcoming). "The Computer Model of the Mind".
Boolos, George S. and Jeffrey, Richard C. (1989). Computability
and Logic. Cambridge University Press, Cambridge Mass.
Dennett, Daniel C. (1978). Brainstorms: Philosophical
Essays on Mind and Psychology. MIT Press, Cambridge, Mass.
Feigenbaum, E.A. and Feldman, J. ed. (1963). Computers
and Thought. McGraw-Hill Company, New York and San Francisco.
Fodor, J. (1975) The Language of Thought.
Thomas Y. Crowell, New York.
Haugeland, John ed. (1981). Mind Design. MIT
Press, Cambridge, Mass.
Johnson-Laird, P. N. (1988). The Computer and the Mind. Harvard
University Press, Cambridge Mass.
Kuffler, Stephen W. and Nicholls, John G. (1976). From
Neuron to Brain. Sinauer Associates, Sunderland, Mass.
Lettvin, J.Y., Maturana, H.R., McCulloch, W.S., and
Pitts, W.H. (1959). What the Frog's Eye Tells the Frog's Brain. Proceedings
of the Institute of Radio Engineers, 47, 1940-1951, reprinted in
McCulloch, 1965, pp.230-255.
Marr, David (1982). Vision. W.H. Freeman and
Company, San Francisco.
McCulloch, Warren S. (1965). The Embodiments of Mind. MIT Press,
Cambridge, Mass.
Pylyshyn (1985). Computation and Cognition.
MIT Press, Cambridge, Mass.
Searle, John R. (1980). "Minds, Brains and
Programs", The Behavioral and Brain Sciences.3, pp. 417-424.
Searle, John R. (1984). Minds, Brains and Science.
Harvard University Press, Cambridge, Mass.
Sharples, M., Hogg, D., Hutchinson, C., Torrance, S.,
and Young, D. (1988). Computers and Thought. MIT Press, Cambridge,
Mass. and London.
Shepherd, Gordon M. (1983). Neurobiology.
Oxford University Press, New York and Oxford.
Turing, Alan (1950). "Computing Machinery and
Intelligence." Mind, 59, 433-460.