harnad@phoenix.Princeton.EDU (S. R. Harnad) (05/31/90)
Abstracts for SPP Symposium On Searle's Chinese Room
and
Workshop on Symbol Grounding
Saturday, June 9, University of Maryland, College Park
(Full Program Follows in Next Posting)
For information: andrewsj@vassar.bitnet
----------------------------------------------------------------------
(1) SYNTAX, SEMANTICS AND SYSTEMS
John McCarthy <JMC@SAIL.Stanford.EDU>
Artificial Intelligence
Stanford University
John Searle begins his (1990) "Consciousness, Explanatory Inversion and
Cognitive Science" with:
"Ten years ago in this journal I published an
article (Searle, 1980a and 1980b) criticising what I
call Strong AI, the view that for a system to have
mental states it is sufficient for the system to
implement the right sort of program with right inputs
and outputs. Strong AI is rather easy to refute and
the basic argument can be summarized in one sentence:
`a system, me for example, could implement a program for
understanding Chinese, for example, without
understanding any Chinese at all.' This idea, when
developed, became known as the Chinese Room Argument."
The Chinese Room Argument can be refuted in one sentence:
"Searle confuses the mental qualities of one computational process,
himself for example, with those of another process that the first
process might be interpreting, a process that understands Chinese, for
example."
That accomplished, the lecture will discuss the ascription of mental
qualities to machines with special attention to the relation between
syntax and semantics, i.e. questions suggested by the Chinese Room
Argument. I will deal explicitly with Searle's four ``axioms'',
which, although they don't have a unique interpretation, suggest
various ideas worth discussing.
-----------------------------------------------------------
(2) SEARLE'S CHINESE ROOM
Pat Hayes <hayes@parc.xerox.com>
The basic flaw in Searle's argument is a widely accepted misunderstanding
about the nature of computers and computation: the idea that a computer
is a mechanical slave that obeys orders. This popular metaphor suggests
a major division between physical, causal hardware which acts, and
formal symbolic software, which gets read. This distinction runs
through much computing terminology, but one of the main conceptual
insights of computer science is that it is of little real scientific
importance. Computers running programs just aren't like the Chinese
room.
Software is a series of patterns which, when placed in the proper
places inside the machine, cause it to become a causally different
device. Computer hardware is by itself an incomplete specification of a
machine, which is completed - i.e. caused to quickly reshape its
electronic functionality - by having electrical patterns moved within
it. The hardware and the patterns together become a mechanism which
behaves in the way specified by the program.
This is not at all like the relationship between a reader obeying some
instructions or following some rules. Unless, that is, he has somehow
absorbed these instructions so completely that they have become part of
him, become one of his skills. The man in Searle's room who has done
this to his program now understands Chinese.
--------------------------------------------------------------------
(3)E MEASUREMENT PROBLEM IN PHSYICS AND BRAIN THEORY
H.H. Pattee
Department of Systems Science
TJ Watson School of Engineering
SUNY
Binghamton NY 13901
The measurement problem in physics is a special case of the
symbol-grounding problem of brain theory, which in turn is a special
case of the epistemological problem of relating the knower and the
known. In quantum theory the measurement problem is notoriously
obscure, because the results of measuring quantum events are
nonclassically observer-dependent and yet must be expressed only in
classical language. Even the measurement of classical events cannot be
completely described by physical laws, because measurement involves
intent, i.e., the what, where, and when of the measurements must be
determined by an "observer," not by the laws. It also makes no sense
to say that a measurement has occurred unless there is a "result."
At issue in physical theory are the necessary conditions for
"observer," and "result." Working physicists evade the philosophical
issues by using language confined to formal, operational symbol
systems that restrict measurements to numerical results and
prediction to computation. Formal symbol systems do allow unambiguous
predictions, but only at the cost of generating logical antinomies and
conceptual puzzles like "Schroedinger's Cat" and "Wigner's friend."
If one relaxes the syntactic precision of formal symbols, however, and
extends the concepts of "observer" and "result" to the subsymbolic,
dynamical level of measurement-control constraints in simpler
organisms, some of these puzzles are resolved. The problem I raise for
brain theory is that formal symbolic models of the brain may likewise
produce unambiguous results only at the cost of conceptual puzzles like
Searle's "Chinese Room."
------------------------------------------------------------------
(4) METAPHOR AND SYMBOL
David Powers <powers@informatik.uni-kl.de>
Robotics
University of Kaiserslautern FRG
The presentation is based on the monograph [Powe89] on
Machine Learning of Natural Language and Ontology,
updated with recent developments and indication of
consolidation of earlier work into a coherent approach.
Hypotheses concerning the role of contrast and
similarity, and the relationship of these mechanisms to linguistic
concepts of metaphor and paradigm, neural self-organization,
psycholinguistic paradoxes concerning negative information,
and consideration of language as part of the entire ontology
have lead to a series of experiments in machine learning of
aspects of language both individually or in combination. The work
makes use of a simulated robot world as well as textual input.
Significantly, similar results have been achieved with neural and
conventional techniques applied to the same task, with simulated
neurons being clearly associated with words and classes, and
with particular grammatical rules being associated with
particular synapses.
These results suggest three possible resolutions of the symbol
grounding problem: the symbol/non-symbol distinction
is not meaningful; neural networks can exhibit 'symbolic' behaviour
and structure; and, a sensory-motor environment can provide
grounding.
REFERENCES
[Powe89] David M. W. Powers, C. C. R. Turk,
Machine Learning of Natural Language, Springer Verlag,
London/Berlin, 1989.
-------------------------------------------------------------------
(5) CATEGORIZATION AS A PSYCHOPHYSICAL PROBLEM
Michel Treisman treisman@vax.oxford.ac.uk
Psychology Department
University of Oxford
If the same observer has to categorize the same stimulus on two
different occasions he may make different decisions each time. Why is
this? In some situations the observer will show an excessive tendency
to repeat previous responses to similar stimuli; this is sometimes
referred to as 'assimilation'. At other times he or she may avoid
previous responses: 'contrast'. Why should categorization be so
unreliable? Or what does this observation tell us about the process of
making a judgment? A psychophysical model will be outlined which
provides an explanation for these phenomena in terms of mechanisms
which tend to optimize uncertain judgments, and the relations between
different types of categorization, at different levels of complexity,
will be considered.
---------------------------------------------------------------------
(6) GROUNDING MENTAL SYMBOLS IN OBJECT IMAGES
Irving Biederman <PSYIRV@vx.acs.umn.edu> and John E. Hummel
Psychology Department
University of Minnesota
We describe a neural net (NN) implementation of a theory of real time
visual shape recognition that takes as input the edges corresponding to
the occlusional and orientation discontinuities in an image. As output
the model activates a unit that is selective for a specified
arrangement of simple volumes (or geons) and thus achieves a basic (or
entry) level classification according to Biederman's (1987)
Recognition-by-Components theory of object recognition. The output unit
can qualify as a symbol of the object in that it reflects the major
invariances of visual object recognition. The model solves four
fundamental problems in object recognition that likely confront all
attempts at visual basic-level symbol grounding:
1) Translational, size, and orientation invariance: The same output
unit(s), corresponding to the object, are activated no matter where the
image falls in the visual field, the size of the image, and the
orientation in depth (up to parts occlusion),
2) Appropriate grouping (or organization) of image elements into
appropriate parts,
3) A basis of determining invariant object centered relations (such as
TOP- OF or SIDE-CONNECTED), and
4) A basis for computing the similarity (or equivalence) of object
images.
These problems all required a solution to the "binding problem"--
determining what groups with what. In the present case, for example,
how are the various segments of the parts of an object grouped
according to their appropriate parts. Most NN models have employed
enumeration, assigning a unit to each attribute combination. Such
enumerative schemes are unsatisfactory in that they require a
prohibitively large number of units to represent even modest input
domains. Moreover, they do not express the equivalence of inputs. By
employing different units to represent the different locations of an
object, for example, the information that it is the same object in the
different locations is not represented. Our model achieves binding
through phase locking of the oscillatory activity of cells that are
tuned to oriented image edges. The phase locking (or synchrony) is
established by "fast enabling links" (FELs) between pairs of a)
collinear, b) coterminating, and c) parallel adjacent edge cells. These
units then activate invariant representations of geons and relations in
intermediate layers.
The model offers some perspective on what it is that makes a category
"basic." A category such as chair will encompass a number of
distinguishably different geon models that, perceptually, may be as
distant as members of different classes.
------------------------------------------------------
(7) INVARIANT STIMULUS FEATURES AND THE
CORTICAL REPRESENTATION OF VISUAL INFORMATION
J. Anthony Movshon tony@cortex.psych.nyu.edu
Center for Neural Science
New York University,
New York, NY 10003.
The responses of visual cortical neurons depend upon a number of
different features of the visual stimuli that fall within their
receptive fields. In most cases there is no difference in the nature
of the response produced by varying the stimulus along different
dimensions. For this reason, it is commonly recognized that the
firing of an individual neuron cannot be used unambiguously to infer
the character of a particular visual stimulus. Rather, it is necessary
to examine the distribution of activity across a population of
neurons. In considering how the multi-dimensional nature of visual
neural signals might most readily be disambiguated, it seems that
special significance might be attached to those stimulus dimensions
for which particular groups of neurons show an invariant selectivity.
An invariant selectivity is a selectivity for the value of a stimulus
along some dimension that is independent of the value of the stimulus
along other dimensions. For example, the selectivity of neurons in
the primary visual cortex (V1) for such stimulus variables as
orientation and spatial frequency is largely independent of the
precise stimulus conditions used to measure them. On the other hand,
their selectivity for the direction of motion of targets depends in a
complex way on the spatial and temporal composition of the target, and
is therefore not invariant.
In the visual cortex of the macaque monkey, many distinct visual
areas have been identified with electrophysiological and anatomical
techniques. A number of ``lower-order'' cortical areas seem to
contain neurons whose activity is primarily controlled by signals of
retinal origin - prominent among these are areas V2, V3, V4 and MT, as
well as the primary visual cortex, V1. The responses of neurons in all
these areas seem to depend on the same collection of visual stimulus
dimensions, including spatial location and size, contour orientation,
spatial frequency, chromatic composition, drift rate, direction of
motion, and binocular disparity. Neurons in different areas can have
more or less sensitivity to variations in one or another of these
parameters, so that in quantitative terms it may be argued that
signals from one area carry more information than signals from another
about particular stimulus features. It is largely on the basis of
quantitative arguments of this kind that a particular role for one or
another area in a particular aspect of visual processing has been
asserted - qualitative differences in the way that visual signals are
represented in different cortical areas have not received much
attention.
In this paper I will argue that a special significance is, in
fact, attached to the particular stimulus dimensions that are the
subjects of invariant representation within an area. For example,
despite the fact that neurons in V1 carry signals about the direction
and speed of motion of objects, the fact that their invariant
selectivity is for orientation, spatial and temporal frequency makes
it impossible for them to carry invariant information about speed and
direction. Neurons in MT, on the other hand, carry invariant
information about motion, at the expense of losing the invariant
representation of spatial and temporal parameters. Analogous
reorganization of signals about color, stereoscopic depth, and other
stimulus features may explain the existence of other representations
of the visual image in the visual cortex.
-----------------------------------------------------------
(8) COLOR AND COLOR CONSTANCY
Laurence T. Maloney ltm@xp.psych.nyu.edu
Department of Psychology
Center for Neural Science
New York University
The initial visual information that determines color appearance in
human vision depends as much on the lighting in a scene as on the
spectral properties of surfaces in the scene. A visual system that
bases color appearance on the properties of the surface, discounting
the contribution of the illuminant, is termed \fIcolor constant\fP.
I describe a class of algorithms designed to allow vision systems to
estimate information (analogous to color) about surface properties
despite changes in the illuminant. These linear model algorithms
include work by Brill, Buchsbaum, Maloney and Wandell, and others.
These algorithms share strong assumptions about the range of possible
illuminants and possible surface reflectances present in a scene. I
describe evidence suggesting that many common surfaces and illuminants
satisfy the constraints required by linear model algorithms.
Hilbert (1987, Chap. 7) discusses the consequences of this work
for philosophy.
Maloney, L. T., and Wandell, B. A., Color constancy: A computational
method for recovering surface spectral reflectance. Journal of the
Optical Society of America A, 1986, \fB3\fP, 29-33.
Maloney, L. T., Evaluation of linear models of surface spectral
reflectance with small numbers of parameters. Journal of the Optical
Society of America A, 1986, \fB3\fP, 1673-1683.
Hilbert, D. R., \fIColor and Color Perception; A Study in
Anthropocentric Realism.\fP (Stanford, CA: CSLI, 1987).
---------------------------------------------------
(9) PERCEPTUAL MEMORY CATEGORIZATION IN PRIMARY SENSORY CORTEX
Richard Granger granger@ICS.UCI.EDU
Center for the Neurobiology of Learning and Memory
University of California, Irvine
Recent results from neurobiological simulation work have led to a novel
hypothesis: that the physiological operation of a primary sensory
cortical area (olfactory (piriform) cortex) automatically organizes
learned perceptual cues into a hierarchical memory (Ambros-Ingerson,
Granger and Lynch, Science, 1990). In the simulation, repetitive
perceptual samples ("sniffs", or "glances") of learned cues traverse
the constructed hierarchy, such that initial samples yield relatively
coarse-grained category responses whereas later samples yield
increasingly finer-grained information about the cue. The resulting
iterative recognition of cues shares many characteristics with the
robust psychological phenomenon of "basic levels": within a
hierarchically nested set of categories such as "animal-bird-robin",
there is a specific level of abstraction that is more readily processed
(e.g., recognized faster) than the others; "bird" in this example
(Mervis and Rosch, Ann.Rev.Psych., 1981). The correspondence raises the
possibility that aspects of this psychological phenomenon may arise
from fundamental physiological mechanisms in primary sensory cortex.
-------------------------------------------------------------------
(10) ICONS AND TEMPORAL PATTERNS:
A DYNAMIC CONNECTIONIST SOLUTION TO SYMBOL GROUNDING
Harnad (1990) proposes that categories can be grounded by their direct
relationship to a physical icon of the input stimulation. The notion of
an icon has a clear meaning in the case of a visual display: it is a
pattern of activity in a field of neurons that is physically isomorphic
with the pattern of light in 2D. The standard proposal for a physical
icon of TIME is physical distance. Such a model is naturally
implemented by delay lines in a network (see Lang, 1990, NN).
But delay lines are not a good model for human behavior. When subjects
are trained on a complex temporal pattern, like a random sequence of
tones, they can develop a detailed perceptual representation (Spiegel &
Watson 1981, Watson and Foyle, 1985, JASA). Some skills that should be
easy are very difficult -- eg, recognizing an absolute time interval in
the face of randomly varying intervening sounds. This should be easy
since absolute time differences are represented by weights on specific
delay lines. So edges from inputs that have a random relation to the
categorical identification should learn random weights. On the other
hand, skills that should be very difficult turn out to be easy, such as
detecting serial order of familiar patterns in the face of changes in
rate of presentation. This should be difficult since a pattern that
appears at different rates will be distributed differently across the
range of delays, and should thus be learned only slowly.
On other hand, we have been developing dynamic network models that
represent learned temporal patterns of tones as stable equilibria in
the activation space of a group of fully recurrent nodes
(Port-Anderson, 1990, Anderson-Port, 1990). These systems were trained
(with real-time recurrent learning) to recognize particular tone
sequences. They are highly resistant to noise and continue to recognize
patterns even when the rate of presentation is varied by a factor of 2
faster or slower. Watson's results and our simulations suggest that
brains do not produce an icon of auditory patterns in time. It implies
that direct contact with stimulation that is distributed in time is not
possible. Although I do not disagree with Harnad on the importance of
symbol grounding for an account of perception, apparently, the
grounding of categories does not require an `icon' in the sense that
Harnad has proposed.
-----------------------------------------------------------------
--
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