Professional Documents
Culture Documents
196
High-Level Perception, Representation, and Analogy
197
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
top-down influences in perception, claiming structure, this would leave unresolved two
that perceptual processes are "cognitively important problems, namely:
impenetrable" or "informationally encapsu- The problem of relevance: How is it de-
lated". These arguments are highly contro- cided which subsets of the vast amounts of
versial, but in any case they apply mostly to data from the environment get used in vari-
relatively low-level sensory perception.Few ous parts of the representational
would dispute that at the higher, conceptual structure?Naturally, much of the information
level of perception, top-down and contextu- content at the lowest level will be quite irrel-
al influences play a large role. evant at the highest representational level.To
determine which parts of the data are rele-
2 ARTIFICIAL INTELLIGENCE AND vant to a given representation, a complex
THE PROBLEM OF filtering process is required.
REPRESENTATION The problem of: How are these data put
into the correct form for the representation?
The end product of the process of per- Even if we have determined precisely which
ception, when a set of raw data has been data are relevant, and we have determined
organized into a coherent and structured the desired framework for the representa-
whole, is a representation.Representations tion- a frame-based representation, for in-
have been the object of much study and de- stance C we still face the problem of orga-
bate within the field of artificial intelligence, nizing the data into the representational form
and much is made of the "representation in a useful way.The data do not come pre-
problem".This problem has traditionally been packaged as slots and fillers, and organizing
phrased as "What is the correct structure for them into a coherent structure is likely to be
mental representations?", and many possi- a highly non-trivial task.
bilities have been suggested, ranging from These questions, taken together, amount
predicate calculus through frames and scripts in essence to the problem of high-level per-
to semantic networks and more. We may di- ception, translated into the framework of
vide representations into two kinds: long- artificial intelligence.
term knowledge representations that are The traditional approach in artificial in-
stored passively somewhere in the system, telligence has been to start by selecting not
and short-term representations that are ac- only a preferred type of high-level represen-
tive at a given moment in a particular men- tational structure, but also the data assumed
tal or computational process. (This distinc- to be relevant to the problem at hand. These
tion corresponds to the distinction between data are organized by a human programmer
long-term memory and working memory.) who appropriately fits them into the chosen
In this discussion, we will mostly be con- representational structure. Usually, research-
cerned with short-term, active representa- ers use their prior knowledge of the nature
tions, as it is these that are the direct prod- of the problem to hand-code a representa-
uct of perception. tion of the data into a near-optimal form.
The question of the structure of repre- Only after all this hand-coding is completed
sentations is certainly an important one, but is the representation allowed to be manipu-
there is another, related problem that has not lated by the machine. The problem of repre-
received nearly as much attention.This is that sentation-formation, and thus the problem
of understanding how such a representation of high-level perception, is ignored. (These
could be arrived at, starting from environ- comments do not, of course, apply to work
mental data.Even if it were possible to dis- in machine vision, speech processing, and
cover an optimal type of representational other perceptual endeavors. However, work
198
High-Level Perception, Representation, and Analogy
in these fields usually stops short of model- surface for inscription.... But if I wished to
ing processes at the conceptual level and is light a fire, and no other materials were by,
thus not directly relevant to our critique of the essential way of conceiving the paper
high-level cognitive modeling.) would be as a combustible material..,. The
The formation of appropriate represen- essence of a thing is that one of its proper-
tations lies at the heart of human high-level ties which is so important for my interests
cognitive abilities. It might even be said that that in comparison with it I may neglect the
the problem of high-level perception forms rest.... The properties which are important
the central task facing the artificial-intelli- vary from man to man and from hour to hour.
gence community: the task of understanding ... many objects of daily use - as paper, ink,
how to draw meaning out of the world. It butter, overcoat - have properties of such
might not be stretching the point to say that constant unwavering importance, and have
there is a "meaning barrier", which has rare- such stereotyped names, that we end by be-
ly been crossed by work in AI. On one side lieving that to conceive them in those ways
of the barrier, some models in low-level per- is to conceive them in the only true way.
ception have been capable of building prim- Those are no truer ways of conceiving them
itive representations of the environment, but than any others; there are only more fre-
these are not yet sufficiently complex to be quently serviceable ways to us." (James 1890,
called "meaningful". On the other side of the pp. 222-224)
barrier, much research in high-level cogni- James is saying, effectively, that we have
tive modeling has started with representations different representations of an object or sit-
at the conceptual level, such as propositions uation at different times. The representation-
in predicate logic or nodes in a semantic al process adapts to fit the pressures of a giv-
network, where any meaning that is present en context.
is already built in. There has been very little Despite the work of philosopher-psychol-
work that bridges the gap between the two. ogists such as James, the early days of artifi-
cial intelligence were characterized by an
OBJECTIVISM AND TRADITIONAL AI objectivist view of perception, and of the rep-
resentation of objects, situations, and cate-
Once AI takes the problem of represen- gories. As the linguist George Lakoff has
tation-formation seriously, the next stage will characterized it, "On the objectivist view,
be to deal with the evident flexibility of hu- reality comes complete with a unique cor-
man high-level perceptual processes. As we rect, complete structure in terms of entities,
have seen, objects and situations can be com- properties and relations. This structure ex-
prehended in many different ways, depend- ists, independent of any human understand-
ing on context and top-down influences. We ing." (Lakoff 1987, p. 159) While this objec-
must find a way of ensuring that AI repre- tivist position has been unfashionable for
sentations have a corresponding degree of decades in philosophical circles (especially
flexibility. William James, in the late nine- after Wittgenstein's work demonstrating the
teenth century, recognized this aspect of cog- inappropriateness of a rigid correspondence
nitive representations: between language and reality), most early
'There is no property ABSOLUTELY work in AI implicitly accepted this set of as-
essential to one thing. The same property sumptions.
which figures as the essence of a thing on The Physical Symbol System Hypothesis
one occasion becomes a very inessential fea- (Newell & Simon 1976), upon which most
ture upon another. Now that I am writing, it of the traditional AI enterprise has been built,
is essential that I conceive my paper as a posits that thinking occurs through the ma-
199
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
nipulation of symbolic representations, which ers" that can respond to diverse aspects of
are composed of atomic symbolic primitives. various situations.
Such symbolic representations are by their In these models, a flexible perceptual pro-
nature somewhat rigid, black-and-white en- cess has been integrated with an equally flex-
tities, and it is difficult for their representa- ible dependence of action upon representa-
tional content to shift subtly in response to tional content, yielding models that respond
changes in context. The result, in practice - to diverse situations with a robustness that is
irrespective of whether this was intended by difficult to match with traditional methods.
the original proponents of this framework - Nonetheless, the models are still somewhat
is a structuring of reality that tends to be as primitive, and the representations they de-
fixed and absolute as that of the objectivist velop are not nearly as complex as the hand-
position outlined above. coded, hierarchically-structured representa-
By the mid-seventies, a small number of tions found in traditional models; still, it
AI researchers began to argue that in order seems to be a step in the right direction. It
to progress, the field would have to part ways remains to be seen whether work in more
with its commitment to such a rigid repre- traditional AI paradigms will respond to this
sentational framework. One of the strongest challenge by moving toward more flexible
early proponents of this view was David and robust representational forms.
Marr, who noted that "the perception of an
event or object must include the simultaneous ON THE POSSIBILITY OF A
computation of several different descriptions REPRESENTATION MODULE
of it, that capture diverse aspects of the use,
purpose or circumstances of the event or It might be granted that given the diffi-
object." (Marr 1977, p. 44) culty of the problem of high-level percep-
Recently, significant steps have been tak- tion, AI researchers could be forgiven for
en toward representational flexibility with the starting with their representations in a made-
advent of sophisticated connectionist mod- to-order form. They might plausibly claim
els whose distributed representations are that the difficult problem of representation-
highly context-dependent (Rumelhart & formation is better left until later. But it must
McClelland 1986). In these models, there are be realized that behind this approach lies a
no representational primitives in internal tacit assumption: that it is possible to model
processing. Instead, each representation is a high-level cognitive processes independent-
vector in a multi-dimensional space, whose ly of perceptual processes. Under this as-
position is not anchored but can adjust flex- sumption, the representations that are cur-
ibly to changes in environmental stimuli. rently, for the most part, tailored by human
Consequently, members of a category are not hands, would eventually be built up by a sep-
all represented by identical symbolic struc- arate lower-level facility C a "representation
tures; rather, individual objects will be rep- module" whose job it would be to funnel data
resented in subtly different ways depending into representations. Such a module would
upon the context in which they are present- act as a "front end" to the models of the cog-
ed. In networks with recurrent connections nitive processes currently being studied, sup-
(Elman 1990), representations are even sen- plying them with the appropriately-tailored
sitive to the current internal state of the mod- representations.
el. Other recent work taking a flexible ap- We are deeply skeptical, however, about
proach to representation includes the classi- the feasibility of such a separation of per-
fier-system models of Holland (1986) and ception from the rest of cognition. A repre-
his colleagues, where genetically-inspired sentation module that, given any situation,
methods are used to create a set of "classifi- produced the single "correct" representation
200
Iligh-Level Perception, Representation, and Analogy
for it, would have great difficulty emulating ables it to discover complex laws involving
the flexibility that characterizes human per- many terms". BACON was able to "discov-
ception. For such flexibility to arise, the rep- er", among other things, Boyle's law of ideal
resentational processes would have to be gases, Kepler's third law of planetary mo-
sensitive to the needs of all the various cog- tion, Galileo's law of uniform acceleration,
nitive processes in which they might be used. and Ohm's law.
It seems most unlikely that a single repre- Such claims clearly demand close scruti-
sentation would suffice for all purposes. As ny. We will look in particular at the program's
we have seen, for the accurate modeling of "discovery" of Kepler's third law of plane-
cognition it is necessary that the representa- tary motion. Upon examination, it seems that
tion of a given situation can vary with vari- the success of the program relies almost en-
ous contextual and top-down influences. tirely on its being given data that have al-
This, however, is directly contrary to the "rep- ready been represented in near-optimal form,
resentation module" philosophy, wherein rep- using after-the-fact knowledge available to
resentations are produced quite separately the programmers.
from later cognitive processes, and then sup- When BACON performed its derivation
plied to a "task-processing" module. of Kepler's third law, the program was given
To separate representation-building from only data about the planets' average distances
higher-level cognitive tasks is, we believe, from the sun and their periods. These are
impossible. In order to provide the kind of precisely the data reqiredto drive the law. The
flexibility that is apparent in cognition, any program is certainly not "starting with essen-
fully cognitive model will probably require a tially the same initial conditions as the hu-
continual interaction between the process of man discoverers", as one of the authors of
representation-building and the manipulation BACON has claimed (Simon 1989, p. 375).
of those representations. If this proves to be The authors'claim that BACON used "orig-
the case, then the current approach of using inal data" certainly docs not mean that it used
hand-coded representations not only is post- ALL of the data available to Kepler at the
poning an important issue but will, in the long time of his discovery, the vast majority of
run, lead up a dead-end street. which were irrelevant, misleading, distract-
We will consider this issue in greater ing, or even wrong.
depth later, when we discuss current research This pre-selection of data may at first
in the modeling of analogical thought. For seem quite reasonable: after all, what could
now, we will discuss in some detail one well- be more important to an astronomer-math-
known AI program for which great claims ematician than planetary distances and peri-
have been made. We argue that these claims ods? But here our after-the-fact knowledge
represent a lack of appreciation of the im- is misleading us. Consider for a moment the
portance of high-level perception. times in which Kepler lived. It was the turn
of the seventeenth century, and Copernicus'
BACON: A CASE STUDY De Revolutionibus Orbium Coelestium was
still new and far from universally accepted.
A particularly clear case of a program in Further, at that time there was no notion of
which the problem of representation is by- the forces that produced planetary motion;
passed is BACON, a well-known program the sun, in particular, was known to produce
that has been advertised as an accurate model light but was not thought to influence the
of scientific discovery (Langley et al 1987). motion of the planets. In that prescientific
The authors of BACON claim that their sys- world, even the notion of using mathemati-
tem is "capable of representing information cal equations to express regularities in na-
at multiple levels of description, which en- ture was rare. And Kepler believed - in fact,
201
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
his early fame rested on the discovery of this of the human nervous system (!). In an in-
surprising coincidence - that the planets' dis- teresting juxtaposition to this, researchers
tances from the sun were dictated by the fact in a recent study (Qin & Simon 1990) found
that the five regular polyhedra could be fit that starting with the data that BACON was
between the five "spheres" of planetary mo- given, university students could make essen-
tion around the sun, a fact that constituted tially the same "discoveries" within an hour-
seductive but ultimately misleading data. long experiment. Somewhat strangely, the
Within this context, it is hardly surprising authors (including one of the authors of
that it took Kepler thirteen years to realize BACON) take this finding to support the
that conic sections and not Platonic solids, plausibility of BACON as an accurate mod-
that algebra and not geometry, that ellipses el of scientific discovery. It seems more rea-
and not Aristotelian "perfect" circles, that the sonable to regard it as a demonstration of
planets' distances from the sun and not the the vast difference in difficulty between the
polyhedra in which they fit, were the rele- task faced by BACON and that faced by
vant factors in unlocking the regularities of Kepler, and thus as a reductio ad absurdum
planetary motion. In making his discoveries, of the BACON methodology.
Kepler had to reject a host of conceptual So many varieties of data were available
frameworks that might, for all he knew, have to Kepler, and the available data had so many
applied to planetary motion, such as religious different ways of being interpreted, that it is
symbolism, superstition, Christian cosmolo- difficult not to conclude that in presenting
gy, and teleology. In order to discover his their program with data in such a neat form,
laws, he had to make all of these creative the authors of BACON are inadvertently
leaps. BACON, of course, had to do nothing guilty of 20-20 hindsight. BACON, in short,
of the sort. The program was given precisely works only in a world of hand-picked, pre-
the set of variables it needed from the outset structured data, a world completely devoid
(even if the values of some of these variables of the problems faced by Kepler or Galileo
were sometimes less than ideal), and was or Ohm when they made their original dis-
moreover supplied with precisely the right coveries. Similar comments could be made
biases to induce the algebraic form of the about STAHL, GLAUBER, and other mod-
laws, it being taken completely for granted els of scientific discovery by the authors of
that mathematical laws of a type now recog- BACON. In all of these models, the crucial
nized by physicists as standard were the de- role played by high-level perception in sci-
sired outcome. entific discovery, through the filtering and
It is difficult to believe that Kepler would organization of environmental stimuli, is ig-
have taken thirteen years to make his dis- nored.
covery if his working data had consisted en- It is interesting to note that the notion of
tirely of a list where each entry said "Planet a "paradigm shift", which is central to much
X: Mean Distance from Sun Y, Period Z". If scientific discovery (Kuhn 1970), is often
he had further been told "Find a polynomial regarded as the process of vieving the world
equation relating these entities", then it might in a radically different way. That is, scien-
have taken him a few hours. Addressing the tists' frameworks for representing available
question of why Kepler took thirteen years world knowledge are broken down, and their
to do what BACON managed within min- high-level perceptual abilities are used to
utes, Langley et al (1987) point to "sleeping organize the available data quite differently,
time, and time for ordinary daily chores", and building a novel representation of the data.
other factors such as the time taken in set- Such a new representation can be used to
ting up experiments, and the slow hardware draw different and important conclusions in
High-Level Perception. Representation, and Analogy
a way that was difficult or impossible with Let us consider two analogies involving
the old representation. In this model of sci- DNA. The first is an analogy between DNA
entific discovery, unlike the model present- and a zipper. When we are presented with
ed in BACON, the process of high-level per- this analogy, the image of DNA that comes
ception is central. to mind is that of two strands of paired nu-
The case of BACON is by no means iso- cleotides (which can come apart like a zip-
lated - it is typical of much work in AI, which per for the purposes of replication). The sec-
often fails to appreciate the importance of ond analogy involves comparing DNA to the
the representation-building stage. We will see source code (i.e., non-executable high-level
this in more depth in the next section, in code) of a computer program. What comes
which we take a look at the modeling of anal- to mind now is the fact that information in
ogy. the DNA gets "compiled" (via processes of
transcription and translation) into enzymes,
3 MODELS OF ANALOGICAL which correspond to machine code (i.e., ex-
THOUGHT ecutable code). In the latter analogy, the
perception of DNA is radically different - it
Analogical thought is dependent on high- is represented essentially as an information-
level perception in a very direct way. When bearing entity, whose physical aspects, so
people make analogies, they are perceiving important to the first analogy, are of virtual-
some aspects of the structures of two situa- ly no consequence.
tions C the essences of those situations, in In cases such as these, it seems that no
some sense C as identical. These structures, single, rigid representation can capture what
of course, are a product of the process of is going on in our heads. It is tme that we
high-level perception. probably have a single rich representation
The quality of an analogy between two of DNA sitting passively in long-term mem-
situations depends almost entirely on one's ory. However, in the contexts of different
perception of the situations. If Ronald Re- analogical mappings, very different facets of
agan were to evaluate the validity of an anal- this large representational structure are se-
ogy between the U.S. role in Nicaragua and lected out as being relevant, by the pressures
the Soviet Union's role in Afghanistan, he of the particular context. Irrespective of the
would undoubtedly see it as a poor one. Oth- passive content of the long-term representa-
ers might consider the analogy excellent. The tion of DNA, the active content that is pro-
difference would come from different per- cessed at a given time is determined by a
ceptions, and thus representations, of the sit- flexible representational process.
uations themselves. Reagan's internal rep- Furthermore, not only is analogy-making
resentation of the Nicaraguan situation is dependent on high-level perception, but the
certainly quite different from Daniel Orte- reverse holds true as well: perception is of-
ga's. ten dependent on analogy-making itself. The
Analogical thought further provides one high-level perception of one situation in
of the clearest illustrations of the flexible terms of another is ubiquitous in human
nature of our perceptual abilities. Making an thought. If we perceive Nicaragua as "another
analogy requires highlighting various differ- Vietnam", for example, the making of the
ent aspects of a situation, and the aspects analogy is fleshing out our representation of
that are highlighted are often not the most Nicaragua. Analogical thought provides a
obvious features. The perception of a situa- powerful mechanism for the enrichment of
tion can change radically, depending on the a representation of a given situation. This is
analogy we are making. well understood by good educators and writ-
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
ers, who know that there is nothing like an ping; but each mapping process requires a
analogy to provide a better mental picture perceptual process to precede it, whereas it
of a given situation. Analogies affect our is not the case that each perceptual process
perception all the time: in a love affair, for necessarily depends upon mapping. There-
instance, it is difficult to stop parallels with fore the perceptual process is conceptually
past romances from modulating one's per- prior, although perception and mapping pro-
ception of the current situation. In the large cesses are often temporally interwoven. If the
or the small, such analogical perception - appropriate representations are already
the grasping of one situation in terms of an- formed, the mapping process can often be
other - is so common that we tend to forget quite straightforward. In our view, the most
that what is going on is, in fact, analogy. central and challenging part of analogy-mak-
Analogy and perception are tightly bound ing is the perceptual process: the shaping of
together. situations into representations appropriate to
It is useful to divide analogical thought a given context
into two basic components. First, there is the The mapping process, in contrast, is an
process of situational perception, which in- important object of study especially because
volves taking the data involved with a given of the immediate and natural use it provides
situation, and filtering and organizing them for the products of perception. Perception
in various ways to provide an appropriate produces a particular structure for the rep-
representation for a given context. Second, resentation of a situation, and the mapping
there is the process of mapping. This involves process emphasizes certain aspects of this
taking the representations of two situations structure. Through the study of analogy-mak-
and finding appropriate correspondences ing, we obtain a direct window onto high-
between components of one representation level perceptual processes. The study of
with components of the other to produce the which situations people view as analogous
match-up that we call an analogy. It is by no can tell us much about how people repre-
means apparent that these processes are sent those situations. Along the same lines,
cleanly separable; they seem to interact in a the computational modeling of analogy pro-
deep way. Given the fact that perception vides an ideal testing-ground for theories of
underlies analogy, one might be tempted to high-level perception. Considering all this,
divide the process of analogy-making sequen- one can see that the investigation of analog-
tially: first situation perception, then map- ical thought has a huge role to play in the
ping. But we have seen that analogy also plays understanding of high-level perception.
a large role in perception; thus mapping may
be deeply involved in the situation-percep- CURRENT MODELS OF
tion stage, and such a clean division of the ANALOGICAL THOUGHT
processes involved could be misleading. Lat-
er, we will consider just how deeply inter- In light of these considerations, it is some-
twined these two processes are. what disheartening to note that almost all
Both the situation-perception and map- current work in the computational modeling
ping processes are essential to analogy-mak- of analogy bypasses the process of percep-
ing, but of the two the former is more funda- tion altogether. The dominant approach in-
mental, for the simple reason that the map- volves starting with fixed, preordained rep-
ping process requires representations to work resentations, and launching a mapping pro-
on, and representations are the product of cess to find appropriate correspondences
high-level perception. The perceptual pro- between representations. The mapping pro-
cesses that produce these representations cess not only takes center stage; it is the only
may in turn deeply involve analogical map- actor. Perceptual processes are simply ig-
204
High-Level Perception, Representation, and Analogy
Figure 1 The representations used by SME in finding an analogy between the solar system and the
atom. (From Falkenhainer et al, 1990.)
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
the analogy becomes somewhat hollow. Since electrons are labrled as "objects", while mass
the representations are tailored (perhaps and charge, for instance, are considered to
unconsciously) to the problem at hand, it is be "attributes". However, it seems most un-
hardly surprising that the correct structural clear that this representational division is so
correspondences are not difficult to find. A clean in human thought. Many concepts,
few pieces of irrelevant information are psychologically, seem to float back and forth
sometimes thrown in as decoys, but this between being objects and attributes, for
makes the task of the mapping process only example. Consider a model of economics:
slightly more complicated. The point is that should we regard "wealth" as an object that
if appropriate representations come presup- flows from one agent, or as an attribute of
plied, the hard part of the analogy-making the agents that changes with each transac-
task has already been accomplished. tion? There does not appear to be any obvi-
Imagine what it would take to devise a ous a priori way to make the decision. A sim-
representation of the solar system or an atom ilar problem arises with the SME treatment
independent of any context provided by a of relations, which are treated as n-place
particular problem. There are so many data predicates. A 3-place predicate can be
available: one might, for instance, include mapped only to a 3-place predicate, and nev-
information about the moons revolving er to a 4-place predicate, no matter how se-
around the planets, about the opposite elec- mantically close the predicates might be. So
tric charges on the proton and the electron, it is vitally important that every relation be
about relative velocities, about proximities represented by precisely the right kind of
to other bodies, about the number of moons, predicate structure in every representation.
about the composition of the sun or the com- It eems unlikely that the human mind makes
position of the nucleus, about the fact that a rigid demarcation between 3-place and 4-
the planets lie in one plane and that each place predicates - rather, this kind of thing
planet rotates on its axis, and so on. It comes is probably very blurry.
as no surprise, in view of the analogy sought, Thus, when one is designing a represen-
that the only relations present in the repre- tation for SME, a large number of somewhat
sentations that SME uses for these situations arbitrary choices have to be made. The per-
are the following: "attracts", "revolves formance of the program is highly sensitive
around", "gravity", "opposite-sign" and "great- to each of these choices. In each of the pub-
er" (as well as the fundamental relation lished examples of analogies made by SME,
"cause"). These, for the most part, are pre- these representations were designed in just
cisely the relations that are relevant factors the right way for the analogy to be made. It
in this analogy. The criticisms of BACON is difficult to avoid the conclusion that at least
discussed earlier apply here also: the repre- to a certain extent, the representations given
sentations used by both programs seem to to SME were constructed with those specif-
have been designed with 20-20 hindsight. ic analogies in mind. This is again reminis-
A related problem arises when we con- cent of BACON.
sider the distinction that Centner makes be- In defense of SME, it must be said that
tween objects,attributes,and relations This there is much of interest about the mapping
distinction is fundamental to the operation process itself; and unlike the creators of
of SME, which works by mapping objects BACON, the creators of SME have made
exlusively to objects and relations to rela- no great claims for their program's "insight".
tions, while paying little attention to at- It seems a shame, however, that they have
tributes. In the atom/solar - system analogy paid so little attention to the question of just
such things as the nucleus, the sun, and the how the SME's representations could have
High-Level Perception, Representation, and Analogy
been formed. Much of what is interesting in connectionist network, the program has no
analogy-making involves extracting structural ability to change its representations under
commonalities from two situations , finding pressure. This constitutes a serious impedi-
some "essence" that both share. In SME, this ment to the attempts of Holyoak and
problem of high-level perception is swept Thagard to capture the flexibility of human
under the rug, by starting with preformed analogical thought.
representations of the situations. The essence
of the sit- uations has been drawn out in ad- THE NECESSITY OF INTEGRATING
vance in the formation of these representa- HIGH-LEVEL PERCEPTION WITH
tions, leaving only the relatively easy task of MORE ABSTRACT COGNITIVE
discovering the correct mapping. It is not that PROCESSING
the work done by SME is necessarily wrong:
it is simply not tackling what are, in our opin- The fact that most current work on ana-
ion, the really difficult issues in analogy-mak- logical thought has ignored the problem of
ing. representation-formation is not necessarily
Such criticisms apply equally to most oth- a damning charge: researchers in the field
er work in the modeling of analogy. It is in- might well defend themselves by saying that
teresting to note that one of the earliest com- this process is far too difficult to study at the
putational models of analogy, Evans' ANAL- moment. In the meantime, they might argue,
OGY (Evans 1968), attempted to build its it is reasonable to assume that the work of
own representations, even if it did so in a high-level perception could be done by a
fairly rigid manner. Curiously, however, al- separate "representation module", which
most all major analogy-making programs takes raw situations and converts them into
since then have ignored the problem of rep- structured representations. Just how this
resentation-building. The work of Kedar- module might work, they could say, is not
Cabelli (1988) takes a limited step in this their concern. Their research is restricted to
direction by employing a notion of "purpose" the mapping process, which takes these rep-
to direct the selection of relevant informa- resentations as input. The problem of rep-
tion, but still starts with all representations resentation, they might claim, is a complete-
pre-built. Other researchers, such as Burst- ly separate issue. (In fact, Forbus, one of the
ein (1986), Carbonell (1986), and Winston authors of SME, has also worked on mod-
(1982), all have models that differ in signif- ules that build representations in "qualitative
icant respects from the work outlined above, physics". Some preliminary work has been
but none of these addresses the question of done on using these representations as input
perception. to SME.)
The ACME program of Holyoak and This approach would be less ambitious
Thagard (1989) uses a kind of connectionist than trying to model the entire perception-
network to satisfy a set of "soft constraints" mapping cycle, but lack of ambition is cer-
in the mapping process, thus determining the tainly no reason to condemn a project a pri-
best analogical correspondences. Neverthe- ori . In cognitive science and elsewhere, sci-
less, their approach seems to have remained entists usually study what seems within their
immune to the connectionist notion of con- grasp, and leave problems that seem too dif-
text-dependent, flexible representations. The ficult for later. If this were all there was to
representations used by ACME are preor- the story, our previous remarks might be
dained, frozen structures of predicate logic; read as pointing out the limited scope of
the problem of high-level perception is by- the present approaches to analogy, but at
passed. Despite the flexibility provided by a the same time applauding their success in
207
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
making progress on a small part of the prob- tener a perception of an evil, aggressive fig-
lem. There is, however, more to the story ure. On the other hand, if one were sympa-
than this. thetic, one might describe him as being like
By ignoring the problem of perception in Robin Hood. This could produce in the lis-
this fashion, artificial-intelligence research- tener a perception of a relatively generous
ers are making a deep implicit assumption - figure, redistributing the wealth of the Ku-
namely, that the processes of perception and waitis to the rest of the Arab population.
of mapping are temporally separable. As we Not only, then, is perception an integral
have already said, we believe that this as- part of analogy-making, but analogy-making
sumption will not hold up. We see two com- is also an integral part of perception. From
pelling arguments against such a separation this, we conclude that it is impossible to split
of perception from mapping. The first argu- analogy-making into "first perception, then
ment is simpler, but the second has a broad- mapping". The mapping process will often
er scope. be needed as an important part of the pro-
The first argument stems from the obser- cess of perception. The only solution is to
vation, made earlier, that much perception give up on any clean temporal division be-
is dependent on processes of analogy. Peo- tween the two processes, and instead to rec-
ple are constantly interpreting new situations ognize that they interact deeply.
in terms of old ones. Whenever they do this, The modular approach to the modeling
they are using the analogical process to build of analogy stems, we believe, from a percep-
up richer representations of various situa- tion of analogical thought as something quite
tions. When the controversial book "The separate from the rest of cognition. One gets
Satanic Verses" was attacked by Iranian the impression from the work of most re-
Moslems and its author threatened with searchers that analogy-making is conceived
death, most Americans were quick to con- of as a special tool in reasoning or problem-
demn the actions of the Iranians. Interest- solving, a heavy weapon wheeled out occa-
ingly, some senior figures in American Chris- sionally to deal with difficult problems. Our
tian churches had a somewhat different re- view, by contrast, is that analogy-making is
action. Seeing an analogy between this book going on constantly in the background of the
and the controversial film "The Last Temp- mind, helping to shape our perceptions of
tation of Christ", which had been attacked everyday situations. In our view, analogy is
in Christian circles as blasphemous, these not separate from perception: analogy-mak-
figures were hesitant about condemning the ing itself is a perceptual process.
Iranian action. Their perception of the situ- For now, however, let us accept this view
ation was significantly altered by such a sa- of mapping as a "task" in which representa-
lient analogy. tions, the products of the perceptual pro-
Similarly, seeing Nicaragua as analogous cess, are used. Even in this view, the tem-
to Vietnam might throw a particular perspec- poral separation of per-ception from map-
tive on the situation there, while seeing the ping is, we believe, a misguided effort, as
Nicaraguan rebels as "the moral equivalent the following argument will demonstrate.
of the founding fathers" is likely to give quite This second argument, unlike the previous
a different picture of the situation. Or con- one, has a scope much broader than just
sider rival analogies that might be used to the field of analogy-making. Such an argu-
explain the role of Saddam Hussein, the Ira- ment could be brought to bear on almost
qi leader who invaded Kuwait, to someone any area within artificial intelligence, dem-
who knows little about the situation. If one onstrating the necessity for "task-oriented"
were unsympathetic, one might describe him processes to be tightly integrated with high-
as analogous to Hitler, producing in the lis- level perception.
208
High-Level Perception, Representation, and Analogy
Consider the implications of the separa- as input large representations that include
tion of perception from the mapping process, all available information would require a
by the use of a separate representation mod- radical change in their design.
ule. Such a module would have to supply a The problem is simply that a vast over-
single "correct" representation for any given supply of information would be available in
situation, independent of the context or the such a representation. To determine precisely
task for which it is being used. Our earlier which pieces of that information were rele-
discussion of the flexibility of human repre- vant would require a complex process of fil-
sentations should already suggest that this tering and organizing the available data from
notion should be treated with great suspi- the representation. This process would in fact
cion. The great adaptability of high-level be tantamount to high-level perception all over
perception suggests that no module that pro- again. This, it would seem, would defeat the
duced a single context-independent represen- purpose of separating the perceptual process-
tation could ever model the complexity of es into a specialized module.
the process. Let us consider what might be going on
To justify this claim, let us return to the in a human mind when it makes an analogy.
DNA example. To understand the analogy Presumably people have somewhere in long-
between DNA and a zipper, the representa- term memory a representation of all their
tion module would have to produce a repre- knowledge about, say, DNA. But when a
sentation of DNA that highlights its physi- person makes a particular analogy involving
cal, base-paired structure. On the other hand, DNA, only certain information about DNA
to understand the analogy between DNA and is used. This information is brought from
source code, a representation highlighting long-term memory and probably used to
DNA's information-carrying properties form a temporary active representation in
would have to be constructed. Such repre- working memory. This second representa-
sentations would clearly be quite different tion will be much less complex, and conse-
from each other. quently much easier for the mapping pro-
The only solution would be for the repre- cess to manipulate. It seems likely that this
sentation module to always provide a repre- smaller representation is what corresponds
sentation all-encompassing enough to take to the specialized representations we saw
in every possible aspect of a situation. For used by SME above. It is in a sense a projec-
DNA, for example, we might postulate a sin- tion of the larger representation from long-
gle representation incorporating information term memory - with only the relevant as-
about its physical, double-helical structure, pects being projected. It seems psychologi-
about the way in which its information is used cally implausible that when a person makes
to build up cells, about its properties of rep- an analogy, their working memory is hold-
lication and mutation, and much more. Such ing all the information from an all-encom-
a representation, were it possible to build it, passing representation of a situation. Instead,
would no doubt be very large. But its very it seems that people hold in working memo-
size would make it far too large for immedi- ry only a certain amount of relevant infor-
ate use in processing by the higher-level task- mation with the rest remaining latent in long-
oriented processes for which it was intended term storage.
- in this case, the mapping module. The map- But the process of forming the appropri-
ping processes used in most current computer ate representation in working memory is
models of analogy-making, such as SME, all undoubtedly not simple. Organizing a rep-
use very small representations that have the resentation in working memory would be
relevant information selected and ready for another specific example of the action of the
immediate use. For these programs to take high-level perceptual processes - filtering and
209
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
This does not mean, however, that one will have made genuine progress towards
must admit defeat. There is another route to understanding the same phenomena in the
the goal. The real world may be too com- unrestricted real world.
plex, but if one restrict the domain, some The model that we will examine here
understanding may be within our grasp. If, works in a domain of alphabetical letter-
instead of using the real world, one carefully strings. This domain is simple enough that
creates a simpler, artificial world in which to the problems of low-level perception are
study high-level perception, the problems avoided, but complex enough that the main
become more tractable. In the absence of issues in high-level perception arise and can
large amounts of pixel-by-pixel information, be studied. The model, the "Copycat" pro-
one is led much more quickly to the prob- gram (Hofstadter 1984; Mitchell 1990; Hof-
lems of high-level perception, which can then stadter and Mitchell 1992), is capable of
be studied in their own right. building up its own representations of situa-
Such restricted domains, or micro- tions in this domain, and does so in a flexi-
domains, can be the source of much insight. ble, context- dependent manner. Along the
Scientists in all fields throughout history have way, many of the central problems of high-
chosen or crafted idealized domains to study level perception are dealt with, using mech-
particular phenomena. When researchers anisms that have a much broader range of
attempt to take on the full complexity of the application than just this particular domain.
real world without first having some ground- Such a model may well serve as the basis for
ing in simpler domains, it often proves to be a later, more general model of high-level
a misguided enterprise. Unfortunately, mi- perception.
crodomains have fallen out of favor in artifi- This highly parallel, non-deterministic
cial intelligence. The "real world" modeling architecture builds its own representations
that has replaced them, while ambitious, has and finds appropriate analogies by means
often led to misleading claims (as in the case of the continual interaction of perceptual
INTEGRATES of BACON), or to limited models (as we saw structuring-agents with an associative con-
:EPTION WITH with models of analogy). Furthermore, while cept network. It is this interaction between
IAKING "real world" representations have impressive perceptual structures and the concept net-
labels - such as "atom" or "solar system" - work that helps the model capture part of
perception is clearly attached to them, these labels conceal the the flexibility of human thought. The Copy-
stacle lies in the way. fact that the representations are nothing but cat program is a model of both high-level
level perception to simple structures in predicate logic or a sim- perception and analogical thought, and it
st be firmly founded ilar framework. Programs like BACON and uses the integrated approach to situation
perception. But the SME are really working in stripped-down perception and mapping that we have been
ition available in the domains of certain highly idealized logical advocating.
roblem of low-level forms - their domains merely appear to have The architecture could be said to fall
;ly complex one, and the complexity of the real world, thanks to somewhere on the spectrum between the
inderstandably been the English words attached to these forms. connectionist and symbolic approaches to
perception poses so While microdomains may superficially artificial intelligence, sharing some of the
- POW, the modeling seem less impressive than "real world" do- advantages of each. On the one hand, like
e. ^rception of the mains, the fact that they are explicitly ideal- connectionist models, Copycat consists of
>al. The gap between ized worlds allows the issues under study to many local, bottom-up, parallel processes
;eption (cells on the be thrown into clear relief - something that from whose collective action higher-level
:reen, waveforms of generally speaking is not possible in a full- understanding emerges. On the other hand,
svel (conceptual pro- scale real-world problem. Once we have it shares with symbolic models the ability to
iplex structured rep- some understanding of the way cognitive deal with complex hierarchically-structured
nt too wide to bridge. processes work in a restricted domain, we representations.
21:
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
We shall use Copycat to illustrate possi- Before we begin a discussion of the de-
ble mechanisms for dealing with five impor- tails of Copycat, we should note that the pro-
tant problems in perception and analogy. gram knows nothing about the shapes of let-
These are: ters, their sounds, or their roles in the En-
• the gradual building-up of represen- glish language. It does know the order of the
tations; letters in the alphabet, both forwards and
• the role of top-down and contextu- backwards (to the program, the alphabet is
al influences; in no sense "circular"). The alphabet consists
• the integration of perception and of 26 "platonic" letter entities, each with no
mapping; explicit relation to anything except its imme-
• the exploration of many possible diate neighbors. When instances of these sim-
paths toward a representation; ple concepts, the letters, are combined into
• the radical restructuring of percep- strings of various lengths, quite complex "sit-
tions, when necessary. uations" can result. The task of the program
The description of Copycat given here will is to perceive structure in these situations,
necessarily be brief and oversimplified, but and to use this structure to make good anal-
further details are available elsewhere (Hof- ogies.
stadter 1984; Mitchell and Hofstadter 1990; The architecture used by the program,
Mitchell 1990; Hofstadter and Mitchell incidentally, is applicable much more wide-
1992). ly than to just the particular domain used
here. For instance, the architecture has also
THE COPYCAT DOMAIN been implemented to deal with the problem
of perceiving structure and making analo-
The task of the Copycat program is to gies involving the dinner implements on a
make analogies between strings of letters. For tabletop (a microdomain with a more "real
instance, it is clear to most people that abc world" feel) (French 1988). An application
and iijjkkll share common structure at some involving perception of the shapes and styles
level. The goal of the program is to capture of visual letterforms, and generation of new
this by building, for each string, a represen- letterforms sharing the given style, has also
tation that highlights this common structure, been proposed (Hofstadter et al, 1987).
and by finding correspondences between the
two representations. BUILDING UP REPRESENTATIONS
The program uses the result of this cor-
respondence-making to solve analogy prob- The philosophy behind the model under
lems of the following form: "If abc changes discussion is that high-level perception
to abd, what does iijjkkll change to?" Once emerges as a product of many independent
the program has discovered common struc- but cooperating processes running in paral-
ture in the two strings abc and iijjkkll, de- lel. The system is at first confronted with a
ciding that the letter a in the first corre- raw situation, about which it knows almost
sponds to the group ii in the second and nothing. Then a number of perceptual agents
that c corresponds to 11, it is relatively swarm over and examine the situation, each
straightforward for it to deduce that the best discovering small amounts of local structure
answer must be iijjkkmm. The difficult task adding incrementally to the system's percep-
for the program - the part requiring high- tion, until finally a global understanding of
level perception - is to build the represen- the situation emerges.
tations in the first place. We will shortly These perceptual agents, called codeletes,
examine in more detail just how these rep- are the basic elements of Copycat's percep-
resentations are built. tual processing. Each codelet is a small piece
212
High-Level Perception, Representation, and Analogy
tion. As a particular feature of the situa- groups on either side, which will activate the
tion is noted, the node representing the node representing the concept of "sameness
concept that corresponds to the feature is group". On the other hand, in the string ab-
activated in the concept network. In turn, cijkpqrrst, the presence of the groups abc
the activation of this concept has a biasing and ijk will cause the node representing "suc-
effect on the perceptual processing that cessor group" (a group consisting of alpha-
follows. Specifically, it causes the creation betically successive letters) to be active, mak-
of some number of associated codelets, ing it more likely that pqr and rst will be
which are placed in the pool of codelets perceived in the same way. Here, then, it is
waiting to run. For instance, if the node more likely that the two adjacent rDs will be
corresponding to the concept of "alphabetic perceived separately, as parts of two differ-
successor" is activated in the Slipnet, then ent "successor groups" of three letters each.
several codelets will be spawned whose task The way in which two neighboring r's are
is to look for successorship relations else- perceived (i.e., as grouped or not) is highly
where in the situation. dependent on the context that surrounds
Further, the activation of certain nodes them, and this contextual dependence is me-
means that it is more likely that associated diated by the Slipnet.
perceptual processes will succeed. If the This two-way interaction between the
"successor" node is highly active, for ex- perceptual process and the concept net-
ample, not only is it more likely that code- work is a combination of top-down and
lets that try to build successorship relations bottom-up processing. The perceptual
will be sprawned, but it is also more likely work performed by the codelets is an in-
that once they run, they C rather than some herently bottom-up process, achieved by
competing type of codelet C will suceed in competing and cooperating agents each of
building a lasting relation as part of the which acts locally. The Slipnet, however,
representation. In both of these ways, per- by modulating the action of the codelets,
ceptual processing that has already been acts as a top-down influence on this bot-
completed can have a contextual, top-down tom-up process. The Slipnet can thus be
influence on subsequent processing regarded as a dynamic controller, allow-
through activation of concepts in the Slip- ing global properties such as the activation
net. of concepts to influence the local action of
For instance, in the string kkrrtt it is likely perceptual agents. This top-down influence
that the two r's will be perceived as a "same- is vitally important, as it it ensures that
ness group" (a group all of whose members perceptual processes do not go on inde-
are the same); such a perception will be re- pendently of the system's understanding of
inforced by the presence of two similar the global context.
Mign-Levei rercepiion, jKepresemanon, ana Anaiugy
number that measures the amount and qual- system concentrates more and more on de- representatior
ity of structure present in the current repre- veloping the structure that exists. Eventual- that are more
sentation of the situation. Relevant structures ly, the program will converge on a good rep- Maier's two-s
here include bonds, groups, and correspon- resentation of a given situation. In practice, example of rac
dences, as well as some others. The term Copycat frequently comes up in different to forget abou
"quality of structure" refers to how well dif- runs with different representations for the a pair of plien
ferent parts of the structure cohere with each same situation, but these representations usu- and instead se
other. Computational temperature is used to ally seem to be cognitively plausible. Its fi- In Copyca
control the amount of randomness in the lo- nal "solutions" to various analogy problems been built up, i
cal action of codelets. If a large amount of are distributed in a fashion qualitatively sim- ily gone dowr
good structure has been built up, the tem- ilar to the distributions found with human change to anc
perature will be low and the amount of ran- subjects (Mitchell 1990; Hofstadter and obviously not;
domness allowed will be small. Under these Mitchell 1992). to keep a repn
circumstances, the system will proceed in a The process of exploring many possibili- a solution. For
fairly deterministic way, meaning that it sticks ties and gradually focusing on the most prom- a special set of
closely to a single pathway with few rival side- ising ones has been called a "parallel terraced situations.
explorations being considered. On the other scan" (Hofstadter 1984; Hofstadter and For instanc
hand, if there is little good structure, the tem- Mitchell 1992). The process is akin to the the analogy "If
perature will be high, which will lead to di- solution to the "two-armed bandit" problem xyz change to'i
verse random explorations being carried out (Holland 1975) where a gambler has access representation
by codelets. to two slot machines with fixed but distinct groups, and qui
At the start of a run, before any structure probabilities of payoff. These payoff proba- to the oth^j- ace
has been built, the temperature is maximally bilities are initially unknown to the gambler, to z, etc, ,ut r
high, so the system will behave in a very ran- who wishes to maximize payoffs over a se- the transformal
dom way. This means that many different ries of trials. The best strategy is to start by ogous to abc 1
pathways will be explored in parallel by the sampling both machines equally, but to grad- blocked, since
perceptual processes. If no promising struc- ually focus one's resources probabilistically successor of thi
tural organization emerges, then the temper- on the machine that appears to be giving the bet is non-circi
ature will remain high and many different better payoff. The Copycat program has to "snag"; the onl)
possibilities will continue to be explored. perform an analogous task. To function flex- an alternative n
Gradually, in most situations, certain struc- ibly, it has to sample many representational The progran
tures will prove more promising, and these possibilities and choose those that promise ly by raising th
are likely to form the basis of the final rep- to lead to the most coherent worldview, grad- ture shoots up ti
resentation. At any given moment, a single ually settling down to a fixed representation duces a great de
structural view is dominant, representing the of the situation. In both the two-armed ban- let level. Secoi
system's current most coherent worldview, dit and in Copycat, it takes time for certain brought in for tl
but many other tentative structures may be possibilities to emerge as the most fruitful, ing representati
present in the background, competing with and a biased stochastic sampling technique representations
it. is optimal for this purpose. up are broken d
As good structures build up, the temper- activation is po
ature grawdually falls and so the system's RADICAL RESTRUCTURING senting the soui
exploratory behavior becomes less random. Z-andn, -ip<
This mean that structures that have already Sometimes representations that have been on the specific 2
been built have a lowered chance of being built up for a given situation turn out to be very salient and
replaced by new ones and are thus favored. inappropriate, in that they do not lead to a causes a signific
The more coherent a global structure, the solution to the problem at hand. When peo- tation-building
less likely parts of it are to be broken. As ple find themselves in this situation, they need around. To mak
structure builds up and temperature falls, the to be able to completely restructure their gram is thereby
216
example of radical restructuring; people have sor group, starting from z and going back-
to forget about their initial representation of wards. Under this new representation, the a
a pair of pliers as a tool for bending things, in the first string is mapped to the z in the
and instead see it as a heavy weight. second.
In Copycat, when a representation has Now if the program attempts to complete
been built up, the temperature has necessar- its task, it discovers that the appropriate
ily gone down, which makes it difficult to transformation on xyz is to take the prede-
change to another representation. But it is cessor of the leftmost letter, and it comes up
obviously not advantageous for the program with the insightful answer wyz. (We should
to keep a representation that does not led to stress that the program, being nondetermin-
a solution. For this reason, the program has istic, does not always or even consistently
a special set of mechanisms to deal with such come up with this answer. The answer xyd is
situations. actually given more often than wyz.) Further
For instance.when the program is given details are given by Mitchell and Hofstadter
the analogy "If abc changes toabc, what does (1990).
xyz change to?", it usually builds a detailed This process of reperception can be re-
representation of abc and xyz as successor garded as a stripped-down model of a "sci-
groups, and quite reasonably maps one string entific revolution" (Kuhn 1970) in a micro-
to the other accordingly (a maps tox, c maps domain. According to this view, when a field
to z, etc). But now, when it tries to carry out of science comes up against a problem it
the transformation forxyz that it feels is anal- cannot solve, clamor and confusion result in
ogous to abc becoming abd, it finds itself the field, culminating in a "paradigm shift"
blocked, since it is impossible to take the where the problem is viewed in a completely
successor of the letter z (the Copycat alpha- different way. With the new worldview, the
bet is non-circular). The program has hit a problems may be straightforward. The radi-
"snag"; the only way to deal with it is to find cal restructuring involved in the above let-
an alternative representation of the situation. ter-string problem seems quite analogous to
The program deals with the problem first- this scientific process.
ly by raising the temperature. The tempera-
ture shoots up to its maximal value. This pro- WHAT COPYCAT DOESNT DO
duces a great deal of randomness at the code-
let level. Secondly, "breaker codelets" are Some have argued that in employing
brought in for the express purpose of destroy- hand-coded mechanisms such as codelets
ing representations. The result is that many and the Slipnet, Copycat is guilty of 20-20
representations that have been carefully built hindsight in much the same fashion as BA-
up are broken down. At the same time, much CON and SME. But there is a large differ-
activation is poured into the concept repre- ence: BACON and SME use fixed represen-
senting the source of the snag - the concept tations, whereas Copycat develops flexible
Z - and much perceptual attention is focused representations using fixcdperceptual mech-
on thespecific z inside xyz (that is, it becomes anisms. Whereas we have seen that the use
very salient and attracts many codelets). This of fixed representations is cognitively implau-
causes a significant change in the represen- sible, it is clear that human beings at any giv-
tation-building process the second time en time have a fixed repertoire of mecha-
around. To make a long story short, the pro- nisms available to the perceptual process.
gram is thereby able to come up with a com- One might justifiably ask where these mech-
217
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
anisms, and the corresponding mechanisms visual modality than Copycat's high-level
in Copycat, come from, but this would be a mechanisms, but still abstracting away from
question about learning. Copycat is not in- the messy low-level details), although the
tended as a model of learning: its per- representations developed are less sophisti-
formance, for instance, does not improve cated than Copycat's. Along similar lines,
from one run to the next. It would be a very Shrager (1990) has investigated the central
interesting further step to incorporate learn- role of perceptu al processes in scientific
ing processes into Copycat, but at present thought, and has developed a program that
the program should be taken as a model of builds up representations in the domain of
the perceptual processes in an individual understanding the operation of a laser, start-
agent at a particular time. ing from idealized two-dimensional inputs.
There are other aspects of human cogni-
tion that are not incorporated into Copycat. 5 CONCLUSION
For instance, there is nothing in Copycat that
corresponds to the messy low-level percep- It may sometimes be tempting to regard
tion that goes on in the visual and auditory perception as not truly "cognitive", something
systems. It might well be argued that just as that can be walled off from higher processes,
high-level perception exerts a strong influ- allowing researchers to study such processes
ence on and is intertwined with later cogni- without getting their hands dirtied by the com-
tive processing, so low-level perception is plexity of perceptual processes. But this is al-
equally intertwined with high-level percep- most certainly a mistake. Cognition is infused
tion. In the end, a complete model of high- with perception. This has been recognized in
level perception will have to take low-level psychology for decades, and in philosophy for
perception into account, but for now the com- longer, but artificial-intelligence research has
plexity of this task means that key features been slow to pay attention.
of the high-level perceptual processes must Two hundred years ago, Kant provoca-
be studied in iso- lation from their low-level tively suggested an intimate connection be-
base. tween concepts and perception. "Concepts
The Tabletop program (French and Hof- without percepts", he wrote, "are empty; per-
stadter 1991; French 1992) takes a few steps cepts without concepts are blind." In this
towards lower-level perception, in that it must paper we have tried to demonstrate just how
make analogies between visual structures in true this statement is, and just how depen-
a two-dimensional world, although this world dent on each other conceptual and percep-
is still highly idealized. There is also a small tual processes are in helping people make
amount of related work in AI that attempts sense of their world.
to combine perceptual and cognitive process- "Concepts without percepts are empty."
es. It is interesting to note that in this work, Research in artificial intelligence has often
microdomains are almost always used. Chap- tried to model concepts while ignoring per-
man's "Sonja" program (Chapman 1991), for ception. But as we have seen, high-level per-
instance, functions in the world of a video ceptual processes lie at the heart of human
game. Starting from simple graphical infor- cognitive abilities. Cognition cannot succeed
mation, it develops representations of the without processes that build up appropriate
situation around it and takes appropriate representations. Whether one is studying
action. As in Tabletop, the input to Sonja's analogy-making, scientific discovery, or some
perceptual processes is a little more complex other area of cognition, it is a mistake to try
than in Copycat, so that these processes can to skim off conceptual processes from the
justifiably be claimed to be a model of "in- perceptual substrate on which they rest, and
termediate vision" (more closely tied to the with which they are tightly intermeshed.
218
"Percepts without concepts are blind." Chapman, D. (1991). Vision, instruction, and
Our perception of any given situation is guid- action. Cambridge, MA: MIT Press.
ed by constant top-down influence from the Elman, J. L. (1990). Finding structure in time.
conceptual level. Without this conceptual Cognitive Science, 14: 179-212.
influence, the representations that result from Evans, T. G. (1968). A program for the so-
such perception wiD be rigid, inflexible, and lution of a class of geometric-analogy
unable to adapt to the problems provided by intelligence-test questions. In M. Min-
many different contexts. The flexibility of sky (ed.), Semantic information pro-
human perception derives from constant in- cessing (Cambridge, MA: MIT Press).
teraction with the conceptual level. We hope Falkenhainer, B., Forbus, K. D., and Gent-
that the model of concept-based perception ner, D. (1990). The structure-mapping
that we have described goes some way to- engine. Artificial Intelligence, 41:1-63.
wards drawing these levels together. Fodor, J. A. (1983). The modularity of mind
Recognizing the centrality of perceptual (Cambridge, MA: MIT Press).
processes makes artificial intelligence more French, R. M., and Hofstadter, D. R. (1991).
difficult, but it also makes it more interest- Tabletop:A stochastic, emergent mod-
ing. Integrating perceptual processes into a el of analogy-making. Proceedings of
cognitive model leads to flexible representa- the 13th annual conference of the Cog-
tions, and flexible representations lead to nitive Science Society. Hillsdale, NJ:
flexible actions. This is a fact that has only Lawrence Erlbaum / this volume.
recently begun to permeate artificial intelli- French, R. M. (1992). Tabletop: A stochas-
gence, through such models as connection- tic, emergent model of analogy-mak-
ist networks, classifier systems, and the ar- ing. Doctoral dissertation, University
chitecture presented here. Future advances of Michigan.
in the understanding of cognition and of per- Gentner, D. (1983). Structure-mapping: A
ception are likely to go hand in hand, for the theoretical framework for analogy.
two types of process are inextricably inter- Cognitive Science, 7(2).
twined. Hofstadter, D. R. (1984). The Copycat
project: An experiment in nondeter-
REFERENCES minism and creative analogies. M.I.T.
AI Laboratory memo 755.
Bruner, J. (1957). On perceptual readiness. Hofstadter, D. R., and Mitchell, M. (1992).
Psychological Review,64:123-152. An overview of the Copycat project.
Burstein, M. (1986). Concept formation by In K. J. Holyoak and J. Barnden
incremental analogical reasoning and (eds.), Connectionist Approaches to
debugging. In R. S. Michalski, J. G. Analogy, Metaphor, and Case-Based
Carbonell, and T. M. Mitchell (eds.), Reasoning (Norwood, NJ: Ablex).
Machine learning: An artificial intelli- Hofstadter, D. R., Mitchell, M., and French,
gence approach, Vol. 2 (Los Altos, R. M. (1987). Fluid concepts and cre-
CA: Morgan Kaufmann). ative analogies: A theory and its com-
Carbonell, J. G. (1986). Learning by analo- puter implementation. CRCC Tech-
gy: Formulating and generalizing nical Report 18, Indiana University.
plans from past experience. In R. S. Holland, J. H. (1975). Adaptation in natural
Michalski, J. G. Carbonell, and T. M. and artificial systems (Ann Arbor, MI:
Mitchell (eds.), Machine learning: An University of Michigan Press).
artificial intelligence approach, Vol. 2 Holland, J. H. (1986). Escaping brittleness:
(Los Altos, CA: Morgan Kaufmann). The possibilities of general-purpose
219
David J. Chalmers, Robert M. French, Douglas R. Hofstadter