This Is a Publication of the American Association for Artificial Intelligence

This Is a Publication of The American Association for Artificial Intelligence

This electronic document has been retrieved from the American Association for Artificial Intelligence 445 Burgess Drive Menlo Park, California 94025 (415) 328-3123 (415) 321-4457 [email protected] http://www.aaai.org

(For membership information, consult our web page)

The material herein is copyrighted material. It may not be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from AAAI. Presidential Address What Are Intelligence? And Why? 1996 AAAI Presidential Address

Randall Davis

■ This article, derived from the 1996 American Asso- that take seriously their visual nature. I speculate ciation for Artificial Intelligence Presidential as well that thinking may be a form of reliving, Address, explores the notion of intelligence from a that re-acting out what we have experienced is one variety of perspectives and finds that it “are” many powerful way to think about and solve problems things. It has, for example, been interpreted in a in the world. In this view, thinking is not simply variety of ways even within our own field, ranging the decontextualized manipulation of abstract from the logical view (intelligence as part of math- symbols, powerful though that may be. Instead, ematical logic) to the psychological view (intelli- some significant part of our thinking may be the gence as an empirical phenomenon of the natural reuse or simulation of our experiences in the envi- world) to a variety of others. One goal of this arti- ronment. In keeping with this, I suggest that it cle is to go back to basics, reviewing the things that may prove useful to marry the concreteness of rea- we, individually and collectively, have taken as soning in a model with the power that arises from given, in part because we have taken multiple dif- reasoning abstractly. ferent and sometimes inconsistent things for granted. I believe it will prove useful to expose the tacit assumptions, models, and metaphors that we elax, there’s no mistake in the title. I’ve carry around as a way of understanding both what we’re about and why we sometimes seem to be at indulged a bit of British-English that I’ve odds with one another. Ralways found intriguing: the use of the Intelligence are also many things in the sense that plural verb with collective nouns (as in it is a product of evolution. Our physical bodies are “Oxford have won the Thames regatta”). in many ways overdetermined, unnecessarily com- The selection of verb sense is purposeful and plex, and inefficiently designed, that is, the pre- captures one of the main themes of the article: dictable product of the blind search that is evolu- I want to consider intelligence as a collective tion. What’s manifestly true of our anatomy is also noun. I want to see what we in AI have thought likely true of our cognitive architecture. Natural of it and review the multiple ways in which intelligence is unlikely to be limited by principles we’ve conceived of it. My intention is to make of parsimony and is likely to be overdetermined, explicit the assumptions, metaphors, and mod- unnecessarily complex, and inefficiently designed. In this sense, intelligence are many things because els that underlie our multiple conceptions. it is composed of the many elements that have I intend to go back to basics here, as a way of been thrown together over evolutionary time- reminding us of the things that we, individual- scales. I suggest that in the face of that, searching ly and collectively, have taken as given, in part for minimalism and elegance may be a diversion, because we have taken multiple different, and for it simply may not be there. Somewhat more sometimes inconsistent, things for granted. I crudely put: The human mind is a 400,000-year- believe it will prove useful to expose the tacit old legacy application…and you expected to find assumptions, models, and metaphors that we structured programming? carry around, as a way of understanding both I end with a number of speculations, suggesting what we’re about and why we sometimes seem that there are some niches in the design space of to be at odds with one another. That’s the first intelligences that are currently underexplored. One example is the view that thinking is in part part of the article. visual, and hence it might prove useful to develop In the second part of the article, I’ll ask you representations and reasoning mechanisms that to come along on a natural history tour—I’m reason with diagrams (not just about them) and going to take you away, back to a time around

Mathematical Psychology Biology Statistics Economics Logic Aristotle Descartes Boole James Laplace Bentham, Pareto Frege Bernoulli Friedman Peano Hebb Lashley Bayes Goedel Bruner Rosenblatt Post Miller Ashby Tversky Von Neumann Church Newell Lettvin Kahneman Simon Turing Simon McCulloch, Pitts Raiffa Davis Heubel, Weisel Putnam Robinson

LOGIC SOAR CONNECTIONIS M Causal networks Rational agents PROLOG Knowledge-based systems A-life Frames

Table 1. Views of Intelligent Reasoning and Their Intellectual Origins.

4 million years ago when the first hominids iors lie at its core. Four behaviors are common- arose and consider how intelligence came to ly used to distinguish intelligent behavior be. We’ll take an evolutionary view, consider from instinct and stimulus-response associa- intelligence as a natural phenomenon, and ask tions: (1) prediction, (2) response to change, why it arose. The vague answer—that it provid- (3) intentional action, and (4) reasoning. ed enhanced survival—turns out not to be very One core capability is our ability to predict informative; so, we’ll ask, why is intelligence, the future, that is, to imagine how things and more important, what does that tell us might turn out rather than have to try them. about how we might proceed in AI? The essential issue here is imagining, that is, the The third part of the article is concerned disconnection of thought and action. That dis- with what we might call inhuman problem solv- connection gives us the ability to imagine the ing; it explores to what degree intelligence is a consequences of an action before, or instead human monopoly. In this part of the article, AI of, experiencing it, the ability, as Popper and learns about the birds and the bees: What kinds Raimund (1985) put it, to have our hypotheses of animal intelligence are there, and does that, die in our stead. The second element— too, inform our search for human intelligence? response to change—is an essential character- I’ll end by considering how we might istic that distinguishes intelligent action from expand our view, expand our exploration of inalterable instinct or conditioned reflexes. intelligence by exploring aspects of it that Intentional action refers to having a goal and have received too little attention. AI has been selecting actions appropriate to achieving the doing some amount of consolidation over the goal. Finally, by reasoning, I mean starting with past few years, so it may well be time to specu- some collection of facts and adding to it by late where the next interesting and provoca- any inference method. tive leaps might be made. Five Views of Reasoning Fundamental Elements AI has of course explored all these in a variety of ways. Yet even if we focus in on just one of If AI is centrally concerned with intelligence, them—intelligent reasoning—it soon becomes we ought to start by considering what behav- clear that there have been a multitude of

92 AI MAGAZINE Presidential Address answers explored within AI as to what we cles of the sorts only the gods could make), mean by that, that is, what we mean when we could in fact be married to algebra, a form of say intelligent reasoning. Given the relative calculation, something mere mortals could do. youth of our field, the answers have often By the time of Leibnitz, the agenda is quite come from work in other fields. Five fields in specific and telling: He sought nothing less particular—(1) mathematical logic, (2) psy- than a calculus of thought, one that would per- chology, (3) biology, (4) statistics, and (5) eco- mit the resolution of all human disagreement nomics—have provided the inspiration for five with the simple invocation “let us compute.” distinguishable notions of what constitutes By this time, there is a clear and concrete belief intelligent reasoning (table 1).1 that as Euclid’s once godlike and unreachable One view, historically derived from mathe- geometry could be captured with algebra, so matical logic, makes the assumption that intel- some (or perhaps any) variety of that ephemer- ligent reasoning is some variety of formal calcu- al stuff called thought might be captured in lation, typically, deduction; the modern calculation, specifically logical deduction. exemplars of this view in AI are the logicists. A In the nineteenth century, Boole provided second view, rooted in work in psychology, sees the basis for propositional calculus in his Laws reasoning as a characteristic human behavior of Thought; later work by Frege and Peano pro- and has given rise to both the extensive work vided additional foundation for the modern on human problem solving and the large collec- form of predicate calculus. Work by Davis, Put- tion of knowledge-based systems. A third nam, and Robinson in the twentieth century approach, loosely rooted in biology, takes the provided the final steps in mechanizing deduc- view that the key to reasoning is the architec- tion sufficiently to enable the first automated ture of the machinery that accomplishes it; theorem provers. The modern offspring of this hence, reasoning is a characteristic stimulus- line of intellectual development include the response behavior that emerges from parallel many efforts that use first-order logic as a rep- interconnection of a large collection of very resentation and some variety of deduction as simple processors. Researchers working on sev- the reasoning engine, as well as the large body eral varieties of connectionism are descendants of work with the explicit agenda of making of this line of work; work on artificial life also logical reasoning computational, exemplified has roots in the biologically inspired view. A by Prolog. fourth approach, derived from probability the- Note we have here the underlying premise ory, adds to logic the notion of uncertainty, that reasoning intelligently means reasoning yielding a view in which reasoning intelligently logically; anything else is a mistake or an aber- means obeying the axioms of probability theo- ration. Allied with this is the belief that logical- ry. A fifth view, from economics, adds the fur- ly, in turn, means first-order logic, typically ther ingredients of values and preferences, lead- sound deduction (although other models have ing to a view of intelligent reasoning defined by of course been explored). By simple transitivi- adherence to the tenets of utility theory. ty, these two collapse into one key part of the Briefly exploring the historical development view of intelligent reasoning underlying logic: of the first two of these views will illustrate the Reasoning intelligently means reasoning in the different conceptions they have of the funda- fashion defined by first-order logic. A second mental nature of intelligent reasoning and will important part of the view is the allied belief demonstrate the deep-seated differences in that intelligent reasoning is a process that can mind set that arise—even within our own be captured in a formal description, particular- field—as a consequence. ly a formal description that is both precise and The Logical View: Reasoning as Formal concise. Calculation Consider first the tradition The Psychological View: Reasoning as that uses mathematical logic as a view of intel- Human Behavior But very different views ligent reasoning. This view has its historical of the nature of intelligent reasoning are also origins in Aristotle’s efforts to accumulate and possible. One distinctly different view is catalog the syllogisms, in an attempt to deter- embedded in the part of AI influenced by the mine what should be taken as a convincing psychological tradition. That tradition, rooted argument. (Note that even at the outset, there in the work of Hebb, Bruner, Miller, and is a hint of the idea that the desired form of Newell and Simon, broke through the stimu- reasoning might be describable in a set of for- lus-response view demanded by behaviorism mal rules.) The line continues with Descartes, and suggested instead that human problem- whose analytic geometry showed that Euclid’s solving behavior could usefully be viewed in work, apparently concerned with the stuff of terms of goals, plans, and other complex men- pure thought (lines of zero width, perfect cir- tal structures. Modern manifestations include

SPRING 1998 93 Presidential Address

work on SOAR (Rosenbloom, Laird, and Newell laborative work. Evolutions like this in our 1993) as a general mechanism for producing concept of intelligence have as corollaries a intelligent reasoning and knowledge-based corresponding evolution in our beliefs about systems as a means of capturing human expert where sources of power are to be found. One of reasoning. the things I take Minsky to be arguing in his Where the logicist tradition takes intelligent society of mind theory is that power is going to reasoning to be a form of calculation, typically arise not from the individual components and deduction in first-order logic, the tradition their (individual) capabilities, but from the based in psychology takes as the defining char- principles of organization—how you put acteristic of intelligent reasoning that it is a things (even relatively simple things) together particular variety of human behavior. In the in ways that will cause their interaction to pro- logicist view, the object of interest is thus a duce intelligence. This leads to the view of construct definable in formal terms via math- intelligence as an emergent phenomenon— ematics, while for those influenced by the psy- something that arises (often in a nonobvious chological tradition, it is an empirical phe- fashion) from the interaction of individual nomenon from the natural world. behaviors. If this is so, we face yet another There are thus two very different assump- challenge: If intelligence arises in unexpected tions here about the essential nature of the ways from aggregations, then how will we ever fundamental phenomenon to be captured. engineer intelligent behavior, that is, purpose- One of them makes AI a part of mathematics; fully create any particular variety of it? the other makes it a part of natural science. Consider then the wide variety of views we A second contrast arises in considering the in AI have taken of intelligent reasoning: logi- character of the answers each seeks. The logical and psychological, statistical and econom- cist view has traditionally sought compact and ic, individual and collaborative. The issue here precise characterizations of intelligence, look- is not one of selecting one of these over anoth- ing for the kind of characterizations encoun- er (although we all may have our individual tered in mathematics (and at times in physics). reasons for doing so). The issue is instead the The psychological tradition by contrast sug- significance of acknowledging and being gests that intelligence is not only a natural aware of the different conceptions that are phenomenon, it is an inherently complex nat- being explored and the fundamentally differ- ural phenomenon: as human anatomy and ent assumptions they make. AI has been and physiology are inherently complex systems will continue to be all these things; it can resulting from a long process of evolution, so embrace all of them simultaneously without perhaps is intelligence. As such, intelligence fear of contradiction. may be a large and fundamentally ad hoc col- AI: Exploring the Design Space of Intel- lection of mechanisms and phenomena, one ligences. The temptation remains, of course, for which complete and concise descriptions to try to unify them. I believe this can in fact may not be possible. be done, using a view I first heard articulated The point here is that there are a number of by Aaron Sloman (1994), who suggested con- different views of what intelligent reasoning ceiving of AI as the exploration of the design is, even within AI, and it matters which view space of intelligences. you take because it shapes almost everything, I believe this is a useful view of what we’re from research methodology to your notion of about for several reasons: First, it’s more gener- success. al than the usual conjunction that defines us The Societal View: Reasoning as Emer- as a field interested in both human intelligence gent Behavior AI’s view of intelligent rea- and machine intelligence. Second, the plur- soning has varied in another dimension as al—intelligences—emphasizes the multiple well. We started out with the straightforward, possibilities of what intelligence is (or are, as introspection-driven view that intelligence my title suggests). Finally, conceiving of it in resided in, and resulted from, an individual terms of a design space suggests exploring mind. After all, there seems at first glance to be broadly and deeply, thinking about what kinds only one mind inside each of us. of intelligences there are, for there may be But this, too, has evolved over time, as AI many. has considered how intelligent reasoning can This view also helps address the at-times arise from groups of (more or less) intelligent debated issue of the character of our field: Are entities, ranging from the simple units that we science or engineering, analytic or synthet- make up connectionist networks, to the more ic, empirical or theoretical? The answer of complex units in Minsky’s (1986) society of course is, “yes.” mind, to the intelligent agents involved in col- Different niches of our field have different

94 AI MAGAZINE Presidential Address characters. Where we are concerned with many theories are correct. I suggest you attend human intelligence, our work is likely to be not to the details of each but to the overall more in the spirit of scientific, analytical, and character of each and what it may tell us about empirical undertakings. Where the concern is how the mind might have arisen. more one of machine intelligence, the work Presumably the mind evolved and should as will be more engineering, synthetic, and theo- a consequence have some of the hallmarks of retical. But the space is roughly continuous, it anything produced by that process. Let’s set is large, and all these have their place. the stage then by asking what’s known about the nature of evolution, the process that was presumably in charge of, and at the root of, all Why Is Intelligence? this. Next I’d like to turn to the question, “Why is intelligence?” That is, can we learn from an The Nature of Evolution explicitly evolutionary view? Is there, or could The first thing to remember about evolution is there be, a paleocognitive science? If so, what that it is engaging in a pastime that’s quite would it tell us? familiar to us: blind search. This is sometimes We had best begin by recognizing the diffi- forgotten when we see the remarkable culty of such an undertaking. It’s challenging results—apparently elegant and complex sys- for several reasons: First, few of the relevant tems—that come from a few million years’ things fossilize. I’ve checked the ancient bits worth of search. The issue is put well in the of amber, and sadly, there are no Jurassic title of one article—“The Good Enough Calculi … can we ontologies to be found embedded there; there of Evolving Control Systems: Evolution Is Not are no Paleolithic rule-based systems still Engineering” (Partridge 1982). The article goes learn from available for study; and although there is spec- on to contrast evolution and engineering an explicitly ulation that the cave paintings at Lascaux problem solving: In engineering, we have a were the earliest implementation of JAVA, this defined problem in the form of design require- evolutionary is, of course, speculation. ments and a library of design elements avail- view? The examples may be whimsical, but the able for the solution. But “biology provides no point is real—few of the elements of our intel- definition of a problem until it has been Is there, lectual life from prehistoric times are preserved revealed by the advantage of a solution. With- or could and available for study. There are even those out a predefined problem, there is no prerequi- there be, a who suggest the entire undertaking is doomed site domain, range, form for a solution, or from the start. Richard Lewontin (1990), who coordinates for its evaluation, except that it paleocognitive has written extensively on evolution, suggests provides a statistically improved survival func- science? that “if it were our purpose in this chapter to tion. This filter selects ‘good enough’ new solu- say what is actually known about the evolutions and thereby identifies solved problems” If so, what tion of human cognition, we would stop at the (p. R173). would it end of this sentence” (p. 229). Consider in particular the claim that “biolo- Luckily, he goes on: “That is not to say that gy provides no definition of a problem until it tell us? a good deal has not been written on the sub- has been revealed by the advantage of a solu- ject. Indeed whole books have been devoted to tion.” The warning here is to be wary of inter- discussions of the evolution of human cogni- preting the results of evolution as nature’s clev- tion and its social manifestations, but these erness in solving a problem. It had no problem works are nothing more than a mixture of pure to solve; it was just trying out variations. speculation and inventive stories. Some of The consequences of blind search are famil- these stories might even be true, but we do not iar to us; so, in some ways what follows seems know, nor is it clear…how we would go about obvious, but the consequences are neverthe- finding out” (p. 229). Hence, we had better be less worth attending to.2 modest in our expectations and claims. One consequence of random search is that A second difficulty lies in the data that are evolution wanders about, populating niches available. Most attempts to date phenomena wherever it finds them in the design space and are good only to something like a factor of two the environment. Evolution is not a process of or four. The taming of fire, for example, prob- ascent or descent; it’s a branching search space ably occurred around 100,000 years ago, but it being explored in parallel. might have been 200,000 or even 400,000. A second consequence is that nature is Then there is the profusion of theories about sometimes a lousy engineer. There are, for why intelligence arose (more on those in a example, futile metabolic cycles in our moment). Luckily for our purposes, we don’t cells—apparently circular chemical reactions actually have to know which, if any, of these that go back and forth producing and unpro-

SPRING 1998 95 Presidential Address

ducing the same molecules and depleting ener- in the blood near zero, but it cannot increase gy stores for no apparent purpose (Katz 1985). blood-oxygen saturation past the blood’s nor-

Third, despite the size of the design space, mal limits. As a result, the CO2 level can stay blind search sometimes doubles back on itself, abnormally low past the time that oxygen lev- and evolution rediscovers the same mecha- els have significantly decreased, and the diver nisms. One widely cited example is the eye of will feel no need to breathe even though the mammal and the eye of the octopus. They blood-oxygen levels are low enough to lead to are quite similar but for one quite striking fact: blackout. The human eye is backward compared with the Fifth, evolution sometimes proceeds by octopus (Katz 1985). In the mammalian eye, functional conversion, that is, the adoption of the photoreceptors are in the retinal layer near- an organ or system serving one purpose to est the rear of the eye; as a consequence, light serve another. The premier example here is has to go through the retinal “back plane” bird wings: The structures were originally before it encounters the photoreceptors. developed for thermal regulation (as they are A second striking example arises in the evo- still used in insects) and, at some point, were lution of lungs in mammals and birds. Both coopted for use in flight. appear to have arisen from the swim bladders Finally, evolution is conservative: It adds that fish use to control buoyancy, but birds’ new layers of solutions to old ones rather than lungs are unidirectionally ventilated, unlike the redesigning. This in part accounts for and pro- tidal, bidirectional flow in other vertebrates. (As duces vestigal organs and systems, and the a consequence, avian lungs are much more effi- result is not necessarily pretty from an engi- cient than ours: Himalayan geese have been neering viewpoint. As one author put it, “The observed not only to fly over human climbers human brain is wall-to-wall add-ons, a maze of struggling with their oxygen tanks to reach the dinguses and gizmos patched into the original top of Mt. Everest but to honk as they do so pattern of a primitive fish brain. No wonder it (Encyclopedia Brittannica 1994–1997); presum- isn’t easy to understand how it works” (Bicker- ably this is nature’s way of reminding us of our ton 1995, p. 36). place in the scheme of things.) Evolution then is doing random search, and The differences in end results suggest the the process is manifest in the product. As one different paths that were taken to these results, author put it, yet the remaining similarities in eyes and lungs In the natural realm, organisms are not show that evolution can rediscover the same built by engineers who, with an overall basic mechanisms despite its random search. plan in mind, use only the most appropri- Fourth, there are numerous examples of ate materials, the most effective design, how nature is a satisficer, not an optimizer. For and the most reliable construction tech- instance, one of the reasons cuckoos can get niques. Instead, organisms are patchworks away with dropping off their eggs in the nests containing appendixes, uvulas, earlobes, of other birds is that birds have only a very dewclaws, adenoids, warts, eyebrows, crude algorithm for recognizing their eggs and underarm hair, wisdom teeth, and toe- their chicks (Calvin 1991). The algorithm is nails. They are a meld of ancestral parts good enough, most of the time, but the cuckoo integrated step by step during their devel- takes advantage of its only adequate (manifest- opment through a set of tried and true ly nonoptimal) performance. ontogenetic mechanisms. These mecha- The control of human respiration provides nisms ensure matching between disparate another example. Respiration is, for the most elements such as nerves and muscles, but

part, controlled by the level of CO2 in the they have no overall vision. Natural onto- blood. There appear to be a variety of reasons genies and natural phylogenies are not

for this (for example, controlling CO2 is one limited by principles of parsimony, and way to control pH levels in the blood), but it’s they have no teleology. Possible organisms still only an adequate system. Its limits are well can be overdetermined, unnecessarily known to mountain climbers and divers. complex, or inefficiently designed (Katz Mountain climbers know that they have to be 1985, p. 28). conscious of the need to breathe at altitude The important point here for our purposes because the thin air leaves CO2 levels in the is that what’s manifestly true of our anatomy blood low, eliminating the normal physiologi- may also be true of our cognitive architec- cal cues to breathe, even through blood-oxy- ture. Natural intelligence is unlikely to have gen levels are also low. an overall vision and unlikely to be limited Divers need to understand that hyperventi- by principles of parsimony; like our bodies, lation is dangerous: It can drive the CO2 level it is likely to be overdetermined, unnecessar-

96 AI MAGAZINE Presidential Address

crancial capacity 1600 H. neanderthalensis H. sapiens 1200 H. erectus 800 H. habilus 400 A. africanus

- 4M -3M -2M -1M 0

Time (millions of years)

A. = Australopithecus H.= Homo

Figure 1. The Fossil Record (derived from data in Donald [1991], Eccles [1989], Mithen [1996], and Hyland [1993]). Note: The averages shown in this chart do not make evident the apparent discontinuities in the size increases. As Mitherton (1996) dis- cusses, there were apparently two bursts of brain enlargement, one about two million years ago, at the transition from A. africanus to H. habilis, and another about 500,000 years ago, with the transition from H. erectus. And yes, the brains of H. neanderthalensis were on average larger than those of modern man, though so, too, was its body. Finally, note that the data in this ﬁeld change more rapidly than one might expect: This chart was accurate when drawn in August 1996, but by December 1996 new evidence (Swisher et al. 1996) was reported suggesting that H. erectus did not in fact die off 250,000 years ago and may have lived contemporaneously with H. sapiens and the Neanderthals.

ily complex, and inefficiently designed. change over a short period of time. Simply put, In the face of that, searching for the mini- our brains got very big very fast. malism and elegance beloved by engineers This is interesting in part because brains are may be a diversion, for it simply may not be metabolically very expensive. In the adult, there. Somewhat more crudely put: The about 20 percent of our metabolism goes into human mind is a 400,000-year-old legacy maintaining our brains; in children, the brain application…and you expected to find struc- consumes about 50 percent of metabolic output tured programming? (Bickerton 1995). This makes the question all All that in turn gives us all the more reason the more pressing: Considering how expensive to explore deeply into the design space of large brains are, why do we have them? Why is intelligence, for the human solution, and its intelligence? What beneﬁt arose from it? sources of power, may be extraordinarily A second clear piece of evidence, this time quirky. from current studies of the brain, is lateralization: The standard examples are language The Available Evidence (found in the left hemisphere in approximate- If we can’t rely on the fossil record for pre- ly 93 percent of us) and the rapid sequencing served bits of cognition, can it supply other of voluntary muscles for things such as throw- useful information? One observation from the ing (found on the left in 89 percent) (Calvin record of particular relevance is the striking 1983). This is striking in part because the increase in what’s called the encephalization human brain has very few anatomical asym- quotient—the ratio of brain size to body size. metries; the observed asymmetries are almost Fossil records give clear evidence that the entirely functional (Eccles 1989). It is also encephalization quotient of human ancestors striking because the asymmetry arose with the increased by a factor of three to four over hominids (Homo and our ancestors) and about four million years (Donald 1991). In appears unique to them; the brains of our clos- evolutionary terms, this is an enormous est living relatives—apes and monkeys—are

SPRING 1998 97 Presidential Address

1600 H. neanderthalensis H. sapiens 1200 H. erectus 800 H. habilus 400 A. africanus

-4M -3M

-200K -100K AAAI-96 fire agriculture speech adv. tools Lascaux

Figure 2. A More Detailed Look at the Fossil Record.

symmetrical both anatomically and function- Theories of the Origin of Intelligence ally (Eccles 1989). A variety of theories have been suggested. The interesting question here of course is why, in a world of symmetry, is the human Early Man, the Primal Tool Maker One brain lateralized, even in part? theory is wrapped up in the notion that man One useful way to set the stage for the vari- is a tool maker. The construction of increasing- ous suggested answers is to consider the ly elaborate tools both gave early man a sur- sequence of events that lead to Homo (H.) sapi- vival advantage and produced evolutionary ens. Figure 1 gives an overview of the last four pressure for yet more elaborate tools and the million years, indicating the evolutionary span brains to build them. Unfortunately, another of several of our immediate ancestors and their look at our time scale provides some disquiet- average cranial capacity. ing data. The earliest tools show up around 2.5 If we zoom in on the last 200,000 years, we million years ago and stay largely unchanged see a few additional events of note (figure 2). until about 300,000 years ago (Calvin 1991). Speech arrives quite recently, around 200,000 Yet during all that time our brains are growing to 400,000 years ago; fire doesn’t get tamed quickly. The tool theory thus seems unlikely. until around 100,000 years ago, which is when Early Man and the Killer Frisbee A sec- more advanced tools also begin to appear. The ond theory (Calvin 1991, 1983) is centered on conversion from hunter-gatherers to a settled hunting methods and involves passing a society dependent on the use of agriculture device that is sometimes whimsically referred happens roughly 10,000 to 15,000 years ago, to as the killer frisbee (figure 3). It’s one of the about the same time as the cave paintings at earliest tools and is more properly called a Lascaux. hand ax because it was believed to be a hand- One question to ask about all this is, What held ax. The curious thing about it is that if changed between four million years ago and you look closely, you’ll see that all its edges are now? Four million years ago, there was (pre- sharp—not a very good idea for something sumably) nothing we would recognize as designed to be held in the hand. human-level intelligence; now there is. What One researcher built replicas of these and changed in between? discovered that if thrown like a discus, it flies

98 AI MAGAZINE Presidential Address

Figure 3. An Early Tool: Top and Side Views. Reproduced with permission from Calvin (1991). like a frisbee at first but soon turns on edge and is, those with differing intrinsic average fre- lands with its sharp edge embedded in the quencies and that are individually unreliable earth. Now add to this the fact that many of on a cycle-to-cycle basis). With this arrange- these artifacts have been found in the mud ment, the standard deviation of cycle length near ancient waterholes. This led to the theory between successive firings is proportional to that the artifacts were thrown by our ancestors 1 N at herds of animals gathered at waterholes, so quadrupling the number of elements cuts with the intent of wounding one of them or the standard deviation in half. This might knocking it down. account for the ability of our brains to control But why should throwing things be interest- muscle action to within fractions of a millisec- ing—because throwing accurately requires pre- ond, when individual neurons are an order of cise time control of motor neurons. For exam- magnitude less precise. ple, if you want to throw accurately at a target The theory then is that our brains grew larg- the size of a rabbit that’s 30 feet away (figure er because more neurons produced an increase 4), the motor-control problem is substantial: in throwing accuracy (or an increase in projec- the time window for release of the projectile is tile speed with no reduction in accuracy), and less than 1 microsecond. But individual neu- that in turn offered a major selective advan- rons are not in general that accurate temporal- tage: the ability to take advantage of a food ly. How do we manage? source—small mammals—that was previously One way to get the needed accuracy is to untapped by hominids. A new food source in recruit populations of neurons and synchro- turn means a new ecological niche ripe for nize them: Enright (1980) shows how precise inhabiting. The advantage resulting from even timing can be produced from mutual coupling a limited ability to make use of a new source of of heterogeneous, inaccurate oscillators (that food also provides a stronger and more imme-

SPRING 1998 99 Presidential Address

Figure 4. Throwing Stones. A. At 4 meters, the launch window is 11 microseconds; at 8 meters, it narrows to 1.4 microseconds. Reproduced with permission from Calvin (1991).

diate selective pressure than is likely to have to be good enough at hunting to accumulate arisen from other advantages of a slightly extra food beyond the day-to-day needs (hence enlarged brain (for example, some limited pro- the related utility of being able to throw accu- tolanguage ability). rately), and then it would have had to develop The theory has a number of appealing corol- both the foresight to put aside some of that for laries. It suggests one source of lateralization the winter and the “technology” for doing so. because throwing is fundamentally asymmet- There is, of course, a stiff Darwinian penalty ric: One-armed throwing is far more accurate for failure to be that smart. and effective than two armed for any reason- Early Man, the Primal Frugivore A able-sized projectile (imagine baseball pitchers fourth theory suggests that the crucial element or outfielders propelling the ball overhead was the evolution of early man into a frugivore, with both arms). As a result, only the neurons or fruit eater. Why should this matter— on one side of the brain need be specialized for because you need to be smart to be a frugivore. the operation (for why this turns out, in nearly Fruit comes in relatively small pieces, so you 90 percent of us, to be the left side of the brain, need to collect a lot of it, and it must be col- see Calvin [1983]).3 That lateralization, which lected within a relatively narrow time window. more generally involves precise sequential As a consequence, frugivores need good spatial muscle control, may in turn have been a key maps of their environments (so they know predecessor to language, which also requires where the sources of fruit are) and good tem- fast and accurate control of musculature. poral maps (so they know when to show up). Thus, the brain may have gotten larger to Perhaps this need for good spatial and tempo- allow us to hunt better. The interesting punch- ral maps was a force for the evolution of larger line for our purposes is that thinking may be brains. an extra use of all those neurons that evolved Early Man, the Primal Psychologist Yet for another purpose. another theory suggests that our primary use Early Man and the Killer Climate A of intelligence is not for making tools, hunt- third theory suggests that climate plays a cen- ing, or surviving the winter; it’s to get along tral role (Calvin 1991). The last few hundred with one another (Humphrey 1976; also see thousand years of our history have been Byrne and Whiten [1988]). This theory is marked by a series of ice ages. A being used to sometimes called Machiavellian intelligence. In surviving in a temperate climate would face a this view, the primary function of intelligence considerable collection of challenges as the is the maintenance of social relationships. weather worsened and winters arrived. In The evidence for this comes from several order to survive the winter, it would have had sources, among them the behavior of monkey

100 AI MAGAZINE Presidential Address troops that have been studied extensively. tinction between animal intelligence and They are seen to spend a good proportion of human intelligence. Animal intelligence has a their time servicing and maintaining their here and now character: With animal calls, for relationships within their groups, tending to example, there is an immediate link from the issues of rank and hierarchy and what appear perception to the mind state to the action. If a to be allegiances. monkey sees a leopard, a certain mind state A second source of evidence comes from a ensues, and a certain behavior (giving the study (Dunbar 1992) that plotted group size appropriate call) immediately follows.4 against neocortex ratio (ratio of neocortex size Human thought, by contrast, has an unlim- to the size of the rest of the brain) for a variety ited spatiotemporal reference, by virtue of sev- of animals: a nearly linear relationship eral important disconnections. Human thought emerged. Perhaps this held true for early man involves the ability to imagine, the ability to as well: As early group size grew, along with the think about something in the absence of per- advantages of larger groups came increasing ceptual input, and the ability to imagine with- demands to be able to understand, predict, and out reacting. perhaps even control the behavior of others. In human thought we have the ability, the We saw earlier that prediction was a key com- luxury, of “re-presentation.” The pun is inten- ponent of intelligent behavior; what more tional and probably educational: Representa- complex, fascinating, and useful thing could tions allow us to re-present things to ourselves there be to predict than the behavior of then in the absence of the thing, so that we can … another other humans? think about it, not just react to it. theory Early Man, the Primal Linguist Finally, Enormous things change when we have Bickerton (1995) has suggested that language both thought and language. Thought and its suggests that was the crucial driving force behind the evolu- useful disconnection from immediate stimuli our primary tion of our brains. He starts with the interest- and immediate action is clearly a great ing observation that if we look back at the his- boon—it’s the origin of our ability to have our use of torical time line, we notice that although brain hypotheses die in our stead. But what about intelligence size grows roughly steadily for about three mil- language? For our purposes, the interesting lion years, progress in the development of thing about language is that it makes knowl- is not for modern culture was not nearly so gradual. In edge immortal and makes society, not the indi- making fact, “instead of a steady ascent . . . we find, for vidual, the accumulator and repository of tools, 95% of that period, a monotonous, almost flat knowledge. No longer is an individual’s knowl- line” (Bickerton 1995, p. 47). Almost nothing edge limited to what can be experienced and hunting, or happens. It is well after the appearance of H. learned in a lifetime. Language not only allows surviving sapiens, and well after the leveling off of brain us to think, it allows us to share and accumu- size, that we see the appearance of language late the fruits of that thought. the winter; and all the other elements of what we have But what then caused our brains to grow it’s to get come to call civilization. over the three million or so years during which Bickerton calls these the two most shocking neither language nor thought (as we know along with facts of human evolution: (1) our ancestors them) was present? What was the evolutionary one another. stagnated so long despite their ever-growing pressure? The theory suggests that the life of a brains and (2) human culture grew exponen- successful hunter-gatherer is fact rich and prac- tially only after the brain had ceased to grow. tice rich. In order to survive as a hunter-gath- It appears that we showed our most obvious erer, you need to know a lot of facts about your evidence of intelligence only after our brains world and need to know a fair number of skills. stopped growing. This then is the hypothesized source of pres- What was it that happened to produce that sure: the increasing accumulation of survival- evidence? He suggests that the crucial event relevant information communicated through was some sort of reorganization within the a form of protolanguage. Early man needed to brain, a reorganization that happened well store “the vast amount of lore . . . in the collec- after size stopped increasing. That reorganiza- tive memories of traditional societies: the uses tion made possible two essential things: first, a of herbs, the habits of animals, aphorisms generative syntax, that is, a true language, and about human behavior, detailed knowledge of second, thought, that is, the ability to think the spatial environment, anecdotes, old wives’ about something (like a leopard) without hav- tales, legends and myths” (Bickerton 1995, p. ing to experience the thing perceptually, and 63).5 equally important, without having to react to Where does this collection of theories (fig- it in the way one would on meeting one. ure 5) leave us? One obvious caution is that This leads to what appears to be a crucial dis- they are unlikely to be either independent or

SPRING 1998 101 Presidential Address

it’s time for AI to learn about the birds and the bees. What do animals know, and (how) do they think? Clever Hans and Clever Hands Early man, the primal tool maker Early man and the killer frisbee Before we get too far into this, it would we wise Early man and the killer climate to consider a couple of cautionary tales to Early man, the primal frugivore ensure the appropriate degree of skepticism Early man, the primal psychologist about this difficult subject. The classic caution- Early man, the protolinguist ary tale concerns a horse named Clever Hans, raised in Germany around 1900, that gave every appearance of being able to do arith- metic, tapping out his answers with his feet (Boakes 1984) (figure 6). He was able to give the correct answers even without his trainer in the room and became a focus of a considerable Figure 5. Theories of the Evolution of Intelligence. amount of attention and something of a celebrity. In the end, it turned out that Hans was not mutually exclusive. They may be mutually mathematically gifted; his gift was perceptual. supportive and all true to some extent, with The key clue came when he was asked ques- each of them contributing some amount of the tions to which no one in the room knew the evolutionary pressure toward larger brains and answer; in that case, neither did he. Hans had intelligence. been attending carefully to his audience and A second point to note is that human intel- reacting to the slight changes in posture that ligence is a natural phenomenon, born of evo- occurred when he had given the correct num- lution, and as suggested earlier, the end prod- ber of taps.6 uct likely shows evidence of the process that The clever hands belong to a chimpanzee created it. Intelligence is likely to be a layered, named Washoe who had been trained in multifaceted, and probably messy collection of American Sign Language (Gardner et al. 1989). phenomena, much like the other products of One day Washoe, seeing a swan in a pond, evolution. gave the sign for water and then bird. This It also may be rather indirect. Here’s Lewon- seemed quite remarkable, as Washoe seemed to tin (1990) again: “There may have been no be forming compound nouns—water bird— direct natural selection for cognitive ability at that he had not previously known (Mithen all. Human cognition may have developed as 1996). But perhaps he had seen the pond and the purely epiphenomenal consequence of the given the sign for water, then noticed the swan major increase in brain size, which, in turn, and given the sign for bird. Had he done so in may have been selected for quite other rea- the opposite order—bird water—little excite- sons” (p. 244), for example, any of the reasons ment would have followed. in figure 5. The standard caution from both of these This, too, suggests a certain amount of cau- tales is always to consider the simpler explana- tion in our approach to understanding intelli- tion—trainer effects, wishful interpretation of gence, at least of the human variety: The data, and so on—before being willing to con- human mind is not only a 400,000-year-old sider that animals are indeed capable of legacy application, it may have been written thought. for another purpose and adopted for current usage only after the fact. In light of that, we Narrow Intelligence: Birds and Bees should not be too surprised if we fail to find Given that, we can proceed to explore some of elegance and simplicity in the workings of the varieties of animal intelligence that do intelligence. exist. Several types of rather narrowly defined intelligence are supported by strong evidence. Inhuman Problem Solving Among the birds and the bees, for example, bees are well known to “dance” for their hive As we explore the design space of intelligences, mates to indicate the direction of food sources it’s interesting to consider some of the other they have found. Some birds have a remark- varieties of intelligence that are out there, par- able ability to construct a spatial map. The ticularly the animal sort. With that, let me Clark’s nutcracker, as one example, stores away turn to the third part of my article, in which on the order of 30,000 seeds in 6,000 sites over

102 AI MAGAZINE Presidential Address

Figure 6. Clever Hans, the Mathematical Horse. His owner and trainer is rightmost of the group at the rear. Reproduced with permission from Boakes (1984). the course of the spring and summer and is They also seem to have a vocabulary with able to find about half of those during the win- semantic content—different calls that corre- ter (Balda and Kamil 1992). This is a narrowly spond to the notion of leopard, eagle, and restricted kind of intelligence but, at 6000 python, the three main monkey predators. That locations, nonetheless impressive. the calls are truly referential is suggested by the facts that they are given only when appropri- Broader Intelligence: Primates ate, they are learned by trial and error by the Broader forms of intelligence are displayed by young monkeys, and the troop takes appropri- some primates. One particular variety—the ate action on hearing one of the calls. Hearing vervet monkey—has been studied widely in the eagle call, for instance, all the troop mem- the wild and has displayed a range of intelli- bers will look up, searching for the eagle, then gent-seeming behaviors (Cheney and Seyfarth take cover in the bushes. Note that we have 1990). One of the important elements in the referred to this as a vocabulary, not a language, life of a monkey group is status—your place in because it appears that there is no syntax per- the dominance hierarchy. Vervet monkeys mitting the construction of phrases. give every sign of understanding and being able to reason using relations such as higher- Lies—Do Monkeys Cry Leopard? status-than and lower-status-than. They can, There is also some anecdotal evidence that the for example, do simple transitive inference to monkeys lie to one another. They have been establish the place of others in the hierarchy: If observed to lie by omission when it concerns A can beat up B, and B can beat up C, there’s food: When happening on a modest-sized no need for A and C to fight; the result can be store of food, a monkey may fail to give the inferred (allowing our hypotheses to get bat- standard call ordinarily given when finding tered in our stead). food. Instead, the lone monkey may simply The monkeys also appear capable of classify- consume it. ing relationships as same or different, under- A more intriguing form of misrepresenta- standing, for example, that mother-of is a dif- tion has been observed to occur when two ferent relation from sibling-of. This can matter neighboring monkey troops get into battles because if you fight with Junior, you had better over territory. Some of these battles have end- avoid mother-of(Junior) (who might be tempt- ed when one of the monkeys gives the leopard ed to retaliate), but sibling-of(Junior) presents call—all the combatants scatter, climbing into no such threat. trees to escape the predator, but there is in fact

SPRING 1998 103 Presidential Address

no leopard to be found. The monkeys may be Alex: Color. lying to one another as a way of breaking up Dr. Pepperberg: Good parrot. You’re the fight (Cheney and Seyfarth 1991).7 right, different color. Alright, now look, tell me, what color Psittacine Intelligence: Bird Brains No bigger? What color bigger (same keys)? Longer Alex: Green. One final example of animal intelligence con- Dr. Pepperberg: Green; good boy. Green cerns an African Grey Parrot named Alex, who bigger. Good parrot. has been trained for quite a few years by Dr. Oh you’re a good boy today. Yes, three Irene Pepperberg of the University of Arizona. different questions on the same objects. Alex seems capable of grasping abstract con- Good parrot. cepts such as same, different, color, shape, and Dr. Pepperberg: What we’ve found out is numbers (Pepperberg 1991). that a bird with a brain that is so different A videotape of Alex in action (WNET 1995) from mammals and primates can perform is particularly compelling; even a transcript of at the same level as chimpanzees and dol- the conversation will give you a sense of what’s phins on all the tests that we’ve used and been accomplished. Pay particular attention to performs about at the level of a young, Alex’s ability to deal with, and reason about, say, kindergarten-age child. abstract concepts and relations. This is an interesting bit of animal intelli- Narrator: For 17 years, Alex and Dr. gence, in part because of the careful training Irene Pepperberg have been working on and testing that’s been done, suggesting that, the mental powers of parrots. Their efforts unlike Hans, Alex really does understand cer- at the University of Arizona have pro- tain concepts. This is all the more remarkable duced some remarkable results. given the significant differences between bird Dr. Pepperberg: What shape (holding up and mammalian brains: Parrot brains are quite a red square)? primitive by comparison, with a far smaller Alex: Corners. cerebral cortex. Dr. Pepperberg: Yeah, how many corners? Say the whole thing. Consequences Alex: Four…corners. These varieties of animal intelligence illustrate Dr. Pepperberg: That’s right, four cor- two important points: First, they illuminate for ners. Good birdie. us a number of other distinguishable points in Alex: Wanna nut. the design space of intelligences. The narrow Dr. Pepperberg: You can’t have another intelligences of birds and bees, clearly more nut. limited than our own, still offer impressive evi- OK, what shape? (holding up a green dence of understanding and reasoning about triangle). space. Primate intelligence provides evidence Alex: Three…corners. of symbolic reasoning that, although primi- Dr. Pepperberg: That’s right, three cor- tive, has some of the character of what seems ners; that’s a good boy. central to our own intelligence. Clearly distin- Now tell me, what color (holding the guishable from our own variety of intelligence, same green triangle)? yet impressive on their own terms, these phe- Alex: Green. nomena begin to suggest the depth and Dr. Pepperberg: Green, ok; here’s a nut. breadth of the natural intelligences that have OK, and what toy (holding up a toy evolved. truck)? Second, the fact that even some part of that Alex: Truck. intelligence appears similar to our own suggests Dr. Pepperberg: Truck; you’re a good boy. the continuity of the design space. Human OK, let’s see if we can do something intelligence may be distinct, but it does not sit more difficult alone and unapproachable in the space. There (holding two keys, one green plastic, is a large continuum of possibilities in that one red metal; the green is slightly larger). space; understanding some of our nearest Tell, me, how many? neighbors may help us understand our own Alex: Two. intelligence. Even admitting that there can be Dr. Pepperberg: You’re right, good parrot. near neighbors offers a useful perspective. Alex: Wanna nut. Dr. Pepperberg: Yes, you can have a nut. Primate Celebrities Alright, now look, tell me, what’s differ- I can’t leave the topic of animal intelligence ent (same keys)? without paying homage to one of the true

104 AI MAGAZINE Presidential Address unsung heroes of early AI research. Everyone in AI knows the monkey and bananas problem of course. But what’s shocking, truly shocking, is that so many of us (myself included) don’t know the real origins of this problem. Thus, for the generations of AI students (and faculty) who have struggled with the monkey and bananas problem without knowing its origins, I give you, the monkey (ﬁgure 7):8 This one is named Rana; he and several other chimps were the subjects in an experiment done by gestalt psychologist Wolfgang Kohler (1925) in 1918. Kohler was studying the intelligence of animals, with particular attention to the phenomenon of insight, and gave his subjects a number of problems to solve. Here’s Grande, another of the chimps, hard at work on the most famous of them (ﬁgure 8). Thus, there really was a monkey and a stalk of bananas, and it all happened back in 1918. Just to give you a feeling of how long ago that was, in 1918, Herb Simon had not yet won the Nobel Prize.

Searching Design Space Figure 7. Rana, Star of an Early AI Problem. In this last segment of the article, I’d like to Reproduced with permission from Kohler (1969). consider what parts of the design space of intelligence we might usefully explore more thoroughly. None of these are unpopulated; people are doing some forms of the work I’ll propose. My suggestion is that there’s plenty of room for others to join them and good reason to want to. Thinking Is Reliving One exploration is inspired by looking at alter- natives to the usual view that thinking is a form of internal verbalization. We also seem to be able to visualize internally and do some of our thinking visually; we seem to “see” things internally. As one common example, if I were to ask whether an adult elephant could fit through your bedroom door, you would most likely attempt to answer it by reference to some mental image of the doorway and an elephant. There is more than anecdotal evidence to support the proposition that mental imaging is closely related to perception; a variety of experimental and clinical data also support the notion. As one example, patients who had suf- fered a loss of their left visual ﬁeld as a consequence of a stroke showed an interesting form of mental imagery loss (Bisiach and Luzzatti 1978). These patients were asked to imagine themselves standing at the northern end of a town square that they knew well and asked to Figure 8. Grande Going for the Gold(en) Bananas. report the buildings that they could “see” in Reproduced with permission from Kohler (1925).

SPRING 1998 105 Presidential Address

Bloci A Block B Block B Block C

Figure 9. Are A and B the Same Object; Are B and C? Reprinted with permission from Shepard, R. N., and Metzler, J., Mental Rotation of Three-Dimensional Objects, Science 171:701–703, copyright 1971, American Association for the Advancement of Science.

their mental image when looking south. Inter- people seem to do a form of mental rotation estingly, they report what they would in fact on these images. The primary evidence for this be able to see out of the right half of their visu- is that response time is directly proportional to al field; that is, they report buildings to the the amount of rotation necessary to get the ﬁg- south and west but none to the east. ures in alignment. Even more remarkably, if they are then A second experiment in the same vein asked to imagine themselves on the south end involved mental folding (Shepard and Feng of the square looking north and asked to 1972). The task here is to decide whether the report on what they “see” in their mental two arrows will meet when each of the pieces image, they describe the buildings in what is of paper shown in figure 10 is folded into a now the right half of their visual ﬁeld (that is, cube. buildings to the north and east) and fail com- If you introspect as you do this task, I think pletely to report those on the west side of the you’ll find that you are recreating in your square, even though they had mentioned mind the sequence of actions you would take them only moments earlier. were you to pick up the paper and fold it by The process going on in using the mind’s hand. eye to “see” is thus remarkably similar in some What are we to make of these experiments? ways to what happens in using the anatomical I suggest two things: First, it may be time to eye to see. take seriously (once again) the notion of visual A second source of support for this view reasoning, that is, reasoning with diagrams as comes from the observation of like-modality things that we look at, whose visual nature is a interference. If I ask you to hold a visual image central part of the representation. in your mind while you try to detect either a Second is the suggestion that thinking is a visual or an auditory stimulus, the ability to form of reliving. The usual interpretation of detect the visual stimulus is degraded, but the data from the rotation and folding experi- detection of the auditory stimulus remains the ments is that we think visually. But consider same (Segal and Fusella 1970). some additional questions about the experi- A third source of evidence comes from ments: Why does it take time to do the rota- experiments done in the 1970s that explored tion, and why does the paper get mentally the nature of visual thinking. One well-known folded one piece at a time? In the rotation experiment involved showing subjects images experiment, why don’t our eyes simply look at that looked like figure 9 and then asking each block, compute a transform, then do the whether the two images were two views of the transformation in one step? I speculate that same structure, albeit rotated (Shepard and the reason is because our thought processes Metzler 1971). mimic real life: In solving the problem mental- One interesting result of this work was that ly, we’re re-acting out what we would experi-

106 AI MAGAZINE Presidential Address

Figure 10. Do the Arrows Meet When the Paper Is Folded into a Cube? Reprinted with permission from Shepard, R. N., and Feng, C., A Chronometric Study of Mental Paper Folding, Cognitive Psychology 3:228–243, copyright 1972, American Association for the Advancement of Science.

ence in the physical world. tions between the high-level and low-level That’s my second suggestion: Take seriously visual areas. Roughly speaking, the pathway the notion of thinking as a form of reliving our from the low-level area does data-driven pro- perceptual and motor experiences. That is, cessing, but the opposite pathway does model- thinking is not simply the decontextualized driven processing. One possible mechanism manipulation of abstract symbols (powerful for thinking as reliving is the dominant use of though that may be). Some significant part of the model-driven pathway to recreate the sorts our thinking may be the reuse, or simulation, of excitation patterns that would result from of our experiences in the environment. In this the actual experience. sense, vision and language are not simply One last speculation I’d like to make con- input-output channels into a mind where the cerns the power of visual reasoning and dia- thinking gets done; they are instead a signifi- grams. The suggestion here is that diagrams are cant part of the thought process itself. The powerful because they are, among other same may be true for our proprioreceptive and things, a form of what Johnson-Laird (1983) motor systems: In mentally folding the paper, called reasoning in the model. Roughly speaking, we simulate the experience as it would be were that’s the idea that some of the reasoning we we to have the paper in hand. do is not carried out in the formal abstract There is, by the way, a plausible evolution- terms of predicate calculus but is instead done ary rationale for this speculation that thinking by creating for ourselves a concrete miniworld is a form of reliving. It’s another instance of where we carry out mental actions and then functional conversion: Machinery developed examine the results. for perception turns out to be useful for think- One familiar example is the use of diagrams ing. Put differently, visual thinking is the when proving theorems in geometry. The offline use of our ability to see. We’re making intent is to get a proof of a perfectly general use of machinery that happened to be there for statement, yet it’s much easier to do with a another purpose, as has happened many times concrete, specific model, one that we can before in evolution.9 manipulate and then examine to read off the One further, ambitious speculation con- answers. cerns the neural machinery that might support Consider, for example, the hypothesis that such reliving: Ullman (1996) describes counter- any triangle can be shown to be the union of streams, a pair of complementary, intercon- two right triangles. nected pathways traveling in opposite direc- We might start by drawing a triangle (figure

SPRING 1998 107 Presidential Address

Figure 11. Triangles. A. A random triangle. B. A random triangle with a perpendicular.

11a). The proof of course calls for any triangle, draw a line that was about three inches long or but we find it much easier with a concrete one long enough to reach this other line. in front of us. That’s my last speculation: There may be We might then play with it a bit and even- ways to marry the concreteness of reasoning in tually hit on the idea of dropping a perpendic- the model with the power and generality of ular (figure 11b). abstraction. One early step in this direction is Wary of a proof from a single concrete discussed in Stahov, Davis, and Shrobe (1996), example, we might try a number of other tri- who discuss how a specific diagram can auto- angles and eventually come up with a formal matically be annotated with constraints that abstract proof. But it’s often a lot easier to have capture the appropriate general relationships a concrete example to work with, manipulate, among its parts, but there is plainly much and then examine the results of our manipula- more to be done. tions. What works for something as plainly visual as geometric theorems also seems to work for Summary things that are not nearly so visual, such as syl- With that, let me summarize. I want to suggest logisms. Consider these sentences describing a that intelligence are many things, and this is group of people (Johnson-Laird 1983, p. 5): true in several senses. Even within AI, and Some of the children have balloons. even with the subfield of inference, intelligence has been conceived of in a variety of Everyone with a balloon has a party hat. ways, including the logical perspective, which There’s evidence that when asked to deter- considers it a part of mathematical logic, and mine the logical consequences of these state- the psychological perspective, which considers ments, people imagine a concrete instance of a it an empirical phenomenon from the natural room and some finite collection of people, world. then examine it to determine the answer. One way to get a synthesis of these numer- The good news about any concrete example ous views is to conceive of AI as the study of is its concreteness; the bad news is its concrete- the design space of intelligences. I find this an ness, that is, its lack of generality—as many a inspiring way to conceive of our field, in part high school geometry student has discovered because of its inherent plurality of views and when he/she drew an insufficiently general dia- in part because it encourages us to explore gram. For diagrams in particular, the problem is broadly and deeply about all the full range of compelling: There’s no such thing as an approx- that space. imate diagram. Every line drawn has a precise We have also explored how human intelli- length, every angle a precise measure. The good gence is a natural artifact, the result of the news is that diagrams make everything explicit; process of evolution and its parallel, oppor- the bad news is that they can’t possibly avoid it. tunistic exploration of niches in the design Yet there are times when we’d like to marry space. As a result, it is likely to bear all the hall- the virtues of reasoning in a concrete diagram marks of any product of that process—it is like- with the generality that would allow us to ly to be layered, multifaceted, burdened with

108 AI MAGAZINE Presidential Address vestigal components, and rather messy. This is 5. Humphrey (1976) also touches on this idea. a second sense in which intelligence are many 6. Oskar Phungst, who determined the real nature of things—it is composed of the many elements Hans’s skill, was able to mimic it so successfully that that have been thrown together over evolu- he could pretend to be a mentalist, “reading the tionary timescales. mind” of someone thinking of a number: Pfungst Because of the origins of intelligence and its simply tapped until he saw the subtle changes in posture that were unconscious to the subject (Rosen- resulting character, AI as a discipline is likely to thal 1966). have more in common with biology and 7. For a countervailing view on the question of ani- anatomy than it does with mathematics or mal lying, see the chapter by Nicholas Mackintosh physics. We may be a long time collecting a in Khalfa (1994). wide variety of mechanisms rather than com- 8. A true-life anecdote concerning life in Cambridge: ing upon a few minimalist principles. When I went to a photographer to have this photo In exploring inhuman problem solving, we turned into a slide, the man behind the counter saw that animal intelligence seems to fit in (probably an underpaid psychology graduate stu- some narrowly constrained niches, particular- dent) looked at the old book with some interest, ly for the birds and bees, but for primates (and then laughed at the photo I wanted reproduced. I perhaps parrots), there are some broader vari- pretended to chide him, pointing out that the photo eties of animal intelligence. These other vari- was of a famous contributor to psychological theory. “A famous contributor to psychology?” he said. eties of intelligence illustrate a number of oth- “Then I know who it is.” “Who?” I asked. “Why er distinguishable points in the design space of that’s Noam Chimpsky, of course,” he replied. Yes, it intelligences, suggesting the depth and really happened, just that way. breadth of the natural intelligences that have 9. There has been signiﬁcant controversy concerning evolved and indicating the continuity of that the exact nature and status of mental images; see, for design space. example, Farah (1988), who reviews some of the Finally, I tried to suggest that there are some alternative theories as well as neuropsychological niches in the design space of intelligences that evidence for the reality of mental images. One of the are currently underexplored. There is, for exam- alternative theories suggests that subjects in experi- ple, the speculation that thinking is in part visu- ments of the mental-rotation sort are mentally sim- ulating their experience of seeing rather than actual- al, and if so, it might prove very useful to devel- ly using their visual pathways. For our purposes, op representations and reasoning mechanisms that’s almost as good: Although literal reuse of the that reason with diagrams (not just about them) visual hardware would be a compelling example of and that take seriously their visual nature. functional conversion, there is also something I speculated that thinking may be a form of intriguing in the notion that one part of the brain reliving, that re-acting out what we have expe- can realistically simulate the behavior of other parts. rienced is one powerful way to think about, and solve problems in, the world. And ﬁnally, References I suggested that it may prove useful to marry Balda, R. P., and Kamil, A. C. 1992. Long-Term Spa- the concreteness of reasoning in a model with tial Memory in Clark’s Nutcrackers. Animal Behav- the power that arises from reasoning abstractly iour 44:761–769. and generally. Bickerton, D. 1995. Language and Human Behavior. Seattle, Wash.: University of Washington Press. Bisiach, E., and Luzzatti, C. 1995. Unilateral Neglect Notes of Representational Space. Cortex 14:129–133. 1. Table 1 and some of the text following is from Boakes, R. 1984. From Darwin to Behaviorism. New Davis, Shrobe, and Szolovits (1993). York: Cambridge University Press. 2. For a detailed exploration of the consequences Byrne, R. W., and Whiten A., eds. 1988. Machiavel- and their potentially disquieting implications, see lian Intelligence: Social Expertise and the Evolution of Dennett (1995). Intellect in Monkeys, Apes, and Humans. Oxford, U.K.: Clarendon. 3. In brief, he suggests that it arises from the near- universal habit of women carrying babies in their Calvin, W. H. 1983. The Throwing Madonna. New left arms, probably because the maternal heartbeat is York: McGraw-Hill. easier for the baby to hear on that side. This kept Calvin, W. H. 1991. The Ascent of Mind: Ice Age Cli- their right arms free for throwing. Hence the first mates and the Evolution of Intelligence. New York: Ban- major league hunter-pitcher may have been what he tam. calls the throwing madonna (not incidentally, the title Cheney, D. L., and Seyfarth, R. M. 1991. Truth and of his book). Deception in Animal Communication. In Cognitive 4. That’s why the possibility of monkeys “lying” to Ethology, ed. C. Ristau, 127–151. Hillsdale, N.J.: one another (see later discussion) is so Lawrence Erlbaum. intriguing—precisely because it’s a break in the per- Cheney, D. L., and Seyfarth, R. M. 1990. How Mon- ception-action link. keys See the World: Inside the Mind of Another Species.

SPRING 1998 109 Presidential Address

Chicago: University of Chicago Press. Pepperberg, I. R. 1991. A Communicative Approach Davis, R.; Shrobe, H.; and Szolovits, P. 1993. What Is to Animal Cognition: A Study of Conceptual Abili- a Knowledge Representation? AI Magazine 14(1): ties of an African Grey Parrot. In Cognitive Ethology, 17–33. ed. C. Ristau, 153–186. Hillsdale, N.J.: Lawrence Erl- baum. Dennett, D. C. 1995. Darwin’s Dangerous Idea. New York: Simon and Schuster. Popper, K. R. 1985. Evolutionary Epistemology. In Popper Selections, ed. D. Miller, 78–86. Princeton, N.J.: Diamond, J. 1992. The Third Chimpanzee. New York: Princeton University Press. Harper Collins. Rosenbloom, P.; Laird, J.; Newell, A., eds. 1993. The Donald, M. 1991. Origins of the Modern Mind. Cam- SOAR Papers: Research on Integrated Intelligence. Cam- bridge, Mass.: Harvard University Press. bridge, Mass.: MIT Press. Dunbar, R. I. 1992. Neocortex Size as a Constraint on Rosenthal, R. 1966. Experimenter Effects in Behavioral Group Size in Primates. Journal of Human Evolution Research. New York: Meredith. 20:469–493. Segal S., and Fusella, V. 1970. Influence of Imaged Eccles, J. C. 1989. Evolution of the Brain. New York: Pictures and Sounds on Detection of Visual and Routledge. Auditory Signals. Journal of Experimental Psychology Encyclopedia Brittanica. 1994–1997. Respiration and 83:458–464. Respiratory Systems: General Feature of the Respira- Shepard, R. N., and Feng, C. 1972. A Chronometric tory Process: Dynamics of Vertebrate Respiratory Study of Mental Paper Folding. Cognitive Psychology Mechanisms: Birds. In Encyclopedia Brittanica. Elec- 3:228–243. tronic version. Available at www.eb.com:180/cgi- Shepard, R. N., and Metzler, J. 1971. Mental Rotation bin/g?DocF=macro/5005/50/11.html. of Three-Dimensional Objects. Science 171:701–703. Enright, J. T. 1980. Temporal Precision in Circadian Sloman, A. 1994. Explorations in Design Space. In Systems: A Reliable Neuronal Clock from Unreliable Proceedings of the Eleventh European Conference Components? Science 209:1542–1545. on Artificial Intelligence, ed. A. G. Cohn, 578–582. Farah M. 1980. Is Visual Imagery Really Visual? Over- New York: Wiley. looked Evidence from Neuropsychology. Psychologi- Stahovich T.; Davis, R.; and Shrobe, H. 1996. Gener- cal Review 95(3): 307–317. ating Multiple New Designs from a Sketch. In Pro- Gardner, R. A.; Gardner, B. T.; and vanCantfort, T. ceedings of the Thirteenth National Conference on 1989. Teaching Sign Language to Chimpanzees. New Artificial Intelligence, 1022–1029. Menlo Park, York: State University of New York Press. Calif.: American Association for Artificial Intelli- Humphrey, N . K. 1976. The Social Function of Intel- gence. lect. In Growing Points in Ethology, eds. P. Bateson and Swisher, C. C., III; Rink, W. J.; Antón, S. C.; Schwarcz, R. Hinde, 303–317. New York: Cambridge University H. P.; Curtis, G. H.; Suprijo, A.; and Widiasmoro. Lat- Press. est Homo Erectus of JAVA: Potential Contemporaneity Hyland, M. E. 1993. Size of Human Groups during with Homo Sapiens in Southeast Asia. Science the Paleolithic and the Evolutionary Significance of 274(5294): 1870. Increased Group Size. Behavioral and Brain Sciences Ullman, S. 1996. High-Level Vision. Cambridge, 16(4): 709–710. Mass.: MIT Press. Johnson-Laird, P. N. 1983. Mental Models. Cam- WNET. 1995. Parrots: Look Who’s Talking. New York: bridge, Mass.: Harvard University Press. Nature Video Library. Videocassette. Khalfa, J., ed. 1994. What Is Intelligence? New York: Cambridge University Press. Randall Davis is a professor of Katz, M. J. 1985. On the Wings of Angels: An Explo- computer science at the Massa- ration of the Limits of the Biological Enterprise. Har- chusetts Institute of Technology, vard Magazine 88(1): 25–32. where he works on model-based Kohler, W. 1969. The Task of Gestalt Psychology. reasoning systems for engineering Princeton, N.J.: Princeton University Press. design, problem solving, and trou- Kohler, W. 1925. The Mentality of Apes. New York: bleshooting. He has also been Harcourt, Brace. active in the area of intellectual property and software, serving on Lewontin, R. 1990. The Evolution of Cognition. In a number of government studies and as an adviser to An Invitation to Cognitive Science, Volume 3, eds. D. the court in legal cases. He received his undergradu- Osherson and E. Smith, 229–245. Cambridge, Mass.: ate degree from Dartmouth College and his Ph.D. MIT Press. from Stanford University. He serves on several edito- Minsky, M. 1986. The Society of Mind. New York: rial boards, including those for Artificial Intelligence Simon and Schuster. and AI in Engineering. In 1990, he was named a Mithen, S. 1996. The Prehistory of the Mind. London: founding fellow of the American Association for Thames and Hudson. Artificial Intelligence and served as president of the Partridge, L. D. 1982. The Good Enuf Calculi of association from 1995–1997. His e-mail address is Evolving Control Systems: Evolution Is Not Engi- [email protected]. neering. American Journal of Physiology 242:R173–R177.

110 AI MAGAZINE