Codes, Cues, Clues & Affordances
Presentation by Michael Lissack
At “Emergences of Designs”
March 25, 2006
Washington Academy of Sciences
National Science Foundation
High-Level Perception, Representation, and Analogy:
A Critique of Artificial Intelligence Methodology
David Chalmers, Robert French, Douglas Hofstadter 3
The Science of Art: A Neurological Theory of Aesthetic Experience
V.S. Ramachandran and William Hirstein 39
Cracking the code of art’s allure
Anthony Freeman 60
Art and the Brain: Editorial Introduction
Joseph A. Goguen 70
Three Laws of Qualia: What Neurology Tells Us about the Biological
Functions of Consciousness, Qualia and the Self
V. S. Ramachandran and William Hirstein 85
Some Speculative Hypotheses about the Nature and Perception of
Dance and Choreography
Ivar Hagendoorn 115
Visualization as Interpretive Practice: The Case of Detective Fiction
Andrea K. Laue 147
Code and Myth: An Introduction
Benjamin Bratton 155
High-Level Perception, Representation, and Analogy:
A Critique of Artificial Intelligence Methodology
David J. Chalmers, Robert M. French, Douglas R. Hofstadter
Center for Research on Concepts and Cognition
Bloomington, Indiana 47408
CRCC Technical Report 49 — March 1991
To appear in Journal of Experimental and Theoretical Artificial Intelligence.
High-Level Perception, Representation, and Analogy:
A Critique of Artificial Intelligence Methodology
High-level perception—the process of making sense of complex data at an abstract, conceptual
level—is fundamental to human cognition. Through high-level perception, chaotic environmen-
tal stimuli are organized into the mental representations that are used throughout cognitive pro-
cessing. Much work in traditional artificial intelligence has ignored the process of high-level
perception, by starting with hand-coded representations. In this paper, we argue that this dis-
missal of perceptual processes leads to distorted models of human cognition. We examine some
existing artificial-intelligence models—notably BACON, a model of scientific discovery, and the
Structure-Mapping Engine, a model of analogical thought—and argue that these are flawed pre-
cisely because they downplay the role of high-level perception. Further, we argue that perceptu-
al processes cannot be separated from other cognitive processes even in principle, and therefore
that traditional artificial-intelligence models cannot be defended by supposing the existence of a
“representation module” that supplies representations ready-made. Finally, we describe a model
of high-level perception and analogical thought in which perceptual processing is integrated with
analogical mapping, leading to the flexible build-up of representations appropriate to a given
1 The Problem of Perception
One of the deepest problems in cognitive science is that of understanding how people make
sense of the vast amount of raw data constantly bombarding them from their environment. The
essence of human perception lies in the ability of the mind to hew order from this chaos, whether
this means simply detecting movement in the visual field, recognizing sadness in a tone of voice,
perceiving a threat on a chessboard, or coming to understand the Iran–Contra affair in terms of
It has long been recognized that perception goes on at many levels. Immanuel Kant divided
the perceptual work of the mind into two parts: the faculty of Sensibility, whose job it is to pick
up raw sensory information, and the faculty of Understanding, which is devoted to organizing
these data into a coherent, meaningful experience of the world. Kant found the faculty of
Sensibility rather uninteresting, but he devoted much effort to the faculty of Understanding. He
went so far as to propose a detailed model of the higher-level perceptual processes involved,
dividing the faculty into twelve Categories of Understanding.
Today Kant’s model seems somewhat baroque, but his fundamental insight remains valid.
Perceptual processes form a spectrum, which for convenience we can divide into two
components. Corresponding roughly to Kant’s faculty of Sensibility, we have low-level
perception, which involves the early processing of information from the various sensory
modalities. High-level perception, on the other hand, involves taking a more global view of this
information, extracting meaning from the raw material by accessing concepts, and making sense
of situations at a conceptual level. This ranges from the recognition of objects to the grasping of
abstract relations, and on to understanding entire situations as coherent wholes.
Low-level perception is far from uninteresting, but it is high-level perception that is most
relevant to the central problems of cognition. The study of high-level perception leads us directly
to the problem of mental representation. Representations are the fruits of perception. In order for
raw data to be shaped into a coherent whole, they must go through a process of filtering and
organization, yielding a structured representation that can be used by the mind for any number of
purposes. A primary question about representations, currently the subject of much debate,
concerns their precise structure. Of equal importance is the question of how these representations
might be formed in the first place, via a process of perception, starting from raw data. The
process of representation-formation raises many important questions: How are representations
influenced by context? How can our perceptions of a situation radically reshape themselves when
necessary? Where in the process of perception are concepts accessed? Where does meaning
enter, and where and how does understanding emerge?
The main thesis of this paper is that high-level perception is deeply interwoven with other
cognitive processes, and that researchers in artificial intelligence must therefore integrate
perceptual processing into their modeling of cognition. Much work in artificial intelligence has
attempted to model conceptual processes independently of perceptual processes, but we will
argue that this approach cannot lead to a satisfactory understanding of the human mind. We will
examine some existing models of scientific discovery and analogical thought in support of this
claim, and will argue that the exclusion of perceptual processes from these models leads to
serious limitations. The intimate link between analogical thought and high-level perception will
be investigated in detail, and we will describe a computational model in which the two processes
Low-level and high-level perception
The lowest level of perception occurs with the reception of raw sensory information by
various sense organs. Light impinges on the retina, sound waves cause the eardrum to vibrate,
and so on. Other processes further along the information-processing chain may also be usefully
designated as low-level. In the case of vision, for instance, after information has passed up the
optic nerve, much basic processing occurs in the lateral geniculate nuclei and the primary visual
cortex, as well as the superior colliculus. Included here is the processing of brightness contrasts,
of light boundaries, and of edges and corners in the visual field, and perhaps also location
Low-level perception is given short shrift in this paper, as it is quite removed from the more
cognitive questions of representation and meaning. Nonetheless, it is an important subject of
study, and a complete theory of perception will necessarily include low-level perception as a
The transition from low-level to high-level perception is of course quite blurry, but we may
delineate it roughly as follows. High-level perception begins at that level of processing where
concepts begin to play an important role. Processes of high-level perception may be subdivided
again into a spectrum from the concrete to the abstract. At the most concrete end of the spectrum,
we have object recognition, exemplified by the ability to recognize an apple on a table, or to pick
out a farmer in a wheat field. Then there is the ability to grasp relations. This allows us to
determine the relationship between a blimp and the ground (“above”), or a swimmer and a
swimming pool (“in”). As one moves further up the spectrum towards more abstract relations
(“George Bush is in the Republican Party”), the issues become distant from particular sensory
modalities. The most abstract kind of perception is the processing of entire complex situations,
such as a love affair or a war.
One of the most important properties of high-level perception is that it is extremely flexible.
A given set of input data may be perceived in a number of different ways, depending on the
context and the state of the perceiver. Due to this flexibility, it is a mistake to regard perception
as a process that associates a fixed representation with a particular situation. Both contextual
factors and top-down cognitive influences make the process far less rigid than this. Some of the
sources of this flexibility in perception are as follows.
Perception may be influenced by belief. Numerous experiments by the “New Look” theorists
in psychology in the 1950’s (e.g., Bruner 1957) showed that our expectations play an important
role in determining what we perceive even at quite a low level. At a higher level, that of
complete situations, such influence is ubiquitous. Take for instance the situation in which a hus-
band walks in to find his wife sitting on the couch with a male stranger. If he has a prior belief
that his wife has been unfaithful, he is likely to perceive the situation one way; if he believes that
an insurance salesman was due to visit that day, he will probably perceive the situation quite dif-
Perception may be influenced by goals. If we are trying to hike on a trail, we are likely to
perceive a fallen log as an obstacle to be avoided. If we are trying to build a fire, we may
perceive the same log as useful fuel for the fire. Another example: Reading a given text may
yield very different perceptions, depending on whether we are reading it for content or proof-
Perception may be influenced by external context. Even in relatively low-level perception, it
is well known that the surrounding context can significantly affect our perception of visual
images. For example, an ambiguous figure halfway between an “A” and an “H” is perceived one
way in the context of “C—T”, and another in the context of “T—E”. At a higher level, if we
encounter somebody dressed in tuxedo and bow-tie, our perception of them may differ depending
on whether we encounter them at a formal ball or at the beach.
Perceptions of a situation can be radically reshaped where necessary. In Maier’s well-
known two-string experiment (Maier 1931), subjects are provided with a chair and a pair of
pliers, and are told to tie together two strings hanging from the ceiling. The two strings are too
far apart to be grasped simultaneously. Subjects have great difficulty initially, but after a number
of minutes some of them hit upon the solution of tying the pliers to one of the strings, and
swinging the string like a pendulum. Initially, the subjects perceive the pliers first and foremost
as a special tool; if the weight of the pliers is perceived at all, it is very much in the background.
To solve this problem, subjects have to radically alter the emphasis of their perception of the pair
of pliers. Its function as a tool is set aside, and its weightiness is brought into the foreground as
the key feature in this situation.
The distinguishing mark of high-level perception is that it is semantic: it involves drawing
meaning out of situations. The more semantic the processing involved, the greater the role played
by concepts in this processing, and thus the greater the scope for top-down influences. The most
abstract of all types of perception, the understanding of complete situations, is also the most
Recently both Pylyshyn (1980) and Fodor (1983) have argued against the existence of top-
down influences in perception, claiming that perceptual processes are “cognitively impenetrable”
or “informationally encapsulated”. These arguments are highly controversial, but in any case
they apply mostly to relatively low-level sensory perception. Few would dispute that at the
higher, conceptual level of perception, top-down and contextual influences play a large role.
2 Artificial Intelligence and the Problem of Representation
The end product of the process of perception, when a set of raw data has been organized into
a coherent and structured whole, is a representation. Representations have been the object of
much study and debate within the field of artificial intelligence, and much is made of the
“representation problem”. This problem has traditionally been phrased as “What is the correct
structure for mental representations?”, and many possibilities have been suggested, ranging from
predicate calculus through frames and scripts to semantic networks and more. We may divide
representations into two kinds: long-term knowledge representations that are stored passively
somewhere in the system, and short-term representations that are active at a given moment in a
particular mental or computational process. (This distinction corresponds to the distinction
between long-term memory and working memory.) In this discussion, we will mostly be
concerned with short-term, active representations, as it is these that are the direct product of
The question of the structure of representations is certainly an important one, but there is
another, related problem that has not received nearly as much attention. This is that of
understanding how such a representation could be arrived at, starting from environmental data.
Even if it were possible to discover an optimal type of representational structure, this would leave
unresolved two important problems, namely:
The problem of relevance: How is it decided which subsets of the vast amounts of
data from the environment get used in various parts of the representational structure?
Naturally, much of the information content at the lowest level will be quite irrelevant
at the highest representational level. To determine which parts of the data are relevant
to a given representation, a complex filtering process is required.
The problem of organization: How are these data put into the correct form for the
representation? Even if we have determined precisely which data are relevant, and
we have determined the desired framework for the representation—a frame-based
representation, for instance—we still face the problem of organizing the data into the
representational form in a useful way. The data do not come prepackaged as slots and
fillers, and organizing them into a coherent structure is likely to be a highly non-trivi-
These questions, taken together, amount in essence to the problem of high-level perception,
translated into the framework of artificial intelligence.
The traditional approach in artificial intelligence has been to start by selecting not only a
preferred type of high-level representational structure, but also the data assumed to be relevant to
the problem at hand. These data are organized by a human programmer who appropriately fits
them into the chosen representational structure. Usually, researchers use their prior knowledge
of the nature of the problem to hand-code a representation of the data into a near-optimal form.
Only after all this hand-coding is completed is the representation allowed to be manipulated by
the machine. The problem of representation-formation, and thus the problem of high-level
perception, is ignored. (These comments do not, of course, apply to work in machine vision,
speech processing, and other perceptual endeavors. However, work in these fields usually stops
short of modeling processes at the conceptual level and is thus not directly relevant to our
critique of high-level cognitive modeling.)
The formation of appropriate representations lies at the heart of human high-level cognitive
abilities. It might even be said that the problem of high-level perception forms the central task
facing the artificial-intelligence community: the task of understanding how to draw meaning out
of the world. It might not be stretching the point to say that there is a “meaning barrier”, which
has rarely been crossed by work in AI. On one side of the barrier, some models in low-level
perception have been capable of building primitive representations of the environment, but these
are not yet sufficiently complex to be called “meaningful”. On the other side of the barrier, much
research in high-level cognitive modeling has started with representations at the conceptual level,
such as propositions in predicate logic or nodes in a semantic network, where any meaning that is
present is already built in. There has been very little work that bridges the gap between the two.
Objectivism and traditional AI
Once AI takes the problem of representation-formation seriously, the next stage will be to
deal with the evident flexibility of human high-level perceptual processes. As we have seen,
objects and situations can be comprehended in many different ways, depending on context and
top-down influences. We must find a way of ensuring that AI representations have a
corresponding degree of flexibility. William James, in the late nineteenth century, recognized
this aspect of cognitive representations:
“There is no property ABSOLUTELY essential to one thing. The same property which figures as
the essence of a thing on one occasion becomes a very inessential feature upon another. Now that
I am writing, it is essential that I conceive my paper as a surface for inscription. . . . But if I wished
to light a fire, and no other materials were by, the essential way of conceiving the paper would be
as a combustible material. . . . The essence of a thing is that one of its properties which is so
important for my interests that in comparison with it I may neglect the rest. . . . The properties
which are important vary from man to man and from hour to hour. . . . many objects of daily
use—as paper, ink, butter, overcoat—have properties of such constant unwavering importance,
and have such stereotyped names, that we end by believing that to conceive them in those ways is
to conceive them in the only true way. Those are no truer ways of conceiving them than any
others; there are only more frequently serviceable ways to us.” (James 1890, pp. 222–224)
James is saying, effectively, that we have different representations of an object or situation at dif-
ferent times. The representational process adapts to fit the pressures of a given context.
Despite the work of philosopher-psychologists such as James, the early days of artificial
intelligence were characterized by an objectivist view of perception, and of the representation of
objects, situations, and categories. As the linguist George Lakoff has characterized it, “On the
objectivist view, reality comes complete with a unique correct, complete structure in terms of
entities, properties and relations. This structure exists, independent of any human understanding.”
(Lakoff 1987, p. 159) While this objectivist position has been unfashionable for decades in
philosophical circles (especially after Wittgenstein’s work demonstrating the inappropriateness of
a rigid correspondence between language and reality), most early work in AI implicitly accepted
this set of assumptions.
The Physical Symbol System Hypothesis (Newell & Simon 1976), upon which most of the
traditional AI enterprise has been built, posits that thinking occurs through the manipulation of
symbolic representations, which are composed of atomic symbolic primitives. Such symbolic
representations are by their nature somewhat rigid, black-and-white entities, and it is difficult for
their representational content to shift subtly in response to changes in context. The result, in prac-
tice—irrespective of whether this was intended by the original proponents of this framework—is
a structuring of reality that tends to be as fixed and absolute as that of the objectivist position
By the mid-seventies, a small number of AI researchers began to argue that in order to
progress, the field would have to part ways with its commitment to such a rigid representational
framework. One of the strongest early proponents of this view was David Marr, who noted that
“the perception of an event or object must include the simultaneous computation of several
different descriptions of it, that capture diverse aspects of the use, purpose or circumstances of the
event or object.” (Marr 1977, p. 44)
Recently, significant steps have been taken toward representational flexibility with the advent
of sophisticated connectionist models whose distributed representations are highly context-
dependent (Rumelhart & McClelland 1986). In these models, there are no representational primi-
tives in internal processing. Instead, each representation is a vector in a multi-dimensional space,
whose position is not anchored but can adjust flexibly to changes in environmental stimuli.
Consequently, members of a category are not all represented by identical symbolic structures;
rather, individual objects will be represented in subtly different ways depending upon the context
in which they are presented. In networks with recurrent connections (Elman 1990),
representations are even sensitive to the current internal state of the model. Other recent work
taking a flexible approach to representation includes the classifier-system models of Holland
(1986) and his colleagues, where genetically-inspired methods are used to create a set of
“classifiers” that can respond to diverse aspects of various situations.
In these models, a flexible perceptual process has been integrated with an equally flexible de-
pendence of action upon representational content, yielding models that respond to diverse
situations with a robustness that is difficult to match with traditional methods. Nonetheless, the
models are still somewhat primitive, and the representations they develop are not nearly as
complex as the hand-coded, hierarchically-structured representations found in traditional models;
still, it seems to be a step in the right direction. It remains to be seen whether work in more
traditional AI paradigms will respond to this challenge by moving toward more flexible and ro-
bust representational forms.
On the possibility of a representation module
It might be granted that given the difficulty of the problem of high-level perception, AI
researchers could be forgiven for starting with their representations in a made-to-order form.
They might plausibly claim that the difficult problem of representation-formation is better left
until later. But it must be realized that behind this approach lies a tacit assumption: that it is
possible to model high-level cognitive processes independently of perceptual processes. Under
this assumption, the representations that are currently, for the most part, tailored by human hands,
would eventually be built up by a separate lower-level facility—a “representation module” whose
job it would be to funnel data into representations. Such a module would act as a “front end” to
the models of the cognitive processes currently being studied, supplying them with the
We are deeply skeptical, however, about the feasibility of such a separation of perception
from the rest of cognition. A representation module that, given any situation, produced the single
“correct” representation for it, would have great difficulty emulating the flexibility that
characterizes human perception. For such flexibility to arise, the representational processes
would have to be sensitive to the needs of all the various cognitive processes in which they might
be used. It seems most unlikely that a single representation would suffice for all purposes. As we
have seen, for the accurate modeling of cognition it is necessary that the representation of a given
situation can vary with various contextual and top-down influences. This, however, is directly
contrary to the “representation module” philosophy, wherein representations are produced quite
separately from later cognitive processes, and then supplied to a “task-processing” module.
To separate representation-building from higher-level cognitive tasks is, we believe, impossi-
ble. In order to provide the kind of flexibility that is apparent in cognition, any fully cognitive
model will probably require a continual interaction between the process of representation-
building and the manipulation of those representations. If this proves to be the case, then the
current approach of using hand-coded representations not only is postponing an important issue
but will, in the long run, lead up a dead-end street.
We will consider this issue in greater depth later, when we discuss current research in the
modeling of analogical thought. For now, we will discuss in some detail one well-known AI
program for which great claims have been made. We argue that these claims represent a lack of
appreciation of the importance of high-level perception.
BACON: A case study
A particularly clear case of a program in which the problem of representation is bypassed is
BACON, a well-known program that has been advertised as an accurate model of scientific
discovery (Langley et al 1987). The authors of BACON claim that their system is “capable of
representing information at multiple levels of description, which enables it to discover complex
laws involving many terms”. BACON was able to “discover”, among other things, Boyle’s law
of ideal gases, Kepler’s third law of planetary motion, Galileo’s law of uniform acceleration, and
Such claims clearly demand close scrutiny. We will look in particular at the program’s
“discovery” of Kepler’s third law of planetary motion. Upon examination, it seems that the
success of the program relies almost entirely on its being given data that have already been
represented in near-optimal form, using after-the-fact knowledge available to the programmers.
When BACON performed its derivation of Kepler’s third law, the program was given only
data about the planets’ average distances from the sun and their periods. These are precisely the
data required to derive the law. The program is certainly not “starting with essentially the same
initial conditions as the human discoverers”, as one of the authors of BACON has claimed (Simon
1989, p. 375). The authors’ claim that BACON used “original data” certainly does not mean that
it used all of the data available to Kepler at the time of his discovery, the vast majority of which
were irrelevant, misleading, distracting, or even wrong.
This pre-selection of data may at first seem quite reasonable: after all, what could be more
important to an astronomer-mathematician than planetary distances and periods? But here our
after-the-fact knowledge is misleading us. Consider for a moment the times in which Kepler
lived. It was the turn of the seventeenth century, and Copernicus’ De Revolutionibus Orbium
Cœlestium was still new and far from universally accepted. Further, at that time there was no
notion of the forces that produced planetary motion; the sun, in particular, was known to produce
light but was not thought to influence the motion of the planets. In that prescientific world, even
the notion of using mathematical equations to express regularities in nature was rare. And Kepler
believed—in fact, his early fame rested on the discovery of this surprising coincidence—that the
planets’ distances from the sun were dictated by the fact that the five regular polyhedra could be
fit between the five “spheres” of planetary motion around the sun, a fact that constituted seductive
but ultimately misleading data.
Within this context, it is hardly surprising that it took Kepler thirteen years to realize that
conic sections and not Platonic solids, that algebra and not geometry, that ellipses and not
Aristotelian “perfect” circles, that the planets’ distances from the sun and not the polyhedra in
which they fit, were the relevant factors in unlocking the regularities of planetary motion. In
making his discoveries, Kepler had to reject a host of conceptual frameworks that might, for all he
knew, have applied to planetary motion, such as religious symbolism, superstition, Christian
cosmology, and teleology. In order to discover his laws, he had to make all of these creative
leaps. BACON, of course, had to do nothing of the sort. The program was given precisely the set
of variables it needed from the outset (even if the values of some of these variables were
sometimes less than ideal), and was moreover supplied with precisely the right biases to induce
the algebraic form of the laws, it being taken completely for granted that mathematical laws of a
type now recognized by physicists as standard were the desired outcome.
It is difficult to believe that Kepler would have taken thirteen years to make his discovery if
his working data had consisted entirely of a list where each entry said “Planet X: Mean Distance
from Sun Y, Period Z”. If he had further been told “Find a polynomial equation relating these
entities”, then it might have taken him a few hours. Addressing the question of why Kepler took
thirteen years to do what BACON managed within minutes, Langley et al (1987) point to
“sleeping time, and time for ordinary daily chores”, and other factors such as the time taken in
setting up experiments, and the slow hardware of the human nervous system (!). In an interesting
juxtaposition to this, researchers in a recent study (Qin & Simon 1990) found that starting with
the data that BACON was given, university students could make essentially the same
“discoveries” within an hour-long experiment. Somewhat strangely, the authors (including one of
the authors of BACON) take this finding to support the plausibility of BACON as an accurate
model of scientific discovery. It seems more reasonable to regard it as a demonstration of the vast
difference in difficulty between the task faced by BACON and that faced by Kepler, and thus as a
reductio ad absurdum of the BACON methodology.
So many varieties of data were available to Kepler, and the available data had so many
different ways of being interpreted, that it is difficult not to conclude that in presenting their
program with data in such a neat form, the authors of BACON are inadvertently guilty of 20–20
hindsight. BACON, in short, works only in a world of hand-picked, prestructured data, a world
completely devoid of the problems faced by Kepler or Galileo or Ohm when they made their
original discoveries. Similar comments could be made about STAHL, GLAUBER, and other
models of scientific discovery by the authors of BACON. In all of these models, the crucial role
played by high-level perception in scientific discovery, through the filtering and organization of
environmental stimuli, is ignored.
It is interesting to note that the notion of a “paradigm shift”, which is central to much
scientific discovery (Kuhn 1970), is often regarded as the process of viewing the world in a
radically different way. That is, scientists’ frameworks for representing available world
knowledge are broken down, and their high-level perceptual abilities are used to organize the
available data quite differently, building a novel representation of the data. Such a new
representation can be used to draw different and important conclusions in a way that was difficult
or impossible with the old representation. In this model of scientific discovery, unlike the model
presented in BACON, the process of high-level perception is central.
The case of BACON is by no means isolated—it is typical of much work in AI, which often
fails to appreciate the importance of the representation-building stage. We will see this in more
depth in the next section, in which we take a look at the modeling of analogy.
3 Models of Analogical Thought
Analogical thought is dependent on high-level perception in a very direct way. When people
make analogies, they are perceiving some aspects of the structures of two situations—the essenc-
es of those situations, in some sense—as identical. These structures, of course, are a product of
the process of high-level perception.
The quality of an analogy between two situations depends almost entirely on one’s perception
of the situations. If Ronald Reagan were to evaluate the validity of an analogy between the U.S.
role in Nicaragua and the Soviet Union’s role in Afghanistan, he would undoubtedly see it as a
poor one. Others might consider the analogy excellent. The difference would come from
different perceptions, and thus representations, of the situations themselves. Reagan’s internal
representation of the Nicaraguan situation is certainly quite different from Daniel Ortega’s.
Analogical thought further provides one of the clearest illustrations of the flexible nature of
our perceptual abilities. Making an analogy requires highlighting various different aspects of a
situation, and the aspects that are highlighted are often not the most obvious features. The
perception of a situation can change radically, depending on the analogy we are making.
Let us consider two analogies involving DNA. The first is an analogy between DNA and a
zipper. When we are presented with this analogy, the image of DNA that comes to mind is that of
two strands of paired nucleotides (which can come apart like a zipper for the purposes of
replication). The second analogy involves comparing DNA to the source code (i.e., non-execut-
able high-level code) of a computer program. What comes to mind now is the fact that
information in the DNA gets “compiled” (via processes of transcription and translation) into
enzymes, which correspond to machine code (i.e., executable code). In the latter analogy, the
perception of DNA is radically different—it is represented essentially as an information-bearing
entity, whose physical aspects, so important to the first analogy, are of virtually no consequence.
In cases such as these, it seems that no single, rigid representation can capture what is going
on in our heads. It is true that we probably have a single rich representation of DNA sitting pas-
sively in long-term memory. However, in the contexts of different analogical mappings, very dif-
ferent facets of this large representational structure are selected out as being relevant, by the
pressures of the particular context. Irrespective of the passive content of the long-term represen-
tation of DNA, the active content that is processed at a given time is determined by a flexible
Furthermore, not only is analogy-making dependent on high-level perception, but the reverse
holds true as well: perception is often dependent on analogy-making itself. The high-level
perception of one situation in terms of another is ubiquitous in human thought. If we perceive
Nicaragua as “another Vietnam”, for example, the making of the analogy is fleshing out our
representation of Nicaragua. Analogical thought provides a powerful mechanism for the
enrichment of a representation of a given situation. This is well understood by good educators
and writers, who know that there is nothing like an analogy to provide a better mental picture of a
given situation. Analogies affect our perception all the time: in a love affair, for instance, it is
difficult to stop parallels with past romances from modulating one’s perception of the current
situation. In the large or the small, such analogical perception—the grasping of one situation in
terms of another—is so common that we tend to forget that what is going on is, in fact, analogy.
Analogy and perception are tightly bound together.
It is useful to divide analogical thought into two basic components. First, there is the process
of situation-perception, which involves taking the data involved with a given situation, and
filtering and organizing them in various ways to provide an appropriate representation for a given
context. Second, there is the process of mapping. This involves taking the representations of two
situations and finding appropriate correspondences between components of one representation
with components of the other to produce the match-up that we call an analogy. It is by no means
apparent that these processes are cleanly separable; they seem to interact in a deep way. Given
the fact that perception underlies analogy, one might be tempted to divide the process of analogy-
making sequentially: first situation perception, then mapping. But we have seen that analogy also
plays a large role in perception; thus mapping may be deeply involved in the situation-perception
stage, and such a clean division of the processes involved could be misleading. Later, we will
consider just how deeply intertwined these two processes are.
Both the situation-perception and mapping processes are essential to analogy-making, but of
the two the former is more fundamental, for the simple reason that the mapping process requires
representations to work on, and representations are the product of high-level perception. The per-
ceptual processes that produce these representations may in turn deeply involve analogical map-
ping; but each mapping process requires a perceptual process to precede it, whereas it is not the
case that each perceptual process necessarily depends upon mapping. Therefore the perceptual
process is conceptually prior, although perception and mapping processes are often temporally in-
terwoven. If the appropriate representations are already formed, the mapping process can often
be quite straightforward. In our view, the most central and challenging part of analogy-making is
the perceptual process: the shaping of situations into representations appropriate to a given
The mapping process, in contrast, is an important object of study especially because of the
immediate and natural use it provides for the products of perception. Perception produces a
particular structure for the representation of a situation, and the mapping process emphasizes
certain aspects of this structure. Through the study of analogy-making, we obtain a direct
window onto high-level perceptual processes. The study of which situations people view as
analogous can tell us much about how people represent those situations. Along the same lines,
the computational modeling of analogy provides an ideal testing-ground for theories of high-level
perception. Considering all this, one can see that the investigation of analogical thought has a
huge role to play in the understanding of high-level perception.
Current models of analogical thought
In light of these considerations, it is somewhat disheartening to note that almost all current
work in the computational modeling of analogy bypasses the process of perception altogether.
The dominant approach involves starting with fixed, preordained representations, and launching a
mapping process to find appropriate correspondences between representations. The mapping
process not only takes center stage; it is the only actor. Perceptual processes are simply ignored;
the problem of representation-building is not even an issue. The tacit assumption of such research
is that correct representations have (somehow) already been built.
Perhaps the best-known computational model of analogy-making is the Structure-Mapping
Engine (SME) (Falkenhainer, Forbus, and Gentner 1990), based upon the structure-mapping
theory of Dedre Gentner (1983). We will examine this model within the context of our earlier
remarks. Other models of analogy-making, such as those of Burstein (1986), Carbonell (1986),
Holyoak & Thagard (1989), Kedar-Cabelli (1988), and Winston (1982), while differing in many
respects from the above work, all share the property that the problem of representation-building is
Let us consider one of the standard examples from this research, in which the SME program is
said to discover an analogy between an atom and the solar system. Here, the program is given
representations of the two situations, as shown in Figure 1. Starting with these representations,
SME examines many possible correspondences between elements of the first representation and
elements of the second. These correspondences are evaluated according to how well they
preserve the high-level structure apparent in the representations. The correspondence with the
highest score is selected as the best analogical mapping between the two situations.
A brief examination of Figure 1 shows that the discovery of the similar structure in these
representations is not a difficult task. The representations have been set up in such a way that the
common structure is immediately apparent. Even for a computer program, the extraction of such
common structure is relatively straightforward.
We are in broad sympathy with Gentner’s notion that the mappings in an analogy should
preserve high-level structure (although there is room to debate over the details of the mapping
process). But when the program’s discovery of the correspondences between the two situations is
a direct result of its being explicitly given the appropriate structures to work with, its victory in
finding the analogy becomes somewhat hollow. Since the representations are tailored (perhaps
unconsciously) to the problem at hand, it is hardly surprising that the correct structural
correspondences are not difficult to find. A few pieces of irrelevant information are sometimes
thrown in as decoys, but this makes the task of the mapping process only slightly more
complicated. The point is that if appropriate representations come presupplied, the hard part of
Figure 1. The representations used by SME in finding an analogy between the solar
system and the atom. (From Falkenhainer et al, 1990.)
the analogy-making task has already been accomplished.
Imagine what it would take to devise a representation of the solar system or an atom
independent of any context provided by a particular problem. There are so many data available:
one might, for instance, include information about the moons revolving around the planets, about
the opposite electric charges on the proton and the electron, about relative velocities, about
proximities to other bodies, about the number of moons, about the composition of the sun or the
composition of the nucleus, about the fact that the planets lie in one plane and that each planet
rotates on its axis, and so on. It comes as no surprise, in view of the analogy sought, that the only
relations present in the representations that SME uses for these situations are the following:
“attracts”, “revolves around”, “gravity”, “opposite-sign” and “greater” (as well as the
fundamental relation “cause”). These, for the most part, are precisely the relations that are
relevant factors in this analogy. The criticisms of BACON discussed earlier apply here also: the
representations used by both programs seem to have been designed with 20–20 hindsight.
A related problem arises when we consider the distinction that Gentner makes between
objects, attributes, and relations. This distinction is fundamental to the operation of SME, which
works by mapping objects exclusively to objects and relations to relations, while paying little
attention to attributes. In the atom/solar-system analogy such things as the nucleus, the sun, and
the electrons are labeled as “objects”, while mass and charge, for instance, are considered to be
“attributes”. However, it seems most unclear that this representational division is so clean in
human thought. Many concepts, psychologically, seem to float back and forth between being
objects and attributes, for example. Consider a model of economics: should we regard “wealth”
as an object that flows from one agent, or as an attribute of the agents that changes with each
transaction? There does not appear to be any obvious a priori way to make the decision. A
similar problem arises with the SME treatment of relations, which are treated as n-place
predicates. A 3-place predicate can be mapped only to a 3-place predicate, and never to a 4-place
predicate, no matter how semantically close the predicates might be. So it is vitally important that
every relation be represented by precisely the right kind of predicate structure in every
representation. It seems unlikely that the human mind makes a rigid demarcation between 3-
place and 4-place predicates—rather, this kind of thing is probably very blurry.
Thus, when one is designing a representation for SME, a large number of somewhat arbitrary
choices have to be made. The performance of the program is highly sensitive to each of these
choices. In each of the published examples of analogies made by SME, these representations
were designed in just the right way for the analogy to be made. It is difficult to avoid the
conclusion that at least to a certain extent, the representations given to SME were constructed
with those specific analogies in mind. This is again reminiscent of BACON.
In defense of SME, it must be said that there is much of interest about the mapping process it-
self; and unlike the creators of BACON, the creators of SME have made no great claims for their
program’s “insight”. It seems a shame, however, that they have paid so little attention to the
question of just how the SME’s representations could have been formed. Much of what is
interesting in analogy-making involves extracting structural commonalities from two situations,
finding some “essence” that both share. In SME, this problem of high-level perception is swept
under the rug, by starting with preformed representations of the situations. The essence of the sit-
uations has been drawn out in advance in the formation of these representations, leaving only the
relatively easy task of discovering the correct mapping. It is not that the work done by SME is
necessarily wrong: it is simply not tackling what are, in our opinion, the really difficult issues in
Such criticisms apply equally to most other work in the modeling of analogy. It is interesting
to note that one of the earliest computational models of analogy, Evans’ ANALOGY (Evans
1968), attempted to build its own representations, even if it did so in a fairly rigid manner.
Curiously, however, almost all major analogy-making programs since then have ignored the
problem of representation-building. The work of Kedar-Cabelli (1988) takes a limited step in this
direction by employing a notion of “purpose” to direct the selection of relevant information, but
still starts with all representations pre-built. Other researchers, such as Burstein (1986), Carbonell
(1986), and Winston (1982), all have models that differ in significant respects from the work
outlined above, but none of these addresses the question of perception.
The ACME program of Holyoak and Thagard (1989) uses a kind of connectionist network to
satisfy a set of “soft constraints” in the mapping process, thus determining the best analogical
correspondences. Nevertheless, their approach seems to have remained immune to the
connectionist notion of context-dependent, flexible representations. The representations used by
ACME are preordained, frozen structures of predicate logic; the problem of high-level perception
is bypassed. Despite the flexibility provided by a connectionist network, the program has no
ability to change its representations under pressure. This constitutes a serious impediment to the
attempts of Holyoak and Thagard to capture the flexibility of human analogical thought.
The necessity of integrating high-level perception with more abstract cognitive processing
The fact that most current work on analogical thought has ignored the problem of
representation-formation is not necessarily a damning charge: researchers in the field might well
defend themselves by saying that this process is far too difficult to study at the moment. In the
meantime, they might argue, it is reasonable to assume that the work of high-level perception
could be done by a separate “representation module”, which takes raw situations and converts
them into structured representations. Just how this module might work, they could say, is not
their concern. Their research is restricted to the mapping process, which takes these
representations as input. The problem of representation, they might claim, is a completely
separate issue. (In fact, Forbus, one of the authors of SME, has also worked on modules that
build representations in “qualitative physics”. Some preliminary work has been done on using
these representations as input to SME.)
This approach would be less ambitious than trying to model the entire perception-mapping
cycle, but lack of ambition is certainly no reason to condemn a project a priori. In cognitive sci-
ence and elsewhere, scientists usually study what seems within their grasp, and leave problems
that seem too difficult for later. If this were all there was to the story, our previous remarks might
be read as pointing out the limited scope of the present approaches to analogy, but at the same
time applauding their success in making progress on a small part of the problem. There is,
however, more to the story than this.
By ignoring the problem of perception in this fashion, artificial-intelligence researchers are
making a deep implicit assumption—namely, that the processes of perception and of mapping are
temporally separable. As we have already said, we believe that this assumption will not hold up.
We see two compelling arguments against such a separation of perception from mapping. The
first argument is simpler, but the second has a broader scope.
The first argument stems from the observation, made earlier, that much perception is
dependent on processes of analogy. People are constantly interpreting new situations in terms of
old ones. Whenever they do this, they are using the analogical process to build up richer
representations of various situations. When the controversial book The Satanic Verses was
attacked by Iranian Moslems and its author threatened with death, most Americans were quick to
condemn the actions of the Iranians. Interestingly, some senior figures in American Christian
churches had a somewhat different reaction. Seeing an analogy between this book and the
controversial film The Last Temptation of Christ, which had been attacked in Christian circles as
blasphemous, these figures were hesitant about condemning the Iranian action. Their perception
of the situation was significantly altered by such a salient analogy.
Similarly, seeing Nicaragua as analogous to Vietnam might throw a particular perspective on
the situation there, while seeing the Nicaraguan rebels as “the moral equivalent of the founding
fathers” is likely to give quite a different picture of the situation. Or consider rival analogies that
might be used to explain the role of Saddam Hussein, the Iraqi leader who invaded Kuwait, to
someone who knows little about the situation. If one were unsympathetic, one might describe
him as analogous to Hitler, producing in the listener a perception of an evil, aggressive figure. On
the other hand, if one were sympathetic, one might describe him as being like Robin Hood. This
could produce in the listener a perception of a relatively generous figure, redistributing the wealth
of the Kuwaitis to the rest of the Arab population.
Not only, then, is perception an integral part of analogy-making, but analogy-making is also
an integral part of perception. From this, we conclude that it is impossible to split analogy-
making into “first perception, then mapping”. The mapping process will often be needed as an
important part of the process of perception. The only solution is to give up on any clean temporal
division between the two processes, and instead to recognize that they interact deeply.
The modular approach to the modeling of analogy stems, we believe, from a perception of an-
alogical thought as something quite separate from the rest of cognition. One gets the impression
from the work of most researchers that analogy-making is conceived of as a special tool in rea-
soning or problem-solving, a heavy weapon wheeled out occasionally to deal with difficult
problems. Our view, by contrast, is that analogy-making is going on constantly in the background
of the mind, helping to shape our perceptions of everyday situations. In our view, analogy is not
separate from perception: analogy-making itself is a perceptual process.
For now, however, let us accept this view of mapping as a “task” in which representations, the
products of the perceptual process, are used. Even in this view, the temporal separation of per-
ception from mapping is, we believe, a misguided effort, as the following argument will
demonstrate. This second argument, unlike the previous one, has a scope much broader than just
the field of analogy-making. Such an argument could be brought to bear on almost any area
within artificial intelligence, demonstrating the necessity for “task-oriented” processes to be
tightly integrated with high-level perception.
Consider the implications of the separation of perception from the mapping process, by the
use of a separate representation module. Such a module would have to supply a single “correct”
representation for any given situation, independent of the context or the task for which it is being
used. Our earlier discussion of the flexibility of human representations should already suggest
that this notion should be treated with great suspicion. The great adaptability of high-level
perception suggests that no module that produced a single context-independent representation
could ever model the complexity of the process.
To justify this claim, let us return to the DNA example. To understand the analogy between
DNA and a zipper, the representation module would have to produce a representation of DNA
that highlights its physical, base-paired structure. On the other hand, to understand the analogy
between DNA and source code, a representation highlighting DNA’s information-carrying
properties would have to be constructed. Such representations would clearly be quite different
from each other.
The only solution would be for the representation module to always provide a representation
all-encompassing enough to take in every possible aspect of a situation. For DNA, for example,
we might postulate a single representation incorporating information about its physical, double-
helical structure, about the way in which its information is used to build up cells, about its
properties of replication and mutation, and much more. Such a representation, were it possible to
build it, would no doubt be very large. But its very size would make it far too large for immediate
use in processing by the higher-level task-oriented processes for which it was intended—in this
case, the mapping module. The mapping processes used in most current computer models of
analogy-making, such as SME, all use very small representations that have the relevant
information selected and ready for immediate use. For these programs to take as input large
representations that include all available information would require a radical change in their
The problem is simply that a vast oversupply of information would be available in such a
representation. To determine precisely which pieces of that information were relevant would
require a complex process of filtering and organizing the available data from the representation.
This process would in fact be tantamount to high-level perception all over again. This, it would
seem, would defeat the purpose of separating the perceptual processes into a specialized module.
Let us consider what might be going on in a human mind when it makes an analogy.
Presumably people have somewhere in long-term memory a representation of all their knowledge
about, say, DNA. But when a person makes a particular analogy involving DNA, only certain
information about DNA is used. This information is brought from long-term memory and
probably used to form a temporary active representation in working memory. This second
representation will be much less complex, and consequently much easier for the mapping process
to manipulate. It seems likely that this smaller representation is what corresponds to the
specialized representations we saw used by SME above. It is in a sense a projection of the larger
representation from long-term memory—with only the relevant aspects being projected. It seems
psychologically implausible that when a person makes an analogy, their working memory is
holding all the information from an all-encompassing representation of a situation. Instead, it
seems that people hold in working memory only a certain amount of relevant information with the
rest remaining latent in long-term storage.
But the process of forming the appropriate representation in working memory is undoubtedly
not simple. Organizing a representation in working memory would be another specific example
of the action of the high-level perceptual processes—filtering and organization—responsible for
the formation of representations in general. And most importantly, this process would necessarily
interact with the details of the task at hand. For an all-encompassing representation (in long-term
memory) to be transformed into a usable representation in working memory, the nature of the task
at hand—in the case of analogy, a particular attempted mapping—must play a pivotal causal role.
The lesson to be learned from all this is that separating perception from the “higher” tasks for
which it is to be used is almost certainly a misguided approach. The fact that representations have
to be adapted to particular contexts and particular tasks means that an interplay between the task
and the perceptual process is unavoidable, and therefore that any “modular” approach to analogy-
making will ultimately fail. It is therefore essential to investigate how the perceptual and
mapping processes can be integrated.
One might thus envisage a system in which representations can gradually be built up as the
various pressures evoked by a given context manifest themselves. We will describe such a
system in the next section. In this system, not only is the mapping determined by perceptual pro-
cesses: the perceptual processes are in turn influenced by the mapping process. Representations
are built up gradually by means of this continual interaction between perception and mapping. If
a particular representation seems appropriate for a given mapping, then that representation
continues to be developed, while the mapping continues to be fleshed out. If the representation
seems less promising, then alternative directions are explored by the perceptual process. It is of
the essence that the processes of perception and mapping are interleaved at all stages. Gradually,
an appropriate analogy emerges, based on structured representations that dovetail with the final
mapping. We will examine this system in greater detail shortly.
Such a system is very different from the traditional approach, which assumes the representa-
tion-building process to have been completed, and which concentrates on the mapping process in
isolation. But in order to be able to deal with the great flexibility of human perception and
representation, analogy researchers must integrate high-level perceptual processes into their work.
We believe that the use of hand-coded, rigid representations will in the long run prove to be a
dead end, and that flexible, context-dependent, easily adaptable representations will be recognized
as an essential part of any accurate model of cognition.
Finally, we should note that the problems we have outlined here are by no means unique to
the modeling of analogical thought. The hand-coding of representations is endemic in traditional
AI. Any program that uses pre-built representations for a particular task could be subject to such
a “representation module” argument similar to that given above. For most purposes in cognitive
science, an integration of task-oriented processes with those of perception and representation will
4 A Model that Integrates High-Level Perception with Analogy-Making
A model of high-level perception is clearly desirable, but a major obstacle lies in the way.
For any model of high-level perception to get off the ground, it must be firmly founded on a base
of low-level perception. But the sheer amount of information available in the real world makes
the problem of low-level perception an exceedingly complex one, and success in this area has
understandably been quite limited. Low-level perception poses so many problems that for now,
the modeling of full-fledged high-level perception of the real world is a distant goal. The gap
between the lowest level of perception (cells on the retina, pixels on the screen, waveforms of
sound) and the highest level (conceptual processes operating on complex structured
representations) is at present too wide to bridge.
This does not mean, however, that one must admit defeat. There is another route to the goal.
The real world may be too complex, but if one restricts the domain, some understanding may be
within our grasp. If, instead of using the real world, one carefully creates a simpler, artificial
world in which to study high-level perception, the problems become more tractable. In the
absence of large amounts of pixel-by-pixel information, one is led much more quickly to the
problems of high-level perception, which can then be studied in their own right.
Such restricted domains, or microdomains, can be the source of much insight. Scientists in all
fields throughout history have chosen or crafted idealized domains to study particular phenomena.
When researchers attempt to take on the full complexity of the real world without first having
some grounding in simpler domains, it often proves to be a misguided enterprise. Unfortunately,
microdomains have fallen out of favor in artificial intelligence. The “real world” modeling that
has replaced them, while ambitious, has often led to misleading claims (as in the case of
BACON), or to limited models (as we saw with models of analogy). Furthermore, while “real
world” representations have impressive labels—such as “atom” or “solar system”—attached to
them, these labels conceal the fact that the representations are nothing but simple structures in
predicate logic or a similar framework. Programs like BACON and SME are really working in
stripped-down domains of certain highly idealized logical forms—their domains merely appear to
have the complexity of the real world, thanks to the English words attached to these forms.
While microdomains may superficially seem less impressive than “real world” domains, the
fact that they are explicitly idealized worlds allows the issues under study to be thrown into clear
relief—something that generally speaking is not possible in a full-scale real-world problem. Once
we have some understanding of the way cognitive processes work in a restricted domain, we will
have made genuine progress towards understanding the same phenomena in the unrestricted real
The model that we will examine here works in a domain of alphabetical letter-strings. This
domain is simple enough that the problems of low-level perception are avoided, but complex
enough that the main issues in high-level perception arise and can be studied. The model, the
“Copycat” program (Hofstadter 1984; Mitchell 1990; Hofstadter and Mitchell 1992), is capable of
building up its own representations of situations in this domain, and does so in a flexible, context-
dependent manner. Along the way, many of the central problems of high-level perception are
dealt with, using mechanisms that have a much broader range of application than just this
particular domain. Such a model may well serve as the basis for a later, more general model of
This highly parallel, non-deterministic architecture builds its own representations and finds
appropriate analogies by means of the continual interaction of perceptual structuring-agents with
an associative concept network. It is this interaction between perceptual structures and the
concept network that helps the model capture part of the flexibility of human thought. The
Copycat program is a model of both high-level perception and analogical thought, and it uses the
integrated approach to situation perception and mapping that we have been advocating.
The architecture could be said to fall somewhere on the spectrum between the connectionist
and symbolic approaches to artificial intelligence, sharing some of the advantages of each. On
the one hand, like connectionist models, Copycat consists of many local, bottom-up, parallel
processes from whose collective action higher-level understanding emerges. On the other hand, it
shares with symbolic models the ability to deal with complex hierarchically-structured
We shall use Copycat to illustrate possible mechanisms for dealing with five important
problems in perception and analogy. These are:
• the gradual building-up of representations;
• the role of top-down and contextual influences;
• the integration of perception and mapping;
• the exploration of many possible paths toward a representation;
• the radical restructuring of perceptions, when necessary.
The description of Copycat given here will necessarily be brief and oversimplified, but further
details are available elsewhere (Hofstadter 1984; Mitchell and Hofstadter 1990; Mitchell 1990;
Hofstadter and Mitchell 1992).
The Copycat domain
The task of the Copycat program is to make analogies between strings of letters. For instance,
it is clear to most people that abc and iijjkkll share common structure at some level. The goal of
the program is to capture this by building, for each string, a representation that highlights this
common structure, and by finding correspondences between the two representations.
The program uses the result of this correspondence-making to solve analogy problems of the
following form: “If abc changes to abd, what does iijjkkll change to?” Once the program has
discovered common structure in the two strings abc and iijjkkll, deciding that the letter a in the
first corresponds to the group ii in the second and that c corresponds to ll, it is relatively
straightforward for it to deduce that the best answer must be iijjkkmm. The difficult task for the
program—the part requiring high-level perception—is to build the representations in the first
place. We will shortly examine in more detail just how these representations are built.
Before we begin a discussion of the details of Copycat, we should note that the program
knows nothing about the shapes of letters, their sounds, or their roles in the English language. It
does know the order of the letters in the alphabet, both forwards and backwards (to the program,
the alphabet is in no sense “circular”). The alphabet consists of 26 “platonic” letter entities, each
with no explicit relation to anything except its immediate neighbors. When instances of these
simple concepts, the letters, are combined into strings of various lengths, quite complex
“situations” can result. The task of the program is to perceive structure in these situations, and to
use this structure to make good analogies.
The architecture used by the program, incidentally, is applicable much more widely than to
just the particular domain used here. For instance, the architecture has also been implemented to
deal with the problem of perceiving structure and making analogies involving the dinner
implements on a tabletop (a microdomain with a more “real world” feel) (French 1988). An
application involving perception of the shapes and styles of visual letterforms, and generation of
new letterforms sharing the given style, has also been proposed (Hofstadter et al, 1987).
Building up representations
The philosophy behind the model under discussion is that high-level perception emerges as a
product of many independent but cooperating processes running in parallel. The system is at first
confronted with a raw situation, about which it knows almost nothing. Then a number of
perceptual agents swarm over and examine the situation, each discovering small amounts of local
structure adding incrementally to the system’s perception, until finally a global understanding of
the situation emerges.
These perceptual agents, called codelets, are the basic elements of Copycat’s perceptual
processing. Each codelet is a small piece of code, designed to perform a particular type of task.
Some codelets seek to establish relations between objects; some chunk objects that have been per-
ceived as related into groups; some are responsible for describing objects in particular ways; some
build the correspondences that determine the analogy; and there are various others. Each codelet
works locally on a small part of the situation. There are many codelets waiting to run at any
given time, in a pool from which one is chosen nondeterministically at every cycle. The codelets
often compete with each other, and some may even break structures that others have built up, but
eventually a coherent representation emerges.
When it starts to process a problem in the letter-string domain, Copycat knows very little
about the particular problem at hand. It is faced with three strings, of which it knows only the
platonic type of each letter, which letters are spatially adjacent to each other, and which letters are
leftmost, rightmost, and middle in each string. The building-up of representations of these strings
i i k k
Figure 2. Examples of perceptual structures built by Copycat.
and of their interrelationships is the task of codelets. Given a string such as ppqqrrss, one
codelet might notice that the first and second letters are both instances of the same platonic letter-
type (“P”), and build a “sameness” bond between them. Another might notice that the physically
adjacent letters r and s are in fact alphabetical neighbors, and build a “successor” bond between
them. Another “grouping” codelet might chunk the two bonded letters p into a group, which can
be regarded at least temporarily as a unit. After many such codelets have run, a highly structured
representation of the situation emerges, which might, for instance, see the string as a sequence of
four chunks of two letters each, with the “alphabetic successor” relation connecting each chunk
with its right neighbor. Figure 2 gives an stripped-down example of Copycat’s perceptual struc-
Different types of codelets may come into play at different stages of a run. Certain types of
codelets, for example, can run only after certain types of structures have been discovered. In this
way, the codelets cause structure to be built up gradually, and in a context-sensitive manner. Due
to the highly nondeterministic selection of codelets, several directions can be simultaneously
explored by the perceptual process. Given the string abbccd, for instance, some codelets might
try to organize it as a sequence of “sameness” groups, a-bb-cc-d, while others might simulta-
neously try to organize it quite differently as a sequence of “successor” groups, ab-bc-cd.
Eventually, the program is likely to focus on one or the other of these possibilities, but because of
the nondeterminism, no specific behavior can be absolutely guaranteed in advance. However,
Copycat usually comes up in the end with highly structured and cognitively plausible representa-
tions of situations it is given.
The role of context and top-down influences
As we have seen, one of the most important features of high-level perception is its sensitivity
to context. A model of the perceptual process that proceeds in a manner that disregards context
will necessarily be inflexible.
The Copycat model captures the dependence of perception on contextual features by means of
an associative concept-network (Figure 3), the Slipnet, which interacts continually with the
perceptual process. Each node in this network corresponds to a concept that might be relevant in
the letter-string domain, and each node can be activated to a varying degree depending on the per-
ceived relevance of the corresponding concept to the given situation. As a particular feature of
the situation is noted, the node representing the concept that corresponds to the feature is activat-
ed in the concept network. In turn, the activation of this concept has a biasing effect on the per-
ceptual processing that follows. Specifically, it causes the creation of some number of associated
codelets, which are placed in the pool of codelets waiting to run. For instance, if the node
corresponding to the concept of “alphabetic successor” is activated in the Slipnet, then several
codelets will be spawned whose task is to look for successorship relations elsewhere in the situa-
Further, the activation of certain nodes means that it is more likely that associated perceptual
processes will succeed. If the “successor” node is highly active, for example, not only is it more
likely that codelets that try to build successorship relations will be spawned, but it is also more
likely that once they run, they—rather than some competing type of codelet—will succeed in
building a lasting relation as part of the representation. In both of these ways, perceptual
processing that has already been completed can have a contextual, top-down influence on
subsequent processing through activation of concepts in the Slipnet.
For instance, in the string kkrrtt it is likely that the two r’s will be perceived as a “sameness
group” (a group all of whose members are the same); such a perception will be reinforced by the
presence of two similar groups on either side, which will activate the node representing the
concept of “sameness group”. On the other hand, in the string abcijkpqrrst, the presence of the
groups abc and ijk will cause the node representing “successor group” (a group consisting of
alphabetically successive letters) to be active, making it more likely that pqr and rst will be
perceived in the same way. Here, then, it is more likely that the two adjacent r’s will be
perceived separately, as parts of two different “successor groups” of three letters each. The way
in which two neighboring r’s are perceived (i.e., as grouped or not) is highly dependent on the
context that surrounds them, and this contextual dependence is mediated by the Slipnet.
This two-way interaction between the perceptual process and the concept network is a
combination of top-down and bottom-up processing. The perceptual work performed by the
Figure 3. A small portion of Copycat’s concept network, the Slipnet.
codelets is an inherently bottom-up process, achieved by competing and cooperating agents each
of which acts locally. The Slipnet, however, by modulating the action of the codelets, acts as a
top-down influence on this bottom-up process. The Slipnet can thus be regarded as a dynamic
controller, allowing global properties such as the activation of concepts to influence the local
action of perceptual agents. This top-down influence is vitally important, as it it ensures that per-
ceptual processes do not go on independently of the system’s understanding of the global context.
Integration of perception and mapping in analogy-making
We have already discussed the necessity of a fully-integrated system of perceptual processing
and mapping in analogy-making. The Copycat model recognizes this imperative. The situation-
perception and mapping processes take place simultaneously. Certain codelets are responsible for
building up representations of the given situations, while others are responsible for building up a
mapping between the two. Codelets of both types are in the pool together.
In the early stages of a run, perceptual codelets start to build up representations of the
individual situations. After some structure has been built up, other types of codelets begin to
make tentative mappings between the structures. From then on, the situation-perception and
mapping processes proceed hand in hand. As more structure is built within the situations, the
mapping becomes more sophisticated, and aspects of the evolving mapping in turn exert pressure
on the developing perceptions of the situations.
Consider, for example, two analogies involving the string ppqrss. If we are trying to find an
analogy between this and, say, the string aamnxx, then the most successful mapping is likely to
map the group of p’s to the group of a’s, the group of s’s to the group of x’s, and qr to the succes-
sor group mn. The most natural way to perceive the second string is in the form aa-mn-xx, and
this in turn affects the way that the first string is perceived, as three two-letter groups in the form
pp-qr-ss. On the other hand, if we are trying to find an analogy between ppqrss and the string
aijklx, then recognition of the successor group ijkl inside the latter string is likely to arouse per-
ceptual biases toward seeking successor relations and groups, so that the system will be likely to
spot the successor group pqrs within ppqrss, and to map one successor group to the other. This
leads to the original string being perceived as p-pqrs-s, which maps in a natural way to a-ijkl-x.
Thus we can see that different mappings act as different contexts to evoke quite different
perceptions of the same string of letters. This is essentially what was going on in the two
analogies described earlier involving DNA. In both cases, the representation of a given situation
is made not in isolation, but under the influence of a particular mapping.
We should note that the Copycat model makes no important distinction between structures
built for the purpose of situation-perception (such as bonds between adjacent letters, or groups of
letters), and those built for the purpose of mapping (such as correspondences between letters or
groups in the two strings). Both types of structure are built up gradually over time, and both
contribute to the program’s current understanding of the overall situation. The mapping
structures can themselves be regarded as perceptual structures: the mapping is simply an
understanding of the analogy as a whole.
Exploring different paths and converging on a solution
A model of perception should, in principle, be able to explore all of the different plausible
ways in which a situation might be organized into a representation. Many representations may be
possible, but some will be more appropriate than others. Copycat’s architecture of competing
codelets allows for the exploration of many different pathways toward a final structure. Different
codelets will often begin to build up structures that are incompatible with each other. This is
good—it is desirable that many possibilities be explored. In the end, however, the program must
converge on one particular representation of a given situation.
In Copycat, the goal of homing in on a particular solution is aided by the mechanism of
computational temperature. This is a number that measures the amount and quality of structure
present in the current representation of the situation. Relevant structures here include bonds,
groups, and correspondences, as well as some others. The term “quality of structure” refers to
how well different parts of the structure cohere with each other. Computational temperature is
used to control the amount of randomness in the local action of codelets. If a large amount of
good structure has been built up, the temperature will be low and the amount of randomness
allowed will be small. Under these circumstances, the system will proceed in a fairly
deterministic way, meaning that it sticks closely to a single pathway with few rival side-
explorations being considered. On the other hand, if there is little good structure, the temperature
will be high, which will lead to diverse random explorations being carried out by codelets.
At the start of a run, before any structure has been built, the temperature is maximally high, so
the system will behave in a very random way. This means that many different pathways will be
explored in parallel by the perceptual processes. If no promising structural organization emerges,
then the temperature will remain high and many different possibilities will continue to be
explored. Gradually, in most situations, certain structures will prove more promising, and these
are likely to form the basis of the final representation. At any given moment, a single structural
view is dominant, representing the system’s current most coherent worldview, but many other
tentative structures may be present in the background, competing with it.
As good structures build up, the temperature gradually falls and so the system’s exploratory
behavior becomes less random. This mean that structures that have already been built have a
lowered chance of being replaced by new ones and are thus favored. The more coherent a global
structure, the less likely parts of it are to be broken. As structure builds up and temperature falls,
the system concentrates more and more on developing the structure that exists. Eventually, the
program will converge on a good representation of a given situation. In practice, Copycat fre-
quently comes up in different runs with different representations for the same situation, but these
representations usually seem to be cognitively plausible. Its final “solutions” to various analogy
problems are distributed in a fashion qualitatively similar to the distributions found with human
subjects (Mitchell 1990; Hofstadter and Mitchell 1992).
The process of exploring many possibilities and gradually focusing on the most promising
ones has been called a “parallel terraced scan” (Hofstadter 1984; Hofstadter and Mitchell 1992).
The process is akin to the solution to the “two-armed bandit” problem (Holland 1975) where a
gambler has access to two slot machines with fixed but distinct probabilities of payoff. These
payoff probabilities are initially unknown to the gambler, who wishes to maximize payoffs over a
series of trials. The best strategy is to start by sampling both machines equally, but to gradually
focus one’s resources probabilistically on the machine that appears to be giving the better payoff.
The Copycat program has to perform an analogous task. To function flexibly, it has to sample
many representational possibilities and choose those that promise to lead to the most coherent
worldview, gradually settling down to a fixed representation of the situation. In both the two-
armed bandit and in Copycat, it takes time for certain possibilities to emerge as the most fruitful,
and a biased stochastic sampling technique is optimal for this purpose.
Sometimes representations that have been built up for a given situation turn out to be
inappropriate, in that they do not lead to a solution to the problem at hand. When people find
themselves in this situation, they need to be able to completely restructure their representations,
so that new ones can evolve that are more adequate for the current task. Maier’s two-string
experiment provides an example of radical restructuring; people have to forget about their initial
representation of a pair of pliers as a tool for bending things, and instead see it as a heavy weight.
In Copycat, when a representation has been built up, the temperature has necessarily gone
down, which makes it difficult to change to another representation. But it is obviously not advan-
tageous for the program to keep a representation that does not led to a solution. For this reason,
the program has a special set of mechanisms to deal with such situations.
For instance, when the program is given the analogy “If abc changes to abd, what does xyz
change to?”, it usually builds a detailed representation of abc and xyz as successor groups, and
quite reasonably maps one string to the other accordingly (a maps to x, c maps to z, etc). But
now, when it tries to carry out the transformation for xyz that it feels is analogous to abc
becoming abd, it finds itself blocked, since it is impossible to take the successor of the letter z
(the Copycat alphabet is non-circular). The program has hit a “snag”; the only way to deal with it
is to find an alternative representation of the situation.
The program deals with the problem firstly by raising the temperature. The temperature
shoots up to its maximal value. This produces a great deal of randomness at the codelet level.
Secondly, “breaker codelets” are brought in for the express purpose of destroying representations.
The result is that many representations that have been carefully built up are broken down. At the
same time, much activation is poured into the concept representing the source of the snag—the
concept Z—and much perceptual attention is focused on the specific z inside xyz (that is, it be-
comes very salient and attracts many codelets). This causes a significant change in the represen-
tation-building process the second time around. To make a long story short, the program is
thereby able to come up with a completely new representation of the situation, where abc is still
perceived as a successor group, but xyz is re-perceived as a predecessor group, starting from z
and going backwards. Under this new representation, the a in the first string is mapped to the z in
Now if the program attempts to complete its task, it discovers that the appropriate
transformation on xyz is to take the predecessor of the leftmost letter, and it comes up with the
insightful answer wyz. (We should stress that the program, being nondeterministic, does not
always or even consistently come up with this answer. The answer xyd is actually given more
often than wyz.) Further details are given by Mitchell and Hofstadter (1990).
This process of reperception can be regarded as a stripped-down model of a “scientific
revolution” (Kuhn 1970) in a microdomain. According to this view, when a field of science
comes up against a problem it cannot solve, clamor and confusion result in the field, culminating
in a “paradigm shift” where the problem is viewed in a completely different way. With the new
worldview, the problems may be straightforward. The radical restructuring involved in the above
letter-string problem seems quite analogous to this scientific process.
What Copycat doesn’t do
Some have argued that in employing hand-coded mechanisms such as codelets and the
Slipnet, Copycat is guilty of 20-20 hindsight in much the same fashion as BACON and SME. But
there is a large difference: BACON and SME use fixed representations, whereas Copycat devel-
ops flexible representations using fixed perceptual mechanisms. Whereas we have seen that the
use of fixed representations is cognitively implausible, it is clear that human beings at any given
time have a fixed repertoire of mechanisms available to the perceptual process. One might justifi-
ably ask where these mechanisms, and the corresponding mechanisms in Copycat, come from, but
this would be a question about learning. Copycat is not intended as a model of learning: its per-
formance, for instance, does not improve from one run to the next. It would be a very interesting
further step to incorporate learning processes into Copycat, but at present the program should be
taken as a model of the perceptual processes in an individual agent at a particular time.
There are other aspects of human cognition that are not incorporated into Copycat. For in-
stance, there is nothing in Copycat that corresponds to the messy low-level perception that goes
on in the visual and auditory systems. It might well be argued that just as high-level perception
exerts a strong influence on and is intertwined with later cognitive processing, so low-level per-
ception is equally intertwined with high-level perception. In the end, a complete model of high-
level perception will have to take low-level perception into account, but for now the complexity
of this task means that key features of the high-level perceptual processes must be studied in iso-
lation from their low-level base.
The Tabletop program (French and Hofstadter 1991; French 1992) takes a few steps towards
lower-level perception, in that it must make analogies between visual structures in a two-dimen-
sional world, although this world is still highly idealized. There is also a small amount of related
work in AI that attempts to combine perceptual and cognitive processes. It is interesting to note
that in this work, microdomains are almost always used. Chapman’s “Sonja” program (Chapman
1991), for instance, functions in the world of a video game. Starting from simple graphical infor-
mation, it develops representations of the situation around it and takes appropriate action. As in
Tabletop, the input to Sonja’s perceptual processes is a little more complex than in Copycat, so
that these processes can justifiably be claimed to be a model of “intermediate vision” (more close-
ly tied to the visual modality than Copycat’s high-level mechanisms, but still abstracting away
from the messy low-level details), although the representations developed are less sophisticated
than Copycat’s. Along similar lines, Shrager (1990) has investigated the central role of perceptu-
al processes in scientific thought, and has developed a program that builds up representations in
the domain of understanding the operation of a laser, starting from idealized two-dimensional in-
It may sometimes be tempting to regard perception as not truly “cognitive”, something that
can be walled off from higher processes, allowing researchers to study such processes without
getting their hands dirtied by the complexity of perceptual processes. But this is almost certainly
a mistake. Cognition is infused with perception. This has been recognized in psychology for
decades, and in philosophy for longer, but artificial-intelligence research has been slow to pay
Two hundred years ago, Kant provocatively suggested an intimate connection between
concepts and perception. “Concepts without percepts”, he wrote, “are empty; percepts without
concepts are blind.” In this paper we have tried to demonstrate just how true this statement is,
and just how dependent on each other conceptual and perceptual processes are in helping people
make sense of their world.
“Concepts without percepts are empty.” Research in artificial intelligence has often tried to
model concepts while ignoring perception. But as we have seen, high-level perceptual processes
lie at the heart of human cognitive abilities. Cognition cannot succeed without processes that
build up appropriate representations. Whether one is studying analogy-making, scientific
discovery, or some other area of cognition, it is a mistake to try to skim off conceptual processes
from the perceptual substrate on which they rest, and with which they are tightly intermeshed.
“Percepts without concepts are blind.” Our perception of any given situation is guided by
constant top-down influence from the conceptual level. Without this conceptual influence, the
representations that result from such perception will be rigid, inflexible, and unable to adapt to the
problems provided by many different contexts. The flexibility of human perception derives from
constant interaction with the conceptual level. We hope that the model of concept-based
perception that we have described goes some way towards drawing these levels together.
Recognizing the centrality of perceptual processes makes artificial intelligence more difficult,
but it also makes it more interesting. Integrating perceptual processes into a cognitive model
leads to flexible representations, and flexible representations lead to flexible actions. This is a
fact that has only recently begun to permeate artificial intelligence, through such models as
connectionist networks, classifier systems, and the architecture presented here. Future advances
in the understanding of cognition and of perception are likely to go hand in hand, for the two
types of process are inextricably intertwined.
Bruner, J. (1957). On perceptual readiness. Psychological Review, 64: 123-152.
Burstein, M. (1986). Concept formation by incremental analogical reasoning and debugging. In
R. S. Michalski, J. G. Carbonell, and T. M. Mitchell (eds.), Machine learning: An artificial
intelligence approach, Vol. 2 (Los Altos, CA: Morgan Kaufmann).
Carbonell, J. G. (1986). Learning by analogy: Formulating and generalizing plans from past
experience. In R. S. Michalski, J. G. Carbonell, and T. M. Mitchell (eds.), Machine learning: An
artificial intelligence approach, Vol. 2 (Los Altos, CA: Morgan Kaufmann).
Chapman, D. (1991). Vision, instruction, and action. Cambridge, MA: MIT Press.
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14: 179-212.
Evans, T. G. (1968). A program for the solution of a class of geometric-analogy intelligence-test
questions. In M. Minsky (ed.), Semantic information processing (Cambridge, MA: MIT Press).
Falkenhainer, B., Forbus, K. D., and Gentner, D. (1990). The structure-mapping engine.
Artificial Intelligence, 41: 1-63.
Fodor, J. A. (1983). The modularity of mind (Cambridge, MA: MIT Press).
French, R. M., and Hofstadter, D. R. (1991). Tabletop: A stochastic, emergent model of
analogy-making. Proceedings of the 13th annual conference of the Cognitive Science Society.
Hillsdale, NJ: Lawrence Erlbaum.
French, R. M. (1992). Tabletop: A stochastic, emergent model of analogy-making. Doctoral dis-
sertation, University of Michigan.
Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive
Hofstadter, D. R. (1984). The Copycat project: An experiment in nondeterminism and creative
analogies. M.I.T. AI Laboratory memo 755.
Hofstadter, D. R., and Mitchell, M. (1992). An overview of the Copycat project. In K. J.
Holyoak and J. Barnden (eds.), Connectionist Approaches to Analogy, Metaphor, and Case-Based
Reasoning (Norwood, NJ: Ablex).
Hofstadter, D. R., Mitchell, M., and French, R. M. (1987). Fluid concepts and creative analogies:
A theory and its computer implementation. CRCC Technical Report 18, Indiana University.
Holland, J. H. (1975). Adaptation in natural and artificial systems (Ann Arbor, MI: University of
Holland, J. H. (1986). Escaping brittleness: The possibilities of general-purpose learning
algorithms applied to parallel rule-based systems. In R. S. Michalski, J. G. Carbonell, and T. M.
Mitchell (eds.), Machine learning: An artificial intelligence approach, Vol. 2 (Los Altos, CA:
Holyoak, K. J. and Thagard, P. (1989). Analogical mapping by constraint satisfaction. Cognitive
Science, 13: 295-355.
James, W. (1890). The principles of psychology (Henry Holt & Co).
Kedar-Cabelli, S. (1988). Towards a computational model of purpose-directed analogy. In A.
Prieditis (ed.), Analogica (Los Altos, CA: Morgan Kaufmann).
Kuhn, T. (1970). The structure of scientific revolutions (2nd edition) (Chicago: University of
Lakoff, G. (1987). Women, fire and dangerous things (Chicago: University of Chicago Press).
Langley, P., Simon, H. A., Bradshaw, G. L., and Zytkow, J. M. (1987). Scientific discovery:
Computational explorations of the creative process (Cambridge, MA: MIT Press).
Maier, N. R. F (1931). Reasoning in humans: II. The solution of a problem and its appearance in
consciousness. Cognitive Psychology, 12: 181-194.
Marr, D. (1977). Artificial intelligence—a personal view. Artificial Intelligence, 9: 37-48.
Mitchell, M. (1990). Copycat: A computer model of high-level perception and conceptual
slippage in analogy-making. Doctoral dissertation, University of Michigan.
Mitchell, M., and Hofstadter, D. R. (1990). The emergence of understanding in a computer
model of concepts and analogy-making. Physica D, 42: 322-334.
Newell, A. and H. A. Simon (1976). Computer science as empirical inquiry: Symbols and
search. Communications of the Association for Computing Machinery, 19: 113-126.
Pylyshyn, Z. (1980). Cognition and computation. Behavioral and Brain Sciences, 3: 111-132.
Qin, Y., and Simon, H. A. (1990). Laboratory replication of scientific discovery processes.
Cognitive Science 14: 281-310.
Rumelhart, D. E., McClelland, J. L., and the PDP Research Group (1986). Parallel distributed
processing (Cambridge, MA: MIT Press).
Shrager, J. (1990). Commonsense perception and the psychology of theory formation. In J.
Shrager and P. Langley (eds.) Computational models of scientific discovery and theory formation
(San Mateo, CA: Morgan Kaufmann).
Simon, H. A. (1989). The scientist as problem solver. In D. Klahr and K. Kotovsky (eds.)
Complex information processing (Hillsdale, NJ: Lawrence Erlbaum).
Winston, P. H. (1982). Learning new principles from precedents and exercises. Artificial
Intelligence 19: 321-350.
V.S. Ramachandran and William Hirstein
The Science of Art
A Neurological Theory of Aesthetic Experience
We present a theory of human artistic experience and the neural mechanisms that mediate
it. Any theory of art (or, indeed, any aspect of human nature) has to ideally have three
components. (a) The logic of art: whether there are universal rules or principles; (b) The
evolutionary rationale: why did these rules evolve and why do they have the form that they
do; (c) What is the brain circuitry involved? Our paper begins with a quest for artistic uni-
versals and proposes a list of ‘Eight laws of artistic experience’ — a set of heuristics that
artists either consciously or unconsciously deploy to optimally titillate the visual areas of
the brain. One of these principles is a psychological phenomenon called the peak shift
effect: If a rat is rewarded for discriminating a rectangle from a square, it will respond
even more vigorously to a rectangle that is longer and skinnier that the prototype. We sug-
gest that this principle explains not only caricatures, but many other aspects of art. Exam-
ple: An evocative sketch of a female nude may be one which selectively accentuates those
feminine form-attributes that allow one to discriminate it from a male figure; a Boucher, a
Van Gogh, or a Monet may be a caricature in ‘colour space’ rather than form space. Even
abstract art may employ ‘supernormal’ stimuli to excite form areas in the brain more
strongly than natural stimuli. Second, we suggest that grouping is a very basic principle.
The different extrastriate visual areas may have evolved specifically to extract correla
tions in different domains (e.g. form, depth, colour), and discovering and linking multiple
features (‘grouping’) into unitary clusters — objects — is facilitated and reinforced by
direct connections from these areas to limbic structures. In general, when object-like enti
ties are partially discerned at any stage in the visual hierarchy, messages are sent back to
earlier stages to alert them to certain locations or features in order to look for additional
evidence for the object (and these processes may be facilitated by direct limbic activa
tion). Finally, given constraints on allocation of attentional resources, art is most appeal
ing if it produces heightened activity in a single dimension (e.g. through the peak shift
principle or through grouping) rather than redundant activation of multiple modules.
This idea may help explain the effectiveness of outline drawings and sketches, the savant
syndrome in autists, and the sudden emergence of artistic talent in fronto-temporal
dementia. In addition to these three basic principles we propose five others, constituting a
total of ‘eight laws of aesthetic experience’ (analogous to the Buddha’s eightfold path to
Journal of Consciousness Studies, 6, No. 6-7, 1999, pp. 15–51
Correspondence: V.S. Ramachandran, Center For Brain and Cognition, University of California,
San Diego, La Jolla, CA 92093-0109, USA.
Journal of Consciousness Studies
Due to file size constraints the plates accompanying the electronic
version of this article can be downloaded free of charge from
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
‘Everyone wants to understand art. Why not try to understand the song of a bird?’
If a Martian ethologist were to land on earth and watch us humans, he would be puz
zled by many aspects of human nature, but surely art — our propensity to create and
enjoy paintings and sculpture — would be among the most puzzling. What biological
function could this mysterious behaviour possible serve? Cultural factors undoubt
edly influence what kind of art a person enjoys — be it a Rembrandt, a Monet, a
Rodin, a Picasso, a Chola bronze, a Moghul miniature, or a Ming Dynasty vase. But,
even if beauty is largely in the eye of the beholder, might there be some sort of univer
sal rule or ‘deep structure’, underlying all artistic experience? The details may vary
from culture to culture and may be influenced by the way one is raised, but it doesn’t
follow that there is no genetically specified mechanism — a common denominator
underlying all types of art. We recently proposed such a mechanism (Ramachandran
and Blakeslee, 1998), and we now present a more detailed version of this hypothesis
and suggest some new experiments. These may be the very first experiments ever
designed to empirically investigate the question of how the brain responds to art.
Many consider art to be a celebration of human individuality and to that extent it
may seem like a travesty to even search for universals. Indeed theories of visual art
range from curious anarchist views (or even worse, ‘anything goes’) to the idea that
art provides the only antidote to the absurdity or our existence — the only escape, per-
haps, from this vale of tears (Penrose, 1973). Our approach to art, in this essay, will be
to begin by simply making a list of all those attributes of pictures that people gener-
ally find attractive. Notwithstanding the Dada movement, we can then ask, Is there a
common pattern underlying these apparently dissimilar attributes, and if so, why is
this pattern pleasing to us? What is the survival value, if any, of art?
But first let us clear up some common misconceptions about visual art. When the
English colonizers first arrived in India they were offended by the erotic nudes in
temples; the hips and breasts were grossly hypertrophied, the waist abnormally thin
Similarly the Rajasthani and Moghul miniature paintings were considered
primitive because they lacked perspective. In making this judgement they were, of
course, unconsciously comparing Indian art with the ideals of Western representa
tional art — Renaissance art in particular. What is odd about this criticism though, is
that it misses the whole point of art. The purpose of art, surely, is not merely to depict
or represent reality — for that can be accomplished very easily with a camera — but
to enhance, transcend, or indeed even to distort reality. The word ‘rasa’ appears
repeatedly in Indian art manuals and has no literal translation, but roughly it means
‘the very essence of.’ So a sculptor in India, for example, might try to portray the rasa
of childhood (Plate 2), or the rasa of romantic love, or sexual ecstasy (Plate 3), or
feminine grace and perfection (Plate 4). The artist is striving, in these images, to
strongly evoke a direct emotional response of a specific kind. In Western art, the
‘discovery’ of non-representational abstract art had to await the arrival of Picasso.
His nudes were also grotesquely distorted — both eyes on one side of the face for
example. Yet when Picasso did it, the Western art critics heralded his attempts to
THE SCIENCE OF ART 16
 Plates for the article can be downloaded free from www.imprint.co.uk/rama
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
‘transcend perspective’ as a profound new discovery — even though both Indian and
African art had anticipated this style by several centuries!
We suggest in this essay that artists either consciously or unconsciously deploy cer
tain rules or principles (we call them laws) to titillate the visual areas of the brain.
Some of these laws, we believe, are original to this article — at least in the context of
art. Others (such as grouping) have been known for a long time and can be found in
any art manual, but the question of why a given principle should be effective is rarely
raised: the principle is usually just presented as a rule-of-thumb. In this essay we try
to present all (or many) of these laws together and provide a coherent biological
framework, for only when they are all considered simultaneously and viewed in a bio
logical context do they begin to make sense. There are in fact three cornerstones to
our argument. First, what might loosely be called the ‘internal logic’ of the phenome
non (what we call ‘laws’ in this essay). Second, the evolutionary rationale: the ques
tion of why the laws evolved and have that particular form (e.g. grouping facilitates
object perception). And third, the neurophysiology (e.g. grouping occurs in extrastri
ate areas and is facilitated by synchronization of spikes and direct limbic activation).
All three of these need to be in place — and must inform each other — before we can
claim to have ‘understood’ any complex manifestation of human nature — such as
art. Many earlier discussions of art, in our view, suffer from the shortcoming that they
view the problem from just one or two of these perspectives.
We should clarify at the outset that many aspects of art will not be discussed in this
article — such as matters concerning style. Indeed it may well be that much of art
really has to do with aggressive marketing and hype, and this inevitably introduces an
element of arbitrariness that complicates the picture enormously. Furthermore the
artistic ‘universals’ that we shall consider are not going to provide an instant formula
for distinguishing ‘tacky’ or ‘tourist’ art, that hangs in the lobbies of business execu-
tives, from the genuine thing — even though a really gifted artist could do so instantly
— and until we can do that we can hardly claim to have ‘understood’ art. Yet despite
these reservations, we do believe that there is at least a component to art — however
small — that IS lawful and can be analysed in accordance with the principles or laws
outlined here. Although we initially proposed these ‘laws’ in a playful spirit, we were
persuaded that there is enough merit in them to warrant publication in a philosophical
journal. If the essay succeeds in stimulating a dialogue between artists, visual physi
ologists and evolutionary biologists, it will have adequately served its purpose.
The Essence of Art and the Peak Shift Principle
Hindu artists often speak of conveying the rasa, or ‘essence’, of something in order to
evoke a specific mood in the observer. But what exactly does this mean? What does it
mean to ‘capture the very essence’ of something in order to ‘evoke a direct emotional
response’? The answer to these questions, it turns out, provides the key to understand
ing what art really is. Indeed, as we shall see, what the artist tries to do (either con
sciously or unconsciously) is to not only capture the essence of something but also to
amplify it in order to more powerfully activate the same neural mechanisms that
would be activated by the original object. As the physiologist Zeki (1998) has elo
quently noted, it may not be a coincidence that the ability of the artist to abstract the
‘essential features’ of an image and discard redundant information is essentially iden
tical to what the visual areas themselves have evolved to do.
17 V.S. RAMACHANDRAN AND W. HIRSTEIN
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
Consider the peak shift effect — a well-known principle in animal discrimination
learning. If a rat is taught to discriminate a square from a rectangle (of say, 3:2 aspect
ratio) and rewarded for the rectangle, it will soon learn to respond more frequently to
the rectangle. Paradoxically, however, the rat’s response to a rectangle that is even
longer and skinnier (say, of aspect ratio 4:1) is even greater than it was to the original
prototype on which it was trained. This curious result implies that what the rat is
learning is not a prototype but a rule, i.e. rectangularity. We shall argue in this essay
that this principle holds the key for understanding the evocativeness of much of visual
art. We are not arguing that it’s the only principle, but that it is likely to be one of a
small subset of such principles underlying artistic experience.
How does this principle — the peak shift effect — relate to human pattern recogni
tion and aesthetic preference? Consider the way in which a skilled cartoonist pro
duces a caricature of a famous face, say Nixon’s. What he does (unconsciously) is to
take the average of all faces, subtract the average from Nixon’s face (to get the differ
ence between Nixon’s face and all others) and then amplify the differences to produce
a caricature. The final result, of course, is a drawing that is even more Nixon-like than
the original. The artist has amplified the differences that characterize Nixon’s face in
the same way that an even skinnier rectangle is an amplified version of the original
prototype that the rat is exposed to. This leads us to our first aphorism: ‘All art is cari-
cature’. (This is not literally true, of course, but as we shall see, it is true surprisingly
often.) And the same principle that applies for recognizing faces applies to all aspects
of form recognition. It might seem a bit strange to regard caricatures as art but take a
second look at the Chola bronze — the accentuated hips and bust of the Goddess Par-
vati (Plate 1) and you will see at once that what you have here is essentially a carica-
ture of the female form. There may be neurons in the brain that represent sensuous,
rotund feminine form as opposed to angular masculine form and the artist has chosen
to amplify the ‘very essence’ (the rasa) of being feminine by moving the image even
further along toward the feminine end of the female/male spectrum (Plate 4). The
result of these amplifications is a ‘super stimulus’ in the domain of male/female
differences. It is interesting, in this regard, that the earliest known forms of art are
often caricatures of one sort or another; e.g. prehistoric cave art depicting animals like
bison and mammoths, or the famous Venus ‘fertility’ figures.
As a further example, look at the pair of nudes in Plate 5, a sculpture from Northern
India (circa 800
AD). No normal woman can adopt such contorted postures and yet the
sculpture is incredibly evocative — beautiful — capturing the rasa of feminine poise
and grace. To explain how he achieves this effect, consider the fact that certain pos
tures are impossible (and unlikely) among men but possible in women because of cer
tain anatomical differences that impose constraints on what can or cannot be done.
Now in our view what the artist has done here is to subtract the male posture from the
female posture to produce a caricature in ‘posture space’ thereby amplifying ‘fem
inine posture’ and producing a correspondingly high limbic activation. The same can
be said of the dancer in Plate 6 or for the amorous couple (Plate 7). Again, even
though these particular, highly stylized anatomical poses are impossible (or unlikely)
it is very evocative of the ‘Sringara Rasa’ or ‘Kama rasa’ (sexual and amorous
ecstasy) because the artist is providing a ‘caricature’ that exaggerates the amorous
pose. It is as though the artist was been able to intuitively access and powerfully
stimulate neural mechanisms in the brain that represent ‘amorousness’.
THE SCIENCE OF ART 18
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
A posture space might be realized in the form of a large set of remembered postures
of people one has observed. (Whether one might expect such a memory mapping to
exist in the ‘dorsal’ stream of visual processing, which connects with the agent’s own
body representations, or the ‘ventral’ stream, known to be used for face perception, is
an interesting question; perhaps the answer is, both). There is an obvious need to con
nect these posture representations to the limbic system: it is quite imperative that I
recognize an attack posture, a posture — or body position — which beckons me, or
one which indicates sadness or depression, etc. The sculptors of Plates 5 and 6 relied
on this represented posture space in creating their works. The sculptor knows, con
sciously or not, that the sight of those postures will evoke a certain sort of limbic acti
vation when the posture is successfully represented in the posture space system — he
tells a story in this medium, we might say.
Until now we have considered caricatures in the form domain, but we know from
the pioneering work of many physiologists (Zeki, 1980; see also Livingstone and
Hubel, 1987; Allman & Kaas, 1971; Van Essen & Maunsell, 1980) that the primate
brain has specialized modules concerned with other visual modalities such as colour
depth and motion. Perhaps the artist can generate caricatures by exploiting the peak
shift effect along dimensions other than form space, e.g., in ‘colour space’ or ‘motion
space’. For instance consider the striking examples of the plump, cherub-faced nudes
that Boucher is so famous for. Apart from emphasizing feminine, neotonous baby-
like features (a peak shift in the masculine/feminine facial features domain) notice
how the skin tones are exaggerated to produce an unrealistic and absurd ‘healthy’
pink flush. In doing this, one could argue he is producing a caricature in colour space,
particularly the colours pertaining to male/female differences in skin tone. Another
artist, Robert, on the other hand, pays little attention to colour or even to form, but
tends to deliberately overemphasize the textural attributes of his objects, be they
bricks, leaves, soil, or cloth. And other artists have deliberately exaggerated (‘caric-
atured’ or produced peak shifts in) shading, highlights, illumination etc to an extent
that would never occur in a real image. Even music may involve generating peak
shifts in certain primitive, passionate primate vocalizations such as a separation cry;
the emotional response to such sounds may be partially hard-wired in our brains.
A potential objection to this scheme is that it is not always obvious in a given pic
ture what the artist is trying to caricature, but this is not an insurmountable objection.
Ethologists have long known that a seagull chick will beg for food by pecking at its
mother’s beak. Remarkably, it will peck just as vigorously at a disembodied beak with
no mother attached or even a brown stick with a red dot at the end (the gull’s beak has
a vivid red spot near the tip). The stick with the red dot is an example of a ‘releasing
stimulus’ or ‘trigger feature’ since, as far as the chick’s visual system is concerned
this stimulus is as good as the entire mother bird. What is even more remarkable,
though, was Tinbergen’s discovery (Tinbergen, 1954) that a very long, thin brown
stick, with three red stripes at the end is even more effective in eliciting pecks than the
original beak, even though it looks nothing like a beak to a human observer.
The gull’s form recognition areas are obviously wired-up in such a way that Tinber
gen had inadvertently produced a super stimulus, or a caricature in ‘beak space’ (e.g.
the neurons in the gull’s brain might embody the rule ‘more red contour the better’).
Indeed, if there were an art gallery in the world of the seagull, this ‘super beak’ would
qualify as a great work of art — a Picasso. Likewise, it is possible that some types of
19 V.S. RAMACHANDRAN AND W. HIRSTEIN
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
art such as cubism are activating brain mechanisms in such a way as to tap into or
even caricature certain innate form primitives which we do not yet fully understand.
At present we have no idea what the ‘form primitives’ used by the human visual path
ways are, but we suggest that many artists may be unconsciously producing height
ened activity in the ‘form areas’ in a manner that is not obvious to the conscious mind,
just as it isn’t obvious why a long stick with three red stripes is a ‘super beak’. Even
the sunflowers of Van Gogh or the water lilies of Monet may be the equivalent — in
colour space — of the stick with the three stripes, in that they excite the visual neu
rons that represent colour memories of those flowers even more effectively than a real
sunflower or water lily might.
There is also clearly a mnemonic component of aesthetic perception, including, the
autobiographical memory of the artist, and of her viewer, as well as the viewer’s more
general ‘cognitive stock’ brought to his encounter with the work. This general cogni
tive stock includes the viewer’s memory of his encounters with the painting’s etio
logical forebears, including those works that the artist himself was aware of. Often
paintings contain homages to earlier artists and this concept of homage fits what we
have said about caricature: the later artist makes a caricature of his acknowledged
predecessor, but a loving one, rather than the ridiculing practised by the editorial car-
toonist. Perhaps some movements in the history of art can be understood as driven by
a logic of peak shift: the new art form finds and amplifies the essence of a previous
one (sometimes many years previous, in the case of Picasso and African art).
THE SCIENCE OF ART 20
 Another manifestation of this principle can be seen in the florid sexual displays of birds — that we find
so attractive. It is very likely, as suggested by Darwin, that the grotesque exaggeration of these
displays, for example the magnificent wings of the birds of paradise, is a manifestation of the peak shift
effect during mate choice — sexual selection caused by birds of each generation preferring caricatures
of the opposite sex to mate with (just as humans lean toward Playboy pinups and Chippendale
dancers). Indeed we have recently suggested (Ramachandran and Blakeslee, 1998) that many aspect of
morphological evolution (not just ‘secondary sexual characteristics’ or florid ‘ethological releasers’
and threat displays) may be the outcome of runaway selection, based on the peak shift principle. The
result would be not only the emergence and ‘quantization’ of new species, but also a progressive and
almost comical ‘caricaturization’ of phylogenetic trends of precisely the kind one sees in the evolution
of elephants or ankylosaurs. Even the quirks of fashion design (e.g. corsets becoming absurdly narrow,
shoes becoming smaller and smaller in ancient China, shrinking miniskirts) become more
comprehensible in terms of this perceptual principle. One wonders, also, whether the striking
resemblance between the accumulation of jewellery, shoes and other brightly coloured objects by
humans and the collections of bright pebbles, berries and feathers by bowerbirds building their
enormous nests is entirely coincidental.
 Lastly, consider the evolution of facial expressions. Darwin proposed that a ‘threat gesture’ may have
evolved from the real facial movements one makes before attacking a victim — i.e. the baring of
canines, etc. The same movement may eventually become divorced from the actual act and begin to
serve as a communication of intent — a threat. If the peak shift principle were to operate in the
recipient’s brain it is easy to see how such a ritualized signal would become progressively amplified
across generations. Darwin had a difficult time, however, explaining why gestures such as sadness
(instead of joy) seem to involve the opposite movement of facial features — e.g. lowering the corners
of the mouth — and he came up with his somewhat ad hoc ‘principle of antithesis’, which states that
somehow the opposite emotion is automatically linked to the opposite facial movements. We would
suggest, instead, that the principle of antithesis is, once again, an indirect result of the recipient’s brain
applying the peak shift principle. Once the organism has circuitry in its brain that says K is normal and J
is a smile, then it may follow automatically that L is the expression of the opposite emotion — sadness.
Whether this particular conjecture is correct or not we believe that emotional expressions analysed in
terms of the peak shift effect may begin to make more sense than they have in the past.
Another layer of complexity here is that even the perception of complex postures or actions may
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
Perceptual Grouping and Binding is Directly Reinforcing
One of the main functions of ‘early vision’ (mediated by the thirty or so extrastriate
visual areas) is to discover and delineate objects in the visual field (Marr, 1981;
Ramachandran, 1990; Pinker, 1998; Shepard, 1981) and for doing this the visual
areas rely, once again, on extracting correlations. For instance if a set of randomly
placed spots A is superimposed on another set of randomly placed dots B, they are
seen to mingle to form just a single enormous cluster. But if you now move one of the
clusters (say, A) then all the dots are instantly glued or bound together perceptually to
create an object that is clearly separate from the background cluster B. Similarly if
cluster A is made of red dots (and B is of green dots) we have no difficulty in segregat
ing them instantly.
This brings us to our second point. The very process of discovering correlations
and of ‘binding’ correlated features to create unitary objects or events must be rein
forcing for the organism — in order to provide incentive for discovering such correla
tions (Ramachandran and Blakeslee, 1998). Consider the famous hidden face (Fig. 1)
21 V.S. RAMACHANDRAN AND W. HIRSTEIN
A jumble of splotches or a face?
[If you have difficulty ‘seeing’ the
face, try looking through half-closed
eyes — Editor.]
Initially seen as a jumble of splotches, once the Dalma-
tian is seen, its spots are grouped together — a pleasing
effect, caused perhaps by activation of the limbic sys-
tem by temporal lobe cortex.
require the observer to somehow internally re-enact or ‘rehearse’ the action before it is identified. For
instance, patients with apraxia (inability to perform complex skilled movements resulting from
damage to the left supramarginal gyrus) often, paradoxically, have difficulty perceiving and
recognizing complex actions performed by others. Also, there are cells in the frontal lobes thought to
be involved in the production of complex movements but which also fire when the animal perceives
the same movements performed by a the experimenter (di Pellegrino et al., 1992). This finding —
together with the peak shift effect — would help account for Darwin’s ‘principle of antithesis’, which
would otherwise seem completely mysterious. Such cells may also be activated powerfully when
viewing dynamic figural representations such as the ‘Dancing Devi’ (Plate 6).
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
or Dalmatian dog photo (Fig. 2). This is seen initially as a random jumble of
splotches. The number of potential groupings of these splotches is infinite but once
the dog is seen your visual system links only a subset of these splotches together and it
is impossible not to ‘hold on’ to this group of linked splotches. Indeed the discovery
of the dog and the linking of the dog-relevant splotches generates a pleasant ‘aha’
sensation. In ‘colour space’ the equivalent of this would be wearing a blue scarf with
red flowers if you are wearing a red skirt; the perceptual grouping of the red flowers
and your red skirt is aesthetically pleasing — as any fashion designer will tell you.
These examples suggest that there may be direct links in the brain between the
processes that discover such correlations and the limbic areas which give rise to the
pleasurable ‘rewarding’ sensations associated with ‘feature binding’. So when you
choose a blue matte to frame your painting in order to ‘pick up’ flecks of blue in the
painting you are indirectly tapping into these mechanisms.
How is such grouping achieved? As noted above, the primate brain has over two
dozen visual areas each of which is concerned with a different visual attribute such as
motion, colour, depth, form, etc. These areas are probably concerned with extracting
correlations in ‘higher dimensional’ spaces — such as ‘colour space’ or ‘motion
space’. In a regular topographic map — e.g., in area 17 — features that are close
together in physical space are also close together in the brain (which is all that is
meant by ‘map’). But now think of non-topographic maps — say a map of ‘colour
space’ — in which points that are close together in wavelength are mapped close
together in the colour area of the brain even though they may be distant from each
other physically (Barlow, 1986). Such proximity along different feature dimensions
may be useful for perceptual grouping and ‘binding’ of features that are similar
within that dimension.
This argument sounds plausible, but why should the outputs of separate vision
modules — space, colour, depth, motion, etc. — be sent directly to the limbic system
before further processing has occurred? Why not delay the reinforcement produced
by limbic activation until the object has actually been identified by neurons in infero
temporal cortex? After all, the various Gestalt grouping processes are thought to
occur autonomously as a result of computations within each module itself (Marr,
1981) without benefit of either cross-module or ‘top down’ influences — so why
bother hooking up the separate modules themselves to limbic regions? One resolution
of this paradox might simply be that the serial, hierarchical, ‘bucket brigade’ model
of vision is seriously flawed and that eliminating ambiguity, segmenting the scene
and discovering and identifying objects do indeed rely on top down processes — at
least to some situations (Churchland et al., 1994). The visual system is often called
upon to segment the scene, delineate figure from ground and recognize objects in
very noisy environments — i.e., to defeat camouflage — and this might be easier to
accomplish if a limbic ‘reinforcement’ signal is not only fed back to early vision once
an object has been completely identified, but is evoked at each and every stage in
processing as soon as a partial ‘consistency’ and binding is achieved. This would
explain why we say ‘aha’ when the Dalmatian is finally seen in Fig. 2 — and why it is
difficult to revert back to seeing merely splotches once the dog is seen as a whole: that
particular percept is powerfully reinforced (Ramachandran and Blakeslee, 1998). In
other words, even though the grouping may be initially based on autonomous process
in each module (Marr, 1981), once a cluster of features becomes perceptually salient
THE SCIENCE OF ART 22
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
as a ‘chunk’ with boundaries (i.e. an object), it may send a signal to the limbic centres
which in turn causes you to ‘hold on’ to that chunk to facilitate further computation.
There is physiological evidence that grouping of features leads to synchronization of
the spikes (action potentials) of neurons that extract those features (Singer and Gray,
1995; Crick and Koch, 1998) and perhaps it is this synchrony that allows the signal to
be sent to the limbic pathways. (This, by the way, may be one reason why musical
consonance often involves harmonics — for example, a C-major chord — which, for
physical reasons would tend to emerge from a single object, whereas dissonant notes
are likely to emerge from two or more separate objects.)
The key idea, then, is the following (and it applies to many of our laws, not just
grouping). Given the limited attentional resources in the brain and limited neural
space for competing representations, at every stage in processing there is generated a
‘Look here, there is a clue to something potentially object-like’ signal that produces
limbic activation and draws your attention to that region (or feature) , thereby facili
tating the processing of those regions or features at earlier stages. Furthermore, par
tial ‘solutions’ or conjectures to perceptual problems are fed back from every level in
the hierarchy to every earlier module to impose a small bias in processing and the
final percept emerges from such progressive ‘bootstrapping’ (Ramachandran et al.,
1998). As noted above, consistency between partial high-level ‘hypotheses’ and ear
lier low-level ensembles also generates a pleasant sensation — e.g. the Dalmatian
dog ‘hypothesis’ encourages the binding of corresponding splotches which, in turn,
further consolidate the ‘dog-like’ nature of the final percept and we feel good when it
all finally clicks in place. And what the artist tries to do, is to tease the system with as
many of these ‘potential object’ clues as possible — an idea that would help explain
why grouping and ‘perceptual problem solving (see below) are both frequently
exploited by artists and fashion designers.
The notion that art exploits grouping principles is of course not new (Gombrich,
1973; Arnheim, 1956; Penrose, 1973), but what is novel here is our claim that the
grouping doesn’t always occur ‘spontaneously’; that out of a temporary binding a
signal sent to the limbic system to reinforce the binding, and this is one source of the
aesthetic experience. For example, in Fig. 3, there are two possible stable organiza
23 V.S. RAMACHANDRAN AND W. HIRSTEIN
Gestalt grouping principles. The tokens can be grouped either on the basis of ‘proximity’ (which
produces hourglasses), or ‘closure’. The latter organization is more stable and pleasing to the eye.
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
tions, one with hourglasses, and one with closure and most people find the latter
organization more pleasing than the former because the limbic activation is stronger
with this closure-based object-like percept. When artists speak of composition, or
grouping, they are probably unconsciously tapping into these very same principles.
One obvious prediction that emerges from this theory is that patients with Kluver-
Bucy syndrome — caused by bilateral amygdala destruction — should not only dis
play problems in recognizing objects (visual agnosia) but also in segmenting them
out from noisy backgrounds, an idea that would be relatively easy to test
Isolating a Single Module and Allocating Attention
The third important principle (in addition to peak shift and binding) is the need to iso
late a single visual modality before you amplify the signal in that modality. For
instance, this is why an outline drawing or sketch is more effective as ‘art’ than a full
colour photograph. This seems initially counterintuitive since one would expect that
the richer the cues available in the object the stronger the recognition signal and asso
ciated limbic activation. This apparent objection can be overcome, however, once one
realizes that there are obvious constraints on the allocation of attentional resources to
different visual modules. Isolating a single area (such as ‘form’ or ‘depth’ in the case
of caricature or Indian art) allows one to direct attention more effectively to this one
source of information, thereby allowing you to notice the ‘enhancements’ introduced
by the artist. (And that in turn would amplify the limbic activation and reinforcement
produced by those enhancements). Consider a full-colour illustration of Nixon, with
depth, shading, skin tones and blemishes, etc. What is unique about Nixon is the form
of his face (as amplified by the caricature) but the skin tone — even though it makes
the picture more human-like — doesn’t contribute to making him ‘Nixon like’ and
therefore actually detracts from the efficacy of the form cues. Consequently, one
would predict that a full colour photo of Nixon would actually be less aesthetically
pleasing than a sketchy outline drawing that captures the essential ‘Nixon-like’ attrib
utes of his face.
The idea that outlines are effective in art is hardly new. It has been repeated ad
nauseum by many authors, ever since David Hubel and Torsten Wiesel (1979) origi
nally pointed out that this principle may reflect the fact that cells in the visual path
ways are adequately stimulated by edges and are indifferent to homogeneous regions.
However this would only explain why one can get away with just using outlines —
not why outlines are actually more effective than a full colour half tone photo, which,
after all, has more information. We would argue that when the colour, skin texture,
etc. are not critical for defining the identity of the object in question (e.g. Nixon’s
face) then the extra redundant information can actually distract your limited atten
tional resources away from the defining attributes of that object. Hence the aphorism
‘more is less’ in Art.
Additional evidence for this view comes from the ‘savant syndrome’ — autistic
children who are ‘retarded’ and yet produce beautiful drawings. The animal drawings
of the eight-year old artist Nadia, for instance, are almost as aesthetically pleasing as
THE SCIENCE OF ART 24
 If this theory is correct, we can make the counterintuitive prediction that activation of the face area in
the brain — measured by fMRI — should, paradoxically, be greater for an outline sketch of a face than
for a full colour photo of a face.
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
those of Leonardo da Vinci! (Plate 8). We would argue that this is because the funda
mental disorder in autism is a distortion of the ‘salience landscape’; they shut out
many important sensory channels thereby allowing them to deploy all their atten
tional resources on a single channel; e.g., in ‘visual form representation’ channel in
the case of Nadia. This idea is also consistent with the ingenious theory of Snyder
(Snyder and Thomas, 1997), that savants are able to ‘directly access’ the outputs of
some of their early vision modules because they are less ‘concept driven’: the concep
tual impoverishment that produces autism also, paradoxically, gives them better
access to earlier processes in vision. And finally, we would suggest that the ‘iso-
lation’ principle also explains the efflorescence of artistic talent that is occasionally
seen in fronto-temporal dementia in adults: a clinical phenomenon that is currently
being studied intensively in our laboratory.
These ideas allow us to make certain novel predictions: If you put luminous dots on
a person’s joints and film him or her walking in complete darkness, the complex
motion trajectories of the dots are usually sufficient to evoke a compelling impression
of a walking person — the so-called Johansson effect (Johansson, 1975). Indeed, it is
often possible to tell the sex of the person by watching the gait. However, although
these movies are often comical, they are not necessarily pleasing aesthetically. We
would argue that this is because even though you have isolated a cue along a single
dimension, i.e., motion, this isn’t really a caricature in motion space. To produce a
work of art, you would need to subtract the female motion trajectories from the male
and amplify the difference. Whether this would result in a pleasing work of kinetic art
remains to be seen.
Contrast Extraction is Reinforcing
Grouping, as we have already noted, is an important principle, but the extraction of
features prior to grouping — which involves discarding redundant information and
extracting contrast — is also ‘reinforcing’. Cells in the retina, lateral geniculate body
(a relay station in the brain) and in the visual cortex respond mainly to edges (step
changes in luminance) but not to homogeneous surface colours; so a line drawing or
cartoon stimulates these cells as effectively as a ‘half tone’ photograph. What is fre
quently overlooked though is that such contrast extractions — as with grouping —
may be intrinsically pleasing to the eye (hence the efficacy of line drawings). Again,
though, if contrast is extracted autonomously by cells in the very earliest stages of
processing, why should the process be rewarding in itself? We suggest that the answer
once again has to do with the allocation of attention. Information (in the Shannon
sense) exists mainly in regions of change — e.g. edges — and it makes sense that such
regions would, therefore, be more attention grabbing — more ‘interesting’ — than
homogeneous areas. So it may not be coincidental that what the cells find interesting
is also what the organism as a whole finds interesting and perhaps in some circum
stances ‘interesting’ translates into ‘pleasing’.
For the same reason, contrast along many other stimulus dimensions besides lumi
nance, such as colour or texture, has been exploited by artists (for instance, colour
contrast is exploited by Matisse), and indeed there are cells in the different visual
areas specialized for colour contrast, or motion contrast (Allman and Kaas, 1971).
Furthermore, just as one can speak of a peak shift principle along very abstract dimen
25 V.S. RAMACHANDRAN AND W. HIRSTEIN
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
THE SCIENCE OF ART 26
Camouflage. Notice that the boundary between the two types of texture (vertical vs. horizontal
lines) is clearly visible on the upper pattern (A), but is masked by the luminance boundaries on the
lower (B). (Based on M.J. Morgan; personal communication).
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
sions, contrast can also emerge in dimensions other than luminance or colour. For
instance, a nude wearing baroque (antique) gold jewellery (and nothing else) is aes
thetically much more pleasing than a completely nude woman or one wearing both
jewellery and clothes, presumably because the homogeneity and smoothness of the
naked skin contrasts sharply with the ornateness and rich texture of the jewellery.
Whether the analogy between luminance contrast extracted by cells in the brain and
the contrast between jewels and naked skin is just a play of words or a deep unifying
principle is a question that cannot be answered given what we know about the brain.
But we do know that the attention grabbing effect of contrast must be a very important
principle in nature, since it is often used as a camouflage device by both predators and
their prey. For instance, in Fig. 4A, a texture border is very visible, but in Fig. 4B it is
almost ‘invisible’, camouflaged by the colour (black/white) borders that grab the
lion’s share of your attention.
At first the two principles we have just considered seem antithetical; grouping on
the basis of similarity is rewarding, but if so how can contrast (the very opposite of
grouping) also be rewarding? One clue comes from the fact that the two mechanisms
have different spatial constraints; grouping can occur between similar features (e.g.
colour or motion) even if they are far apart in space (e.g., the spots on the nose and tail
of a leopard). Contrast, on the other hand, usually occurs between dissimilar features
that are physically close together. Thus even though the two processes seem to be
inconsistent, they actually complement one another in that they are both concerned
with the discovery of objects — which is the main goal of vision. (Contrast extraction
is concerned with the object’s boundaries whereas grouping allows recovery of the
object’s surfaces and, indirectly, of its boundaries as well). It is easy to see then why
the two should be mutually reinforcing and rewarding to the organism.
Symmetry, of course, is also aesthetically pleasing as is well known to any Islamic
artist (or indeed to any child looking through a kaleidoscope) and it is thought to be
extracted very early in visual processing (Julesz, 1971). Since most biologically
important objects — such as predator, prey or mate are symmetrical, it may serve as
an early-warning system to grab our attention to facilitate further processing of the
symmetrical entity until it is fully recognised. As such, this principle complements
the other laws described in this essay; it is geared towards discovering ‘interesting’
object-like entities in the world.
Intriguingly, it has recently been shown experimentally that when choosing a mate,
animals and humans prefer symmetrical over asymmetrical ones and evolutionary
biologists have argued that this is because parasitic infestation — detrimental to fer
tility — often produces lopsided, asymmetrical growth and development. If so, it is
hardly surprising that we have a built-in aesthetic preference for symmetry.
The Generic Viewpoint and the Bayesian Logic of Perception
Another less well known principle relates to what AI researchers refer to as ‘the
generic viewpoint’ principle, which is illustrated in Fig. 5A and B and Fig. 6A and B.
In Fig. 5A most people see a square occluding the corner of another square, even
27 V.S. RAMACHANDRAN AND W. HIRSTEIN
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
THE SCIENCE OF ART 28
One square is seen as occluding the other. It is hard to see A as B viewed from a unique vantage
point. The brain ‘prefers’ the generic view.
The flat hexagon with radiating spokes could be a cube but is never seen as one. The ‘generic’ inter
pretation is again the brain’s preferred one.
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
29 V.S. RAMACHANDRAN AND W. HIRSTEIN
The brain’s abhorrence of ‘suspicious coincidences’ (a phrase used by Horace Barlow). Figure B is
pleasing, but A is distasteful to the eye.
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
though it could theoretically be Fig. B seen from a unique view point. The reason is
that there is an infinite set of viewpoints that could produce the class of retinal images
resembling A, but only a single, unique viewpoint that could produce retinal image A,
given the objects in B. Consequently, the visual system rejects the latter interpretation
as being highly improbable and prefers to see A as occlusion. (The same principle
applies to 6A and B; A could depict an outline of a cube seen from one specific van
tage point, but people usually see it as a flat hexagon with spokes radiating from the
middle.) These examples illustrate the universal Bayesian logic of all perception:
your visual system abhors interpretations which rely on a unique vantage point and
favours a generic one or, more generally, it abhors suspicious coincidences (Barlow,
1980). For this reason, Fig. 7B is pleasing, whereas 7A is unattractive (palm tree and
hills). So if an artist is trying to please the eye, he too, should avoid coincidences,
such as those in 7A and 6B. Yet one must be cautious in saying this since every now
and then — given the perverse nature of art and artists — a pleasing effect can be pro
duced by violating this principle rather than adhering to it. For instance, there is a
Picasso nude in which the improbability of the arm’s outline exactly coinciding with
that of the torso grabs the viewer’s attention — and is arguably attractive to him!
We hasten to add that the principles we have discussed so far certainly do not
exhaust all types of artistic experience. We have hardly touched on the purely sym-
bolic or allegorical aspects of some types of paintings or sculpture, or on surrealism
and modern abstract art (e.g., minimalists such as Kandinsky), not to mention
‘counter’ art such as the Dada movement. Also very puzzling is the question of why a
nude hidden by a diaphanous veil is more alluring than one seen directly in the flesh,
as pointed out by Ernst Gombrich (1973). It is as though an object discovered after a
struggle is more pleasing than one that is instantly obvious. The reason for this is
obscure but perhaps a mechanism of this kind ensures that the struggle itself is re-
inforcing — so that you don’t give up too easily — whether looking for a leopard
behind foliage or a mate hidden in the mist. On the other hand, we suspect that surre
alist art really doesn’t have much to do with visual representations per se but involves
playing with links between vision and semantics, thereby taking it closer to the meta
phorical ambiguities of poetry and language than to the purely visual appeal of a
Picasso, a Rodin, or a Chola bronze. For example, in his erotic masterpiece ‘Young
virgin autosodomised by her own chastity’ (1954), Dali has used the male penis to
represent the female buttocks and genitalia. The medium and message ‘resonate’
since they both pertain to sex but they are also in subtle conflict since they depict
‘opposite’ sexes! The result is an image pleasing on many levels simultaneously. This
playful, whimsical, aspect of art, often involving the humorous juxtaposition of com
plementary — or sometimes even incongruous — elements, is perhaps the most enig
matic aspect of our aesthetic experience, one which we have hardly touched upon in
this essay. Another aspect of art that we have not dealt with is style, although one can
see how once a style or trend is set in motion the peak shift principle can certainly help
Art as Metaphor
The use of visual metaphors in art is well known. For instance, in Plate 9, the languor
ous, sensuous pose of the woman mimics the tree branch above — the curves match
her curves and perhaps the tree’s fertility is a metaphor for her youthfulness. (Just as
THE SCIENCE OF ART 30
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
in Plate 4, the fruit in the tree echoes the curve of the breasts as well as the abdomen.)
There are countless examples of this sort in both Eastern and Western art and yet the
question is rarely raised as to why visual ‘puns’ or allegories should be aesthetically
pleasing. A metaphor is a mental tunnel between two concepts or percepts that appear
grossly dissimilar on the surface. When Shakespeare says ‘Juliet is the sun,’ he is
appealing to the fact that they are both warm and nurturing (not the fact that they both
reside in our solar system!). But, again, why should grasping an analogy of this kind
be so rewarding to us? Perhaps the use of a simple concrete example (or one that is
easily visualised, such as the sun) allows us to ignore irrelevant, potentially distract
ing aspects of an idea or percept (e.g. Juliet has nails, teeth and legs) and enables us to
‘highlight’ the crucial aspects (radiance and warmth) that she shares with the sun but
not with other women. Whether this is purely a device for effective communication,
or a basic cognitive mechanism for encoding the world more economically, remains
to be seen. The latter hypothesis may well be correct. There are many paintings that
instantly evoke an emotional response long before the metaphor is made explicit by
an art critic. This suggests that the metaphor is effective even before one is conscious
of it, implying that it might be a basic principle for achieving economy of coding
rather than a rhetorical device. This is also true of poetic metaphors, as when Shake-
speare says of Juliet, ‘Death, that has sucked the honey of thy breath’: the phrase is
incredibly powerful well before one becomes consciously aware of the hidden anal-
ogy between the ‘sting of death’ and the bee’s sting and the subtle sexual connota-
tions of ‘sucking’ and ‘breath’.
Classifying objects into categories is obviously vital for survival, e.g. prey vs.
predator, edible vs. inedible, male vs. female, etc. Seeing a deep similarity—acom-
mon denominator as it were — between disparate entities is the basis of all concept
formation whether the concepts are perceptual (‘Juliet’) or more abstract (‘love’).
Philosophers often make a distinction between categories or ‘types’ and ‘tokens’ —
the exemplars of a type — (e.g. ‘ducks’ vs. ‘that duck’). Being able to transcend
tokens to create types is an essential step in setting up a new perceptual category.
Being able to see the hidden similarities between successive distinct episodes allows
you to link or bind these episodes to create a single super-ordinate category, e.g., sev
eral viewer-centred representations of a chair are linked to form a viewer-
independent abstract representation of ‘chairness’. Consequently, the discovery of
similarities and the linking of superficially dissimilar events would lead to a limbic
activation — in order to ensure that the process is rewarding. It is this basic mechan
ism that one taps into, whether with puns, poetry, or visual art.
Partial support for this view comes from the observation that these mechanisms can
go awry in certain neurological disorders. In Capgras syndrome, for instance, connec
tions from the visual ‘face region’ in the inferotemporal cortex to the amygdala (a part
of the limbic system where activation leads to emotions) are severed so that a familiar
face no longer evokes a warm fuzzy emotional response (Hirstein and
Ramachandran, 1997). Remarkably, some Capgras patients are no longer able to link
successive views of a person’s face to create more general perceptual category of that
particular face. We suggested that in the absence of limbic activation — the ‘glow’ of
recognition — there is no incentive for the brain to link successive views of a face, so
that the patient treats a single person as several people. When we showed our Capgras
patient DS different photos of the same person, he claimed that the photos were of dif
31 V.S. RAMACHANDRAN AND W. HIRSTEIN
Copyright (c) Imprint Academic 2005
For personal use only -- not for reproduction
ferent people, who merely resembled each other! One might predict, therefore, that
patients like DS would also experience difficulty in appreciating the metaphorical
nuances of art, but such a prediction is not easy to test.
An Experimental Test
We conclude by taking up the final test of any theory: does it lead to counterintuitive
predictions that can be tested experimentally? One approach — albeit a laborious one
— would be to do ‘psychophysics’ on a