harnad@phoenix.Princeton.EDU (S. R. Harnad) (05/31/90)
The following three articles on Searle's Chinese Room Argument
and the Symbol Grounding Problem are available by anonymous
ftp from phoenix.princeton.edu in directory /a/ftp/pub/harnad
(see end of message for ftp instructions):
(1) THE SYMBOL GROUNDING PROBLEM
[Physica D 1990, in press]
Stevan Harnad
Department of Psychology
Princeton University
ABSTRACT: There has been much discussion recently about the scope and
limits of purely symbolic models of the mind and about the proper role
of connectionism in cognitive modeling. This paper describes the
"symbol grounding problem" for a semantically interpretable symbol
system: How can its semantic interpretation be made intrinsic to the
symbol system, rather than just parasitic on the meanings in our heads?
How can the meanings of the meaningless symbol tokens, manipulated
solely on the basis of their (arbitrary) shapes, be grounded in
anything but other meaningless symbols? The problem is analogous to
trying to learn Chinese from a Chinese/Chinese dictionary alone.
A candidate solution is sketched: Symbolic representations must be
grounded bottom-up in nonsymbolic representations of two kinds:
(1) iconic representations, which are analogs of the proximal sensory
projections of distal objects and events, and (2) categorical
representations, which are learned and innate feature-detectors that
pick out the invariant features of object and event categories from
their sensory projections. Elementary symbols are the names of these
object and event categories, assigned on the basis of their
(nonsymbolic) categorical representations. Higher-order (3) symbolic
representations, grounded in these elementary symbols, consist of
symbol strings describing category membership relations ("An
X is a Y that is Z").
Connectionism is one natural candidate for the mechanism that learns
the invariant features underlying categorical representations, thereby
connecting names to the proximal projections of the distal objects they
stand for. In this way connectionism can be seen as a complementary
component in a hybrid nonsymbolic/symbolic model of the mind, rather
than a rival to purely symbolic modeling. Such a hybrid model would not
have an autonomous symbolic "module," however; the symbolic functions
would emerge as an intrinsically "dedicated" symbol system as a
consequence of the bottom-up grounding of categories' names in their
sensory representations. Symbol manipulation would be governed not just
by the arbitrary shapes of the symbol tokens, but by the nonarbitrary
shapes of the icons and category invariants in which they are grounded.
---------------------------------------------------------------------
(2) MINDS MACHINES AND SEARLE
J Exper. Theor. A.I. 1(1) pp. 5 - 25 (1989)
Stevan Harnad
Department of Psychology
Princeton University
Princeton NJ 08544
SUMMARY: Searle's celebrated Chinese Room Argument has shaken the
foundations of Artificial Intelligence. Many refutations have been
attempted, but none seem convincing. This paper is an attempt to sort
out explicitly the assumptions and the logical, methodological and
empirical points of disagreement. Searle is shown to have
underestimated some features of computer modeling, but the heart of the
issue turns out to be an empirical question about the scope and limits
of the purely symbolic (computational) model of the mind. Nonsymbolic
modeling turns out to be immune to the Chinese Room Argument. The
issues discussed include the Total Turing Test, modularity, neural
modeling, robotics, causality and the symbol-grounding problem.
Summary and Conclusions
Searle's provocative "Chinese Room Argument" attempted to show that the
goals of "Strong AI" are unrealizable. Proponents of Strong AI are
supposed to believe that (i) the mind is a computer program, (ii) the
brain is irrelevant, and (iii) the Turing Test is decisive. Searle's
argument is that since the programmed symbol-manipulating instructions
of a computer capable of passing the Turing Test for understanding
Chinese could always be performed instead by a person who could not
understand Chinese, the computer can hardly be said to understand
Chinese. Such "simulated" understanding, Searle argues, is not the same
as real understanding, which can only be accomplished by something that
"duplicates" the "causal powers" of the brain. In the present paper the
following points have been made:
(1) Simulation versus Implementation:
Searle fails to distinguish between the simulation of a mechanism,
which is only the formal testing of a theory, and the implementation of
a mechanism, which does duplicate causal powers. Searle's "simulation"
only simulates simulation rather than implementation. It can no more be
expected to understand than a simulated airplane can be expected to
fly. Nevertheless, a successful simulation must capture formally all
the relevant functional properties of a successful implementation.
(2) Theory-Testing versus Turing-Testing:
Searle's argument conflates theory-testing and Turing-Testing.
Computer simulations formally encode and test models for
human perceptuomotor and cognitive performance capacities; they
are the medium in which the empirical and theoretical work
is done. The Turing Test is an informal and open-ended test
of whether or not people can discriminate the performance
of the implemented simulation from that of a real human being.
In a sense, we are Turing-Testing one another all the time, in
our everyday solutions to the "other minds" problem.
(3) The Convergence Argument:
Searle fails to take underdetermination into account. All scientific
theories are underdetermined by their data; i.e., the data are
compatible with more than one theory. But as the data domain grows, the
degrees of freedom for alternative (equiparametric) theories shrink.
This "convergence" constraint applies to AI's "toy" linguistic and
robotic models too, as they approach the capacity to pass the Total
(asymptotic) Turing Test. Toy models are not modules.
(4) Brain Modeling versus Mind Modeling:
Searle also fails to appreciate that the brain itself can be understood
only through theoretical modeling, and that the boundary between brain
performance and body performance becomes arbitrary as one converges on
an asymptotic model of total human performance capacity.
(5) The Modularity Assumption:
Searle implicitly adopts a strong, untested "modularity" assumption to
the effect that certain functional parts of human cognitive performance
capacity (such as language) can be be successfully modeled
independently of the rest (such as perceptuomotor or "robotic"
capacity). This assumption may be false for models approaching the
power and generality needed to pass the Turing Test.
(6) The Teletype Turing Test versus the Robot Turing Test:
Foundational issues in cognitive science depend critically on the truth
or falsity of such modularity assumptions. For example, the "teletype"
(linguistic) version of the Turing Test could in principle (though not
necessarily in practice) be implemented by formal symbol-manipulation
alone (symbols in, symbols out), whereas the robot version necessarily
calls for full causal powers of interaction with the outside world
(seeing, doing AND linguistic competence).
(7) The Transducer/Effector Argument:
Prior "robot" replies to Searle have not been principled ones. They
have added on robotic requirements as an arbitrary extra constraint. A
principled "transducer/effector" counterargument, however, can be based
on the logical fact that transduction is necessarily nonsymbolic,
drawing on analog and analog-to-digital functions that can only be
simulated, but not implemented, symbolically.
(8) Robotics and Causality:
Searle's argument hence fails logically for the robot version of the
Turing Test, for in simulating it he would either have to USE its
transducers and effectors (in which case he would not be simulating all
of its functions) or he would have to BE its transducers and effectors,
in which case he would indeed be duplicating their causal powers (of
seeing and doing).
(9) Symbolic Functionalism versus Robotic Functionalism:
If symbol-manipulation ("symbolic functionalism") cannot in principle
accomplish the functions of the transducer and effector surfaces, then
there is no reason why every function in between has to be symbolic
either. Nonsymbolic function may be essential to implementing minds and
may be a crucial constituent of the functional substrate of mental
states ("robotic functionalism"): In order to work as hypothesized
(i.e., to be able to pass the Turing Test), the functionalist
"brain-in-a-vat" may have to be more than just an isolated symbolic
"understanding" module -- perhaps even hybrid analog/symbolic all the
way through, as the real brain is, with the symbols "grounded"
bottom-up in nonsymbolic representations.
(10) "Strong" versus "Weak" AI:
Finally, it is not at all clear that Searle's "Strong AI"/"Weak AI"
distinction captures all the possibilities, or is even representative
of the views of most cognitive scientists. Much of AI is in any case
concerned with making machines do intelligent things rather than with
modeling the mind.
Hence, most of Searle's argument turns out to rest on unanswered
questions about the modularity of language and the scope and limits of
the symbolic approach to modeling cognition. If the modularity
assumption turns out to be false, then a top-down symbol-manipulative
approach to explaining the mind may be completely misguided because its
symbols (and their interpretations) remain ungrounded -- not for
Searle's reasons (since Searle's argument shares the cognitive
modularity assumption with "Strong AI"), but because of the
transdsucer/effector argument (and its ramifications for the kind of
hybrid, bottom-up processing that may then turn out to be optimal, or
even essential, in between transducers and effectors). What is
undeniable is that a successful theory of cognition will have to be
computable (simulable), if not exclusively computational
(symbol-manipulative). Perhaps this is what Searle means (or ought to
mean) by "Weak AI."
------------------------------------------------------------------
CATEGORY INDUCTION AND REPRESENTATION
[From "Categorical Perception: The Groundwork of Cognition"
Cambridge University Press 1987, S. Harnad, Ed.]
Stevan Harnad
Psychology Department
Princeton University
Princeton NJ 08544
SUMMARY: Categorization is a very basic cognitive activity. It is
involved in any task that calls for differential responding, from
operant discrimination to pattern recognition to naming and describing
objects and states-of-affairs. Explanations of categorization range
from nativist theories denying that any nontrivial categories are
acquired by learning to inductivist theories claiming that most
categories are learned.
"Categorical perception" (CP) is the name given to a suggestive
perceptual phenomenon that may serve as a useful model for
categorization in general: For certain perceptual categories,
within-category differences look much smaller than between-category
differences even when they are of the same size physically. For
example, in color perception, differences between reds and differences
between yellows look much smaller than equal-sized differences that
cross the red/yellow boundary; the same is true of the phoneme
categories /ba/ and /da/. Indeed, the effect of the category boundary
is not merely quantitative, but qualitative.
There have been two theories to explain CP effects. The "Whorf
Hypothesis" explains color boundary effects by proposing that language
somehow determines our view of reality. The "motor theory of speech
perception" explains phoneme boundary effects by attributing them to
the patterns of articulation required for pronunciation. Both theories
seem to raise more questions than they answer, for example: (i) How
general and pervasive are CP effects? Do they occur in other modalities
besides speech-sounds and color? (ii) Are CP effects inborn or can they
be generated by learning (and if so, how)? (iii) How are categories
internally represented? How does this representation generate
successful categorization and the CP boundary effect?
Some of the answers to these questions will have to come from ongoing
research, but the existing data do suggest a provisional model for
category formation and category representation. According to this
model, CP provides our basic or elementary categories. In acquiring a
category we learn to label or identify positive and negative instances
from a sample of confusable alternatives. Two kinds of internal
representation are built up in this learning by "acquaintance": (1) an
ICONIC representation that subserves our similarity judgments and (2)
an analog/digital feature-filter that picks out the invariant
information allowing us to categorize the instances correctly. This
second, CATEGORICAL representation is associated with the category
name. Category names then serve as the atomic symbols for a third
representational system, the (3) SYMBOLIC representations that underlie
language and that make it possible for us to learn by "description."
Connectionism is one possible mechainsm for learning the sensory
invariants underlying categorization and naming.
Among the implications of the model are (a) the "cognitive identity of
(current) indiscriminables": Categories and their representations can
only be provisional and approximate, relative to the alternatives
encountered to date, rather than "exact." There is also (b) no such
thing as an absolute "feature," only those features that are invariant
within a particular context of confusable alternatives. Contrary to
prevailing "prototype" views, however, (c) such provisionally invariant
features MUST underlie successful categorization, and must be
"sufficient" (at least in the "satisficing" sense) to subserve reliable
performance with all-or-none, bounded categories, as in CP. Finally,
the model brings out some basic limitations of the
"symbol-manipulative" approach to modeling cognition, showing how (d)
symbol meanings must be functionally grounded in nonsymbolic,
"shape-preserving" representations -- iconic and categorical ones.
Otherwise, all symbol interpretations are ungrounded and indeterminate.
This amounts to a principled call for a psychophysical (rather than a
neural) "bottom-up" approach to cognition.
-------------------------------------------------------------
To retrieve a file by ftp them from a Unix/Internet site, type:
ftp phoenix.princeton.edu
When you are asked for your login, type:
anonymous
For your password, type:
ident
then change directories with:
cd pub/harnad
Then type:
binary
(This is for retrieving compressed files.)
To show the available files, type:
ls
Next, retrieve the file you want with (for example):
get filename.Z
When you have the file(s) you want, type:
quit
Next uncompress the file with:
uncompress filename.Z
Now the file will be called, simply,
filename
Finally, read or print it.
I also maintain an email group for discussion of the symbol grounding
problem. If you want to be added to the list, please write
harnad@clarity.princeton.edu
Stevan Harnad
--
Stevan Harnad Department of Psychology Princeton University
harnad@clarity.princeton.edu srh@flash.bellcore.com
harnad@elbereth.rutgers.edu harnad@pucc.bitnet (609)-921-7771