[MUSIC PLAYING] Good afternoon, and welcome. I'm Kim Vandiver, the
chairman of the MIT faculty. We gather here today
to honor two people-- James R Killian, Jr. and
Professor Noam Chomsky. The Killian Award
is the highest award that the MIT faculty
bestows upon one of its own, named after the late Jim
Killian, 10th president of MIT. He came to MIT as a
student in the mid-1920s. He joined the staff of
Technology Review in 1926, and he retired as chairman
of the MIT Corporation 45 years later, in 1971. He brought great honor
and distinction to MIT through his many years
of national service. I'll mention two
examples, which are just the tip of the iceberg. In the late 1950s, he was the
first full-time science adviser to a president of
the United States, and provided
important leadership in the period that set
the nation's course to putting a man on the moon. He also served as chairman
of the Carnegie Commission on Educational Television,
which recommended the formation of the Corporation
for Public Broadcasting. He was also a very genuine
and likeable person, as I discovered quite
by accident one day about 20 years ago, when I was
a new graduate student at MIT. I sat down at a table in
the Lobdell dining room with an elderly gentleman
that I didn't know, and we struck up a
conversation, one that was to begin a long series
of enlightening conversations I had with him until the
time of his death in 1988. His easy way with students and
with other members of the MIT community are one of the things
that define Jim Killian to us. This award is
named in his honor. We bestow it upon
a faculty member each year who has also brought
honor and distinction to MIT. Professor Chomsky has
been doing so for 37 years as a member of our faculty. Professor Chomsky, would you
come to the stage, please? In preparing for today, I did
one small piece of research on Professor Chomsky, which
I'd like to share with you. With the help of the
MIT library staff, I was able to determine how
often he is cited in the Arts and Humanities Citation Index. [LAUGHTER] As of the last time the
statistics were compiled, we discovered that
he is in the top 10. It happens that none of the
others are still living. [LAUGHTER] He is the most
cited living author. I thought you might like to
know the names of the company he keeps. They are, in order-- Marx, Lennon-- [LAUGHTER] [APPLAUSE] Shakespeare-- [LAUGHTER] Aristotle, the Bible-- [LAUGHTER] Plato, Freud, Chomsky,
followed by Hegel and Cicero. [LAUGHTER] [APPLAUSE] I would like to
read from the award. It says, "The president and
faculty of the Massachusetts Institute of Technology
have the honor to present the James R Killian,
Jr. Faculty Achievement Award to Noam Chomsky, and to announce
his appointment as Killian Award Lecturer for
the academic year '91, '92, in recognition of
his extraordinary contributions to linguistics, philosophy,
and cognitive psychology. One of the foremost
thinkers of this century, he revolutionized the study of
language and of the human mind. His books, monographs,
and articles constitute major milestones
in the history of linguistics. An inspiring teacher,
he has trained scholars who are among the
most important contributors to the discipline. His presence here has
greatly enriched MIT." Professor Chomsky, the
microphone is yours. [APPLAUSE] He's lucky he didn't read what
those citations say, but-- [LAUGHTER] The title of this talk,
the announced title, has three parts-- refers to language, the
cognitive revolutions. And then there's the quotes
around cognitive revolutions. The relation between language
and cognitive science is natural. That's a reasonable combination. Modern linguistics
developed as part of, what's sometimes called,
the cognitive revolution of the 1950s-- actually to a large extent
here in building 20, thanks initially to Jerry
Wiesner's initiatives. It was also a major
factor in the development of the cognitive sciences ever
since their modern origins. So those two words
are sort of obvious. The quotes are intended to
express a lingering skepticism. And when I say cognitive
revolution from here on in, you're supposed to hear a
so-called cognitive revolution. I'll explain why as I proceed. The skepticism that I feel
does not have anything to do with the
actual work that's done in the various areas
of cognitive science-- so vision, language, cognitive
development, a few others. In these areas, there have
been quite significant advances over the past 40 years. The questions in my mind
arise at a second order level, that is, thinking about the
nature of the disciplines that are concerned with
what were traditionally called mental acts
or mental faculties. And I want to come back to
a few of those questions. Some of them are substantive. Some of them are historical. First, a few words about the
contemporary cognitive sciences and the cognitive revolution,
as they look to me. And let me stress that others
don't see it this way at all. This is not intended to be a
survey, but rather a rather personal point of view. George Miller, who was
one of the leading figures and founders of the modern
field, in a recent article, he traced the modern
cognitive sciences back to a meeting in 1956,
held in Cambridge. I can't remember whether
it was here or at Harvard, but one or the other. It was a meeting of the
Institute of Radio Engineers, which had a series of papers on
experimental human psychology, which made use of information
theory and signal detection, and other then pretty new ideas. It had a paper of
mine on language, which outlined some of the basic
ideas of generative grammar. It had another paper on
problem-solving and reasoning. It was a paper by Herb
Simon and Al Newell, a simulation of theorem
proving in elementary logic. And he argues that that
confluence of research papers essentially initiated
the modern field. Well, there were some shared
ideas among the participants. One shared idea was a kind
of shift in perception or perspective toward the
disciplines, a shift which is not terribly surprising
now, but was pretty striking at that moment. The shift was away from
the study of behavior and the product of behavior,
such as words or sounds or their arrangement
in text, a shift from that focus toward the study
of the states and properties of the mind and the brain
that enter into human thought and action. From the first point of view,
which overwhelmingly prevailed at the time, behavior
and its products were regarded as
the object of study. From the second
point of view, what later came to be called
the cognitive sciences, behavior and its products
were of no particular interest in themselves. They simply provided data,
data which would be interesting insofar as it served as evidence
for what really concerns us now, the inner workings of
the mind, where the phrase mind here simply
is a way of talking about relevant
properties and states and processes of the brain. The data of behavior
and the text, and so on, from
this point of view, are to be used alongside
of other data, whatever might turn out to
be useful, say, electrical activity
of the brain. But it has no inherently
privileged status, and it's not the
focus of interest. That's a shift. It's a shift away
from what was called behavioral science,
or structuralism, or stressed statistical
analysis of texts, and a shift toward something
that might properly be called cognitive science. In my personal
view, it was a shift from a kind of natural
history to at least potential natural science. Now, that shift was highly
controversial at the time, and it still is. And in fact, if you were to
use such measures as amount of government funding,
it would still be very much in the minority. But I think it was
the right move. Well, that was one
kind of shared interest from independent ones of view. There was no communication
among the people prior to the meeting. A second shared
interest was an interest in what we might call
computational representational systems. That means a way of looking
at mental activities and mental faculties, at
what the brain is doing, to looking at them as a kind
of a software problem, that is, a study of the
mechanisms, but viewed from a particular
abstract perspective. Analogies can always
be misleading, but one might compare this
approach to, say, chemistry about a century ago, where
there were notions like valence or periodic table or
organic molecule, and so on, but they were not grounded in
the basic physics of the time. The results, however,
and the ideas did provide guidelines
for what turned out to be quite radical
changes in physics that were necessary to
accommodate these ideas. Chemistry was never
actually reduced to physics. It would be more accurate
to say that physics was expanded, in a certain
sense, to chemistry. Similarly, computation
representational theories of the brain attempt to
establish states and properties that have to be accommodated
somehow in the brain sciences, and might well provide,
insofar as they're correct, they might well provide
guidelines for inquiry into mechanisms. It's possible, again, that
they may provide guidelines for what will have to be
radical changes in the way that the brain and its processes
and states are conceived. How two approaches
to the natural world will be unified, if
ever, can never be guessed in advance really. It could be reduction, as in,
say, the biology and chemistry case, more or less. It could be what you
might call expansion, as in the physics,
chemistry case, or it could be something else. Crucially, there's nothing
mysterious about any of this. It seems, to me at least,
to be just normal science, normal science at a typical
stage before unification of several approaches through
some topic in the natural world before they're unified. Well, again, that was and
remains highly controversial. Assuming it, anyway,
how can we proceed? You can look at the
brain or the person from a number of points of view. Let's begin by
taking what you might call a receptive point of view. So we think of the brain, say,
as being in a certain state. A signal reaches the brain. It produces some kind of
symbolic representation of the signal
internally, presumably, and it moves to
some other state. If we focus on the
interstate transitions, we're looking at
problems of learning or maybe, more
accurately, growth. If we focus on the input signal
and the symbolic representation of it, we're looking
at perception. If we focus on the initial
state of the system prior to any relevant
signals, we're looking at the
biological endowment, the innate structure. We could trace the
processes back further. At that point, we'd be
moving into embryology. And one should stress that
these distinctions are pretty artificial, actually. Typically, the systems
that we're looking at are like physical organs. Or we might say more accurately,
like other physical organs, because we are studying
a physical organ from a particular point of view. They are like physical organs,
other physical organs, in that, say, like the digestive system
or the circulatory system or the visual
system, or whatever, which, again, we often
look at abstractly, as one would expect in
the study of the system of any degree of complexity. We can call these particular
systems mental organs, not implying, again,
anything more than that they are involved in
what were traditionally called mental states, mental
faculties, or mental processes. Well, like other
physical organs, these mental organs
quite typically reach a kind of a stable
state at a certain point. That is, the input
signal that's causing an interstate transition
moves the system around within some kind of equivalence
class of states, that are alike, for purposes at hand. That's obviously
an idealization, but a reasonable one. When you look at
the stable state after the point at which there
isn't a significant change, then you're very
clearly studying perception, when you look at the
signal or result relationship. So for example, if you
take the visual system, if you look at
the initial state, the state prior to
relevant signals, you're looking at the basic
structure of the visual system. If you look at the
interstate transition, you're looking at the
growth of the visual system, the changes in the distribution
of vertical and horizontal receptors in the striate cortex
or the onset of binocular vision after a couple of months
of age in humans, and so on. If you look at a
stable state, you're studying adult perception. Adult perception, of course,
has some kind of access to systems of belief
about the world. Not much is known
about this, but it seems reasonable to
assume that that involves some kind of
computational procedure that specifies an infinite range
of beliefs, which are called upon when needed,
although here, we move into a domain of
almost total ignorance. Turning to the
language system, you can look at it the same way. The initial state of the
system, the initial state of the language
faculty, the study of it is sometimes called universal
grammar, these days. The study of the initial
state of the language faculty is the study of that, is a
study of biological endowment. It's an empirical fact, which
seems pretty well established, that this element of
the brain is pretty uniform across the species. The variations appear to be
quite slight apart from extreme pathology. And it also appears to
be unique in essentials. There's no other known organism
that is anything like it. Also, it's rather different
from other biological systems in many respects. Systems of discrete infinity,
which this plainly is, are pretty rare in
the biological world, so it appears. If you're looking at the
interstate transition as signals, some relevant
information comes in, you're studying
language acquisition. It apparently reaches a
stable state at about puberty. This stable, steady
state incorporates a computational procedure,
which characterizes an infinite set of
expressions, each of them having a certain
array of properties, properties of sound,
properties of meaning, properties of
structural organization. That computational procedure
constitutes person's language. To have such a procedure in the
appropriate part of the brain is to have a language
or, as we sometimes say, to know a language. When I speak of a
language, I simply mean the computational
or generative procedure that specifies the properties
of an infinite range of linguistic
expressions, such as, say, the one I just produced. Since your language
is similar enough to mine, when the signal
that I just produced reaches your brain, your
generative procedure constructs an
expression, an expression with these properties, which
is similar enough to the one that I form, so we understand
one another, more or less. Unlike the system
of belief that's accessed in visual
perception, in this case, we happen to know quite a lot
about the generative procedures that are involved in
the use of language, and even about their
innate, invariant basis. And these, in
fact, are the areas of major progress in
the past generation. Well, what about
perception of language? There are several common
assumptions about this. One major assumption
is that there is something called a parser. A parser is something
that maps a signal into a symbolic representation,
paying no attention to any other aspects of
the relevant environment, of the environment. Of course, it is assumed,
if people are rational, that this parser accesses
the generative procedure. That is, it's assumed that you
use your knowledge of English when you interpret signals. Note, however,
several assumptions. First, one assumption
is that such a thing exists, that a parser exists. There is some faculty
of the mind that interprets signals independently
of any other features of the environment. That's not at all obvious. The existence of an internal,
generative procedure is far more secure and far
more obvious than the question of the existence of a parser. And it's also far
better grounded assumption, far better
grounded empirically, and with much richer
theoretical structure. This is contrary to common
belief, which usually considers the two the other way around. Nevertheless, let's assume
that a parser exists, though it isn't obvious. It's further assumed
that the parser doesn't grow the way a language
grows, from some initial state to a stable state. Now, that assumption, again,
is not very well founded. It's based mainly on ignorance. It's the simplest assumption,
and we have no reason to believe anything about it, so
we might as well believe that. There is some kind
of evidence for it, indirect perhaps,
but interesting. There is a forthcoming book
coming out from MIT Press, by Robert Berwick and Sandiway
Fong, a former student of his, on what you might call a
universal parser, a fixed, invariant parser, which has
the property that if you flip its switches one way,
it interprets English. And if you flip its
switches another way, it interprets Japanese. And if that's the
right idea, then there is a universal parser, which
is unaffected by environmental factors, apart from
the properties that serve to identify one or
another language, that is, the array of
flipping of switches. I'll come back to that. There's a further comment. So that assumption,
the assumption that the parser doesn't grow,
while strange, could be true. And there's only this kind
of indirect evidence for it. A further common assumption
is that parsing is easy, something that you do
very quickly and easily. In fact, that's often
held to be a criterion for the theory of language. It's often held to provide-- the criterion is supposed to
be that the theory of language must provide generative
procedures that satisfy this requirement, that
parsing is easy and quick. Well, that one is flatly
false, uncontroversially. Parsing often fails completely
and quite systematically, and it's often
extremely difficult. Many categories of what
are called parsing failures are known, even with
short, simple expressions. A parsing failure just means
a case where the parser works fine, but it gives some kind
of an interpretation which is not the one that's assigned
to the signal by the language. So it's wrong in that sense. It's perfectly true that when--
and very often, you can easily find categories of expressions,
even pretty simple ones, where the parsing
system just breaks down. People don't know
what they're hearing. It sounds like gibberish,
even though it's perfectly well formed and
has fixed, strict meaning, so on and so forth. Now, it's perfectly
true that one uses expressions that
are readily parsed, but that verges on tautology. What it means is we
use what is usable. Other resources of the
language, we just don't use. There's a related assumption. It's often alleged
that language is well adapted to the function
of communication. It's not clear that that
statement is even meaningful, or any similar statements
about biological organs. But to the extent
that one can give some meaning to that
statement, this one, again, looks just false. The design of language appears
to make it largely unusable, which is fine. We just use those fragments
that are usable for us. It doesn't mean that the
design made it usable somehow. There are similar assumptions
in learnability theory, formal learnability theory. It's often claimed,
as a starting point, that natural languages
must be learnable. In fact, natural language
is often defined. Natural languages
are defined often as the set of languages
that are learnable under standard conditions. Now, notice that
that could be true. We don't know that
that's false, as we know that the comparable
statement about parsing is false. It could be true that
languages are learnable, but it's not a
conceptual necessity. Language design might be such. That's our language
faculty could be such, that it permits all kinds
of inaccessible languages. These would be realizable
in the brain, just as, say, English is, but not
accessible to humans. Again, that wouldn't stop
us from doing what we do. We would just pick up
the accessible ones. In fact, there is some recent
work, quite recent work, which suggests that natural
languages are learnable. But if true, that's an empirical
discovery and, in fact, a rather surprising one. Well, I've said nothing so far
about production of language. There is a reason for that. The reason is that apart
from very peripheral aspect, it's almost a complete mystery. We can learn a good deal
about the mechanisms that are put to use when
we speak, and we can study at least certain
aspects of perception, the mapping of
signals to percept, using the internalized
mechanisms, the generative procedure
of the language. But no one has anything sensible
to say about what I'm doing now or what two people are doing
in a normal conversation. It was observed centuries ago
that the normal use of language has quite curious properties. It is unbounded. It's not random. It is not determined
by external stimuli, and there's no reason
to believe that it's determined by internal states. It's uncaused, but it's somehow
appropriate to situations. It's coherent. It evokes in the hearer
thoughts that he or she might have expressed the same way. That collection of properties--
we can call them the creative aspect of language use-- that set of properties was
regarded by the Cartesians as the best evidence
that another organism has a mind like ours, not
just a body, like a dog or a monkey or a robot. Actually, if you look
back at their argument, it was quite reasonable. The science happened
to be wrong, but it was very reasonable. And within the framework
of scientific reasoning, we might put matters
in different terms. But we haven't made any
progress in understanding these critically important
phenomena, the phenomena that concerned them and
that rely much-- the provide much of the basis
for the traditional mind-body distinction. In my opinion, in fact,
contemporary thought about these topics,
usually phrased in terms of notions
like the Turing test, involves a
considerable regression from the much more
scientific approach of Cartesian psychology. I'm not going to have time
to elaborate properly. I merely mention this to provoke
some possible interest or maybe outrage. We'll see. Well, I've mentioned
several kinds of shared interest in the early
modern cognitive sciences, say, back to the 1956 meeting
that George Miller pointed to. One common interest,
again, is in computational
representational systems, looking at the brain in
terms of software problems. For linguistics, at
least, this turned into the study of the
generative procedures that are put to use in
perception and presumably, an expression of
thought speaking, although the latter again
remains a complete mystery. There's another shared interest. It's what we might call
the unification problem. Now, that's a problem that
has two different aspects. One aspect is the relation
between the hardware and the software,
the relationship between computational
representational theories of the brain and the
entities that they postulate, and what you might
call implementation. And as I suggested,
I think you can regard that as similar to
the problem of the relation between, say, chemistry
and physics 100 years ago or genetics and chemistry
50 years ago, and so on. That's one problem,
the unification generally, one aspect
of the unification problem, the deepest aspect. A less deep aspect has
to do with the relation among the various domains
of cognitive science, say, language and
vision, whatever may be involved in problem solving. This is the question, what
do they have in common? Notice that that's much
less profound a problem, if it turns out the answer is
nothing, just the way it is. On the other hand, we'd be upset
if the answer to the first one was no answer. Well, at this point, some of my
skepticism begins to surface. So let's take the first
one, the deep problem, the serious,
scientific problem-- the relation between roughly
the hardware and software, chemistry and physics. Here, there are two different
approaches you can identify. One approach== let's
call it naturalistic-- says, it goes like this. It says, the mental
organs are viewed as computational
representational systems, and we want to show
that they're real, that their statements in
these theories are true, that is, that there are certain
states and the properties of the brain which these
theories exactly capture, in that's what you try to show. The mental organs can also
be viewed in other ways. For example, you
can look at them in terms of cells or atoms. Or more interestingly,
for the moment, you can look at them in
terms of electrical activity of the brain, so-called event
related potentials, ERPs. Now, rather
surprisingly, there's some recent work which shows
quite dramatic and surprising correlations between
certain properties of the computational
representational systems of language and of
ERPs, evoked potentials. For the moment,
these ERP measures have no status apart from their
correlation with categories of expressions that come out of
computational representational theories. That is, they're just
numbers picked at random, because there's no relevant
theory about them that says, just look at these numbers
and not at some other numbers. In themselves, in other
words, they're curiosities. Still, it's interesting,
because they do-- or there are correlations to
rather subtle properties that have emerged in the
attempt to develop computational representational
theories which explain the form and meaning of language. So that suggests an interesting
direction for research, namely to try to unify these
quite different approaches to the brain, and to
place each of them in an appropriate
theoretical context. For the moment, that's primarily
a problem for the ERPs, but if you can pursue
it, it could be quite an interesting direction. And one can think of
lots of consequences. Well, the same is true of the
relation between CR systems, computational
representational systems, and physiology, at
least in principle. Here, it's just in principle,
because in practice at present, we don't really
have much to look at there except
from the theories of the computational
representational systems. But here, whatever the
difficulties may be, we seem to face ordinary
problems of science before the typical situation when
alternative approaches don't-- there's no understood
way to unify them or connect them. Well, there's a
different approach to this unification
problem, which actually dominates
contemporary thinking about cognitive science-- not necessarily the actual
work, but the thinking about what it amounts to. This approach is to divorce
the study of mental organs from the biological
setting altogether and to ask whether some
simulation can, essentially, fool an outside observer. So we ask whether the simulation
passes some rather arbitrary test, called the Turing test. But Turing test is just
a name for a huge battery of possible tests. And so we pick out one
of those at random, and we now ask
whether some, say, extraterrestrial
invented organism or some actual
programmed computer passes the test that
has fooled somebody. And people write
articles in which they ask, how can we
determine empirically whether an extraterrestrial or
a programmed computer or machine can, say, play chess
or understand Chinese? And the answer is
supposed to be, if it can fool an observer
under some arbitrary conditions, called
the Turing test, then we say we've
empirically established that. Actually, you may have
noticed, last November, there was a big splash about a
running of the Turing test, over at the Boston
Computer Museum, together with the Cambridge
Center for Behavioral Studies and a quite distinguished
panel, that a bunch of programs which were
supposed to try to fool people. There was, I think, a
$100,000 prize offered, a big story in the Boston Globe,
front page story in the New York Times. Science had a big splash on it. I wasn't there, but I'm told
that the program that did best in fooling people into thinking
that it was a human was one that produced clichés at random. [LAUGHTER] Apparently, it did pretty well. I don't know. It may tell you something about
humans, but I'm not sure what. As far as I can see, all of
this is entirely pointless. It's like asking how we
could determine empirically whether an airplane
can fly, the answer being if it can fool
someone into thinking that it's an eagle, say,
under some conditions. But the question whether
an airplane can fly and whether it's really
flying or for that matter, the question whether
high jumpers were really flying at the last
Olympics, that's not an empirical question. That's a matter of decision,
and different languages make decisions differently. So if you're speaking Japanese,
I'm told airplanes, eagles, and high jumpers all
are really flying. If you're talking English,
airplanes and eagles are really flying, but humans
aren't, except metaphorically. If you're talking Hebrew,
eagles are flying, and neither airplanes
nor humans are. But that's just a matter of-- there's no substantive
issue involved. There's nothing to
this to determine, no question to answer. Simulation of an eagle could
have some purposes, of course, say, learning about
eagles or maybe solving some problem in aerodynamics. But fooling an observer under
some variety of the Turing test is not a sensible purpose,
as far as I can see. As far as I'm aware,
there's nothing like this in the sciences. Incidentally, if you go back
to the 17th and 18th centuries, scientists were very much in-- who we call philosophers,
but they didn't distinguish-- they were greatly intrigued
by automata, by machines, for example, automata
that simulated, say, the digestion of a duck. But their goal was
to learn about ducks, not to answer
meaningless questions about whether machines
can really digest, as proven by the fact that
they can fool an observer who's just, say, looking behind
a screen or something, however they may be doing it. Well, this is one
of the respects in which, in my
opinion at least, there's been regression since
the 17th and 18th century. Chess playing programs are
a perfect example of this, in my opinion. Well, anyway, that's
two approaches to the first question
of unification, the sort of deep one. What about the more
shallow question of unification, that is,
the question of unification among the various branches
of the cognitive sciences? Is there any real
commonality among them? Well, that doesn't
seem entirely obvious. So take what you might think
of, what you might call, physical organs below the
neck, metaphorically speaking, meaning not the
ones we call mental. Say, take the circulatory
system and the digestive system, the visual system, and so on. These are organs of the body in
some reasonable sense, viewed abstractly, and
that's a normal way to study any complex system. The study of these
abstract organs does fall together
at some level, we assume, presumably at the
level of cellular biology. But it's not at all clear
that there's any other level, that is, that there is
some kind of organ theory at which these
systems are unified. Maybe, but it's not obvious. Well, the brain is
scarcely understood, but to the extent that
anything's understood about it, it appears to be an
extremely intricate system, heterogeneous, with all
kind of subparts that have special design, and so on. And there's no special
reason to suppose that the mental
organs are unified, in some kind of
mental organ theory, above the level of
cellular biology, that is, above the level at
which they may fall together with the circulatory
system, and so on. If there is no
intermediate level, no such unification, then
there is no cognitive science. Hence, there was no
cognitive revolution except for similarities of
perspective, which could be quite interesting, in fact. Well, there are
various speculations about this question. And in fact, they go right
back to that 1956 meeting. One general line
of speculation is that the mind is what's
called modular, that is, that mental organs
have special design, like everything else we
know in the physical world. The other approach assumes
uniformity, that there are, what are called,
general mechanisms of learning or
thinking, which are just applied to different domains. They're applied to language
or to chess, and so on. And as I say, these differences
didn't exactly surface, but they were more or less
implicit at the 1956 meeting. The paper on generative grammar
took for granted modularity, that is, a rich, specific
genetic endowment, the language faculty's special system, and
no general mental organ theory. The Newell-Simon approach, the
study of the logic machine, that took for granted the
opposite position, that is, that there are general
mechanisms of problem solving which are quite indifferent
as to subject matter-- in our case, theorem proving,
but equivalently, chess or language, or whatever. And you can trace these
different intuitions back quite a way, quite
back 100s of years, in fact. They're very lively
today in the discussions about connectionist
work, for example. My own view is that the
evidence is overwhelmingly in favor of the highly
modular approaches. There are only a few areas
where we have any real insight, and these invariably seem to
involve quite special design. Possibly, there are some
unifying principles, but about these, if they exist,
nothing much is understood. Again, that's a minority
view, but I can only tell you the way it looks to me. Even internal to the
language faculty, there seems to be quite
highly modular design. So let's distinguish
roughly between what we might call grammatical
ability, meaning the ability to produce properly
formed expressions, and conceptual ability, that
is, to produce expressions that somehow make some sense,
and pragmatic ability, that is, the ability
to use expressions in some way appropriate
to the occasion. Well, these capacities have been
discovered to be dissociated developmentally-- like one can
develop and the other is not-- and selectively
impaired under injury. And their properties
are quite different. And they seem to be radically
different from what we find in other cognitive systems. So even internally, just
at the most gross look, we seem to find modularity. And when you look more closely
into subsystems of language, they just seem to behave
rather differently. I don't see why this
should surprise anyone. It's exactly what we find
about every other biological organism, so why shouldn't it
be true of the brain, which is maybe the most
complex one around? But it's considered a very
controversial position. Again, I stress,
as far as I know, all the evidence supports it,
and it shouldn't be surprising. Well, just how modular
is the language faculty? That's an interesting question. It doesn't seem so
special that it's linked to particular
processing systems. Here, a study of sign language
in the last decade or so, maybe 20 years, has
been quite interesting. Rather surprisingly, it
seems that sign language, the normal language of
the deaf, sign language is localized in speech
areas of the brain, in the left hemisphere and
not in the visual processing areas-- which you might
expect, since you see it. You don't hear it. That's shown pretty
well by aphasia studies and again, ERP studies. Also, it turns out
that children who are acquiring sign language
in a natural setting go through steps-- at least
in the early stages, which have been studied--
that are remarkably similar to the acquisition
of spoken language and timing and everything else. They even override
obvious iconic properties of visual symbols. So pointing is
overridden and given a symbolic interpretation. Furthermore, acquisition
of sign language turns out to be quite
different from acquisition of communicative gestures. There's even one case on
record of spontaneous invention of sign language by
three deaf children, who had no environmental
model and no stimulus at all, as far as anyone
can be determined. And that's the
perfect experiment. The system that these
three kids constructed has very standard properties
of natural language, natural spoken language, and the
developmental stages also appear to be similar. That looks like about a pure
expression of the language faculty unaffected
by experience, and it suggests that however
modular the system may be, it's not so modular that it's
keyed to a particular mode of articulation or input. What's now known suggests
that the language faculty simply grows a language, much as
other organs of the body grow. And it does so in a way largely
determined by its inner nature, by properties of
the initial state. External events doubtless
have some kind of an effect, like I'm talking
English, not Japanese. But a rational Martian
scientist who is looking at us would probably not find these
variations very impressive. And the language that
grows in the mind appears to be, at
least partially, independent of sensory modality. Well, at this point, we're
moving toward serious inquiry into modular architecture
of mind and brain, in some of the few areas where
it's possible to undertake it, and with some
surprising results. Well, one last
element of skepticism about the cognitive
revolution has to do with just how much
of a revolution it was. In fact, to a far greater extent
than was realized at that time, or is even generally
recognized today, the shift in
perspective in the 1950s recapitulated and rediscovered
some long-forgotten ideas and insights from what you
might call the first cognitive revolution, which
pretty much coincided with the scientific revolution
of the 17th and 18th centuries. Now, the Galilean-Newtonian
revolution in physics is well known. The Cartesian revolution
and its aftermath in psychology and physiology
is not very well known. But it was quite dramatic,
and it had important effects. It had important consequences. It also developed some of the
major themes that were taken up again since the 1950s, primarily
in the areas of language and vision, which
are perhaps the most successful branches of
contemporary cognitive science, as well. Now, this first
cognitive revolution, though it had a great
deal of interest, and there's much you can learn
from it, it did face barriers. And in fact, they were
insuperable barriers. So for example, in
the study of language, it came to be realized soon
enough that human language is somehow a process, not
just a dead product. It's not a bunch of texts. It's some process that goes on. Furthermore, it's a
process that makes infinite use of finite means,
as Wilhelm von Humboldt put it in the early 19th century. But no sense could be
made of these ideas, and the inquiry aborted,
aborted for well over a century. By this century,
the formal sciences had provided enough
understanding of all of this, so that many traditional themes
could be productively engaged, particularly against the
background of other advances in science and engineering and
anthropological linguistics. And one could
understand quite well what it meant to make
infinite use of finite means. Well, let's turn to
a quick look at some of the kinds of
questions that arise and some of the
kinds of answers that can be offered with respect
to language specifically. So take something
that everybody knows. Start with something really
trivial, very simple. So take a simple phrase,
like, say, a brown house. You and I have
generative procedures which are more or less similar,
and these generative procedures determine what we know
about this expression. For example, we know that
it consists of two words. We know that those two
words have the same vowel for most speakers. We know also that if
something or other, if I point to
something, and I say, that's a brown house, what I
mean is its exterior is brown. If somebody paints
a house brown, we know that they painted
the exterior brown. If they painted
the inside brown, they didn't paint
the house brown. If you see a house,
you see its exterior. So we can't see this
building, as a matter of conceptual necessity. If we were standing
outside, we might. And the same is
true of a whole mass of what are sometimes called
container words, like box or igloo or airplane. You can see an airplane, for
example, if you're inside it, but only if you can look out
the window and see the wing, or if there's a
mirror outside which reflects the exterior
of the airplane, or something like that. Then you could see the
airplane, otherwise not. The same is true with
invented concepts, even impossible concepts. So, say, take a spherical cube. If somebody paints this
spherical cube brown, that means they painted
the exterior brown, not the interior. So all of these, a house
that's [INAUDIBLE] and all these things are
exterior surfaces, which is kind of
curious to start with. However, they're
not just exteriors. So suppose two people
say John and Mary are equidistant from the
exterior regarded as a mathematical object. Suppose they're equidistant
from the exterior, but John is outside it,
and Mary's inside it. We can ask whether John, the guy
outside, is nearer the house, or whatever it is,
and there's an answer, depending on current
criteria for nearness. But we can't ask it about Mary. We're not near this building,
no matter what the criteria are. So it's not just an exterior. It's an exterior plus something
about a distinguished interior. However, the nature
of that interior doesn't seem to
matter very much. So it's the same house if
you, say, take it and fill it with cheese, or
something like that. You change the exterior,
it's the same house. It hasn't changed at all. And similarly, you can
move the walls around, and it stays the same house. On the other hand,
you can clean a house and not touch the
exterior at all. You can only do things
to the interior. So somehow, it's a very
special combination of a abstract, though
somehow concrete, interior, with an abstract exterior. And of course, the house
itself is perfectly concrete. The same is true of my home. That's also perfectly concrete,
but in a quite different way. If a house is a
brown, wooden house, it has a brown exterior
surface, but it doesn't just have a wooden exterior. It's both a concrete object
of some kind and an abstract surface, as well as having
a distinguished interior with weird properties. Well, proceeding
further, we discover that container words,
like, say, house, have extremely weird properties. Certainly, there
can't be any object in the world that has this
combination of properties, nor do we believe that there is. Rather, a word
like house provides a certain quite
complex perspective for looking at what we take
to be things in the world. Furthermore, these properties
are completely unlearned. Hence, they must be universal,
and as far as we know, they are. They're just part of our nature. We didn't learn them. We couldn't have learned them. They're also largely
unknown and unsuspected. Take the most
extensive dictionary you like, say, the big
Oxford English Dictionary, that you read with
a magnifying glass, and you find that
it doesn't dream of such somatic properties. Rather, what the
dictionary does is offer some hints that
allow a person who already knows almost everything
to pick out the intended concept. And the same is true
of words generally. And in fact, the same is true
of the sound system of language. They are largely known in
advance in all their quite remarkable intricacy. Hence, they're a universal,
a species property. The external environment
may fix some details, the way the Oxford
English Dictionary can fix a few details. But language acquisition,
and in fact, probably a good deal of what's
misleadingly called learning, is really a kind of
biological growth, where you fill in
details in an extremely rich, predetermined
structure, that just goes back to your nature-- very much like
physical growth or, as I think we should say, other
aspects of physical growth. Well, if you look
at the structure of more complex expressions,
then these conclusions are just reinforced. Again, let me take pretty
simple cases to illustrate. So take the sentence,
say, John ate an apple, and drop out the word
apple, and you get John ate. And what that means,
everybody knows, is John ate something or other. So if you eliminate a
phrase from the sentence, you interpret it as
something or other. John ate means John
ate something or other, more or less. Take the sentence, a little
bit longer, but not much-- John is too stubborn
to talk to Bill. Well, that means
John is so stubborn, that he won't talk to Bill. And just as in the case
of John ate an apple, drop the last phrase,
Bill, and we get, John is too stubborn to talk to. And by analogy, or by induction,
if such a thing existed, people ought to
interpret it as meaning John is so stubborn, that he
won't talk to someone or other. However, it doesn't mean that. John is too stubborn to talk
to means John is so stubborn, that nobody is going
to talk to him, John. Somehow it all inverts
if I drop the last word. Suppose you make it a
little more complex. John is too stubborn to
expect anyone to talk to Bill. Well, that means John-- you
have to think maybe a little bit here already, because parsing
is not easy and quick. John is too stubborn--
it means John is too stubborn for
him, John, to expect that anyone will talk to Bill. Now I'll drop Bill again. Now we get John is too stubborn
to expect anyone to talk to. And reflect for a
moment, and you'll see that the meaning
shifts radically. It means John is so stubborn,
that somebody or other doesn't expect anyone to
talk to him, John. Everything shifts. Now, all of this is inside you. Nobody could ever learn it. Nobody could possibly learn it. None of it is mentioned, even
ever hinted at, in fact, even in the most comprehensive
grammar books. It's known without experience. It's universal,
as far as we know. In fact, it better
be universal, or it means there are genetic
differences crucially among people. It just grows out
from our nature. A huge variety of
material of that sort has come to light in
the last 30 years or so as a direct consequence
of something new, namely the attempt to actually
discover and to formulate the generative
procedures that grow in the language
faculty of the brain, virtually without experience. That had never been done before. It was always assumed
that it's all obvious. How can it be anything
complicated here? If you go back, say,
to the early '50s, the problem about language
acquisition was supposed to be, why does it take so long? Why does it take so long for
such a trivial thing to happen? Why do children need so
much exposure to material to do this simple process
of habit formation? As soon as you try to
describe the facts, you discover that it's
quite the opposite. There's extremely
intricate things that are known even
in the simplest cases. There's no possible
way of picking them up from experience, and
consequently, they're universal. They're part of our nature. These are just part of
the generative procedures that the brain determines. Well, what do these generative
procedures look like? Apparently, they look
something like this. A language can be
looked at as involving a computational procedure and
what you can call a lexicon. A lexicon is a selection of
concepts that are associated with particular sounds. The concepts appear to
be pretty much invariant. There are things like
house, let's say. And they have very
weird properties, which means that they
must be invariant, because if they have
very weird properties, they've got to be
unlearned, which means that they're going
to be invariant, just part of our nature. The sound variety also
seems quite limited. On the other hand,
the association between a concept and
a sound looks free. So you can say tree in English
and [INAUDIBLE] in German, and so on. So that tells us what one aspect
of language acquisition is. It's determining which of
the given set of concepts, with all their richness
and complexity, is associated with which
of the possible sounds, with their limited variety. Notice that that's a
pretty trivial task. So you can imagine how a
little bit of experience could help fix
those associations, if the concepts were given,
and the structure of the sound was given. Well, the lexicon also
contains what are sometimes called formal elements,
things like inflections, like tense or plural or
case markers of the kind that you find, say, in
German or Latin, and so on. And languages do appear to
differ quite substantially in these respects. However, that seems
to be mostly illusion. At least, the more we learn,
the more it looks like illusion. So for example,
take, say, cases. English doesn't have them. You've got to memorize them
when you study German and Latin. But that looks illusory. It appears to be the case that
English has them fine, in fact, exactly the way Latin has them. It's just that they don't
happen to come out the mouth. They're not articulated. They're in the internal
mental computation, and their effects, which are
quite ramified, are detectable. But that part of
the computation just doesn't happen to link
up to the vocal tract. So the computation
is running along, but where it gets
spelled out of sounds, it just didn't have
to look at this stuff. Well, that's another aspect
of language acquisition, that is, to determine what aspects
of the computational system receive an external
articulation. There are also some differences
among these formal elements that also have to be fixed,
but they appear pretty narrow. Notice, incidentally,
that if you have a very intricate
system with just a few possible
differences in it, which could be right at the core
somewhere, making those changes may produce something that looks
phenomenally quite different. Just look at it
from the outside, and they look
radically different, even though it's basically the
same thing with really only trivial differences. And that's pretty much what
languages seem to look like. Well, what about the
computational system? That's the lexicon. What about the
computational system? It's possible that it's unique,
that is, it's invariant. There's just one of them. There's nothing at
all to be learned. That's not certain, but
it's at least possible, in fact plausible. If this picture is
on the right track, notice we're very
close to concluding that there's only
one human language, at least in its deeper aspects. And incidentally, going back
to this rational Martian scientist, that's what he would
have guessed, in first place. Otherwise, acquisition of
these highly intricate systems, on the basis of
fragmentary experience and the uniformity
of the process, would just be a miracle. So the natural assumption
is, well, there's basically only one of them,
and it's fixed in advance. If, in fact,
acquisition of language is something like a chicken
embryo becoming a chicken or humans undergoing
puberty at a certain age, then it's at least
within the realm of potential scientific
understanding, not a miracle. And that's the way
it seems to be. The more we learn, the
more it looks like that. More precisely,
there does seem to be a certain variety of languages,
apparently a finite variety. That is, it does seem to be
the-- or it's a good guess now that each possible language
that the language faculty makes available, each
possible language is fixed by answering a
finite number of fairly simple questions. You can look at the system as
being like a complex, wired up thing which is just given, with
a switch box associated with it and a finite number of switches. You can flip them up. You can flip them down. You do that on the basis
of the answer to one or another of these
questions, and they've got to be pretty
simple questions. If we could understand
the way that works-- bits and pieces of it
are understood-- if we suppose we could understand
it, then we ought to be able to, say,
deduce Hungarian by setting the switches one way
and deduce Swahili by setting the switches a different way. The questions must be easily
answered for empirical reasons. Children do it highly
efficiently and very fast and with very little evidence--
pretty much the way they grow. It follows that
languages are learnable. That is, if this is
correct, it follows that, as I mentioned
before, languages, in fact, are learnable, which is
an empirical discovery, and a pretty surprising one. No reason why it would have
had-- no biological reason why it would have had to be true. But it looks as
though it may be true, largely because the
variety is so restricted. Well, assuming that
to be true, we now have the following situation. Languages appear to be
learnable, which is surprising. But languages appear
to be unusable, which I don't think
is surprising at all. What that means is that
only scattered parts of them are usable. And of course, those
are the parts we use, so we don't notice it. Now, actually, this
unusability property may be somewhat deeper
than what I suggested. Since the origins of modern
generative linguistics, there have been attempts to show
that the computational system is constrained by certain very
general principles of economy, which have a kind of
global character to them. And there some recent work
that carries these ideas a long step forward. It's still pretty tentative. In fact, it's unpublished. But let me sketch some
leading ideas that indicate where things might go. Suppose we think of a linguistic
expression in something like the following way. The language faculty
selects a certain array of items from the lexicon. That's what it's
going to talk about. And it begins
computing away, using its computational
procedure-- which I'm now assuming to be
invariant across languages. It computes in parallel,
just picks out these items, and goes computing
along with them. It occasionally merges
pieces of computation, so you bring them
together, then maybe it picks something out and starts
going in parallel again. At some point in
the computation, they've all been merged. It just keeps computing. It proceeds merrily on
its way, in other words. At some point after
they've merged, the system, the
language faculty, decides to spell it out,
meaning to provide instructions to the articulatory
and perceptual system. Then the computation
just keeps going on, and it ultimately
provides something like a representation of
meaning, which is probably to be understood as instructions
for other performance systems. Well, proceeding in this
way, the computation ultimately proceeds to
paired symbolic expressions, to paired outputs, instructions
for the articulatory perceptual apparatus, on the one hand,
and instructions for language use, or called
somatic sometimes. So these are the
two, let's call them, interface levels of
the language faculty. The language faculty
seems to interface with other systems of the mind,
performance systems at two points, one having to do with
articulation and perception, the other having to do with the
things you do with language-- referring to things, asking
questions, and so on, roughly somatic. So let's say that's
what happens. It looks like it does. Now, of these
various computations, only certain of them
converge in a certain sense. The sense is that the two
kinds of outputs, the phonetic, or instructions, the
articulatory perceptual system and the somatic, actually yield
interpretable instructions-- like they might compute
along and end up with an output
that doesn't yield interpretable instructions. In that case, if either
one of them fails, we say that it doesn't converge. Now, looking just at the
convergent derivations, the convergent
computations, those that yield interpretable
paired outputs, some of them are blocked by the fact
that others of them are more economical. That is, they involve
less computation. Now, here, you've got to
define amount of computation in a very special
and precise sense. But it's not an unnatural sense. Well, if you can
do this, it's going to turn out that many
instructions that would be perfectly
interpretable just can't be constructed
by the mind, at least by the language faculty,
because the outputs are blocked by more economical derivations. What is the linguistic
expression, then? Well, the linguistic
expression is nothing other than the
optimal realization of certain external
interface conditions. A language is the setting
of certain switches, and an expression
of the language is the optimal realization of
universal interface conditions, period. All the work would now be
done by special properties of the nature of computation
and very special notions of economy, which do have
a kind of global character. They have a kind of a
least effort character, but of a global sort. Well, it turns out that quite
a variety of strange things can be explained in
these ways, in terms of a picture of language that
is really rather elegant. It's guided by conditions
of economy of derivation, with very simple and
straightforward operations. They're not only
elegant, but they're pretty surprising for
a biological system. In fact, these
properties are more like the kind, what
one expects to find, for quite unexplained reasons,
in the inorganic world. Well, these economy
conditions, as I mentioned, have a sort of a global
character to them, though not entirely. They have a tendency
to yield high degrees of computational
complexity that would render large parts
of language unusable. Well, that's not necessarily
an empirical objection. We then turn to an
empirical question. Are the parts that
are rendered unusable the parts that can't be used? We know there are plenty of
parts that can't be used. So if it turns out
that parts of language involve irresolvable
computational complexity, irresolvable by a
reasonable device, and those are the parts
you can't use, fine. That's a positive result,
not a negative result. Well, what we might discover
then, if this continues to look as promising as I think
it does now, what we might discover is that languages are
learnable, because there isn't much to learn, that they're
unusable in large measure, but that they're surprisingly
beautiful, which is just another mystery, if it
turns out to be true. Thanks. [APPLAUSE]
