FALL 1995 81

Book Review

Dave Chalmers, Daniel Defays, BobFrench, Gary McGraw, and MelanieMitchell). Although Hofstadter is notthe first author of all the articles,there is a coherent vision to thewhole book, reinforced by the pref-aces Hofstadter wrote for each article.

Two major ideas appear to underliemuch of Hofstadter’s work; althoughthese ideas are related, one is primari-ly a computational claim, and theother is primarily psychological. Thecomputational claim is that cogni-tion can be modeled using a parallelterraced scan, which builds a represen-tation by parallel investigation ofmany possibilities with different lev-els of commitment. The psychologi-cal claim is that cognition is a formof high-level perception that involves

■ Fluid Concepts and Creative Analo-gies, Douglas Hofstadter and theFluid Analogies Research Group,Basic Books, New York, 1995, 518pp., $30, ISBN 0-465-05154-5.

There is an apt analogy for theauthor of this book: DouglasHofstadter could be considered

the Carl Sagan of AI. As Hofstadterpoints out, analogies are fluid, mean-ing that the analogy between twoentities can be drawn differentlydepending on how these entities arerepresented. The analogy that isdrawn, in turn, can change the repre-sentation of the entities being com-pared. Thus, the analogy betweenHofstadter and Sagan can be seen aspositive: Both have explained impor-tant concepts in their fields to a wideaudience and transmitted the excite-ment of these ideas. Both haveinspired a number of people withintheir fields. Unfortunately, a morenegative analogy between Sagan andHofstadter is possible. Among someastronomers, Sagan’s work has notbeen taken seriously. In a similarfashion, Hofstadter’s work has hadsurprisingly little impact on the fieldof AI. Some of the reasons for thislack of acceptance in the field arebased on irrelevancies. For example,Hofstadter is the only researcher inAI with both a Pulitzer Prize and hisown news group on the Internet.However, there are some valid rea-sons for these reservations. This bookillustrates both the negative and thepositive analogies.

Fluid Concepts and Creative Analo-gies is a collection of articles, manyalready published, by Hofstadter andmembers of his Fluid AnalogiesResearch Group (here represented by

is. Chapter 5, “The COPYCAT Project: AModel of Mental Fluidity and Analo-gy Making,” is perhaps the bestdescription of an implementation ofHofstadter’s ideas.

The essential idea behind a parallelterraced scan is that initially, the rep-resentation of an entity consists of anumber of unrelated elements. Therepresentation is then elaborated bymany codelets, small pieces of codethat do specific things. Some codeletsrecognize potential structures but donot build them; others build struc-tures that have been identified. Thesestructures are often simple, such asthe combination of two elements.Because of the parallelism, compet-ing structures can be built; othercodelets can destroy a structure thathas become inconsistent with thehigher-level structure. The structurethat emerges is probabilistic becauseit results from interactions amongthe actions of many codelets that areexecuted probabilistically from acoderack of codelets: Every codeletwaiting on the coderack has somenonzero probability of being execut-ed at any time, but once a codelet isexecuted, it can change the environ-ment so as to change the probabilityof waiting codelets being executed.Executing a codelet can also lead toother codelets being placed on thecoderack; for example, if one codeletrecognizes a potential structure, itcan place a codelet on the coderackto build the structure. As coherentstructure is built, the temperature islowered, making it harder to executecodelets that would tear down theexisting structure. Thus, a coherentstructure emerges without the needfor any top-level control.

A parallel terraced scan is similarin philosophy to other approachesthat consider representation as anemergent property of the interactionof many lower-level processes, suchas neural networks and genetic algo-rithms. The link to genetic algo-rithms seems particularly strong, somuch so that it would not be surpris-ing to see some attempt to learncodelets using genetic algorithms.Such learning might address a poten-tial criticism of codelets; they are pre-coded by the designer of a particular

Fluid Concepts and Creative Analogies: A Review

Bruce Burns

a constant interaction between howwe represent information and howwe process this information. Thesethemes are repeated throughout thebook. Those wanting a quick tourshould read Chapters 2, 4, and 5.Chapter 2, “The Architecture of JUM-BO,” gives a good explanation of aparallel terraced scan. Chapter 4,“High-Level Perception, Representa-tion, and Analogy: A Critique ofArtificial Intelligence Methodology,”explains what high-level perception

Douglas Hofstadter could be considered the

Carl Sagan of AI.

Copyright © 1995, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1995 / $2.00

AI Magazine Volume 16 Number 3 (1995) (© AAAI)

system. However, Hofstadter is lessinterested in the exploration of thealgorithm’s properties than in thefact that a parallel terraced scan pro-duces the behavior that is of realinterest to him—the probabilisticemergence of structural representa-tions. Hofstadter is attempting tocombine the ability of neural net-works to detect similarity with thepower that symbolic systems get frombeing able to represent structure.

The central idea behind Hof-stadter’s work is that cognition ishigh-level perception, an idea cap-tured in the slogan cognition is recog-nition. Low-level perception is definedas the processing of information fromthe sensory modalities, and high-level

with especially tough problems” (p.187); rather, it provides a powerfulway of fleshing out the representa-tion of any given situation. Further,the flexibility of analogies (they allowconcepts to “slip” to related con-cepts) explains the fluidity of con-cepts; Hofstadter therefore usesanalogies to account for creativity.

The line between high-level andlow-level perception is fuzzy, but theimplication Hofstadter draws for AIresearch is clear: Too often AI hasfocused on the processing of rigidrepresentations and has ignored howthese representations are formed andhow their formation interacts withtheir processing. (Hofstadter exemptsmachine vision and language pro-

competing programs have addressedreal-world problems in science andpolitics. Hofstadter and his colleaguesargue with some force that this criti-cism is misguided because there is somuch information available in theworld that a fully developed model ofhigh-level perception is not possibleat this stage. He contends that therelevant issues can be addressed in arestricted domain. To do so mightseem less impressive than solvingreal-world problems, but programsthat address such problems actuallyrestrict their domains, making themtractable. Further, the part played bythe stripped-down nature of realdomains in a program’s success hasnot always been acknowledged. Hof-stadter has a worthwhile point here;AI research has ignored the advan-tages of carefully constructedmicrodomains, but it has allowed theuse of vaguely constructed domainsthat are also restricted. The use ofmicrodomains is not without prob-lems, however, because they, too, cancontain hidden assumptions. It is notclear to what extent the programspresented in Hofstadter’s book suc-ceed because they avoid representa-tional issues that are relegated to low-level perception.

Another possible reason why Hofs-tadter’s impact has been limited canbe traced to his method of engage-ment with those he disagrees with.His form of argument is often basedon analogies and rhetorical questionsrather than the analysis of what aprogram aims to achieve and howwell it achieves its goals. The facilityof his writing sometimes makes it tooeasy for him to stray in the directionof ridicule. Thus, in Preface 4, TheIneradicable ELIZA Effect and Its Dan-gers, Hofstadter claims that theapparent successes of a number of AIprograms are analogous to that ofWeizenbaum’s (1976) ELIZA program,which appeared to be interactingwith a person but was really only giv-ing back canned responses. The dan-ger of a program’s success being basedon its simply giving back what theprogrammer built into it is wellknown, and there is probably no pro-gram that does not partly succeedbecause of what is built into it. How-

82 AI MAGAZINE

Book Review

Hofstadter makes many compelling andevocative points in this book. Why, then,has his work had limited impact on thefield of AI?

perception is extracting the meaningfrom the raw material of low-levelperception by accessing concepts andmaking sense of information at aconceptual level. Such an approachto cognition is not new; it was thecentral idea of the Gestalt psycholo-gists, and Hofstadter traces its rootsto Kant. However, modeling cogni-tion as a form of perception con-trasts with the more dominantapproach in AI, which assumes thatcognition is a process of manipulat-ing symbols. In Hofstadter’s view,cognition is instead modeled as aninteraction between the building ofrepresentations and the manipula-tion of representations. Hofstadter’swork has come to focus on analogy:He claims we are constantly perceiv-ing situations in terms of what wealready know. Analogy, defined thisbroadly, is not just a “heavy weaponwheeled out now and then to deal

cessing from this criticism becausethey stop short of modeling pro-cesses at a conceptual level.) It is notenough to leave representation to ayet-to-be-developed representationmodule because the interactionbetween representation and processis crucial. For example, Hofstadterargues that BACON (Langley et al.1987) misses the point by aiming toshow that given the appropriatedata, the program can derive Kepler’sthird law of planetary motion—because the key to Kepler’s discoverywas identifying the appropriate data.

Hofstadter makes many com-pelling and evocative points in thisbook. Why, then, has his work hadlimited impact on the field of AI?One criticism has been that all theprograms created by his group haveaddressed only microdomains: letter-string analogies, table settings, wordpuzzles, and letter fonts. Meanwhile,

ever, the ELIZA effect is not an all-or-none proposition, and the evaluationof the degree to which a program suc-ceeds for interesting reasons and towhat degree it succeeds for uninter-esting reasons needs to include ananalysis of the program’s assump-tions and goals. Such an analysis alsoneeds to be applied to Hofstadter’sprograms: Do they really solve theproblems that they criticize others forfailing to address? For example,Falkenhainer, Forbus, and Gentner’s(1990) SME analogy-mapping programis repeatedly criticized because thesymbols it uses are not grounded; forexample, it would operate just as wellif the symbols water and ice werereplaced with A and B. This result isunsurprising given that syntax drivesthe mapping process in their theory.However, the same argument turnedon COPYCAT is quickly dismissedbecause if an astute human observercontinually saw the symbol SIGMAinvoked in the presence of successorgroups and successor relationships,then this observer would quicklyfigure out that SIGMA stands for theidea of successor. But the symbolgrounding problem is not so easilysolved. How does the observer knowthat the concept of successor group ispresent? Why could not an equallyastute observer, on seeing water andice, infer that the arbitrary symbolwith them means melts?

Hofstadter’s criticisms raise thewhole issue of how AI programsshould be evaluated, the topic ofPreface 5: The Knotty Problem ofEvaluating Research. In this preface,Hofstadter considers and then dis-misses a number of AI models, oftenexplicitly on the grounds that theyfail to agree with his intuitions (forexample, the reference in the indexto SOAR for this section is “SOAR pro-gram…ambitiousness yet boringnessof” [p. 515]). Hofstadter’s point isthat ultimately, all AI research isbased on intuitions that cannot bedefended. However, he firmlybelieves that AI is a field searching forits foundations; so, it is not surpris-ing that any project is based 90 per-cent on intuition (he acknowledgesthat it is dangerous to make thispoint in print). Hofstadter has a

point here because AI models relyheavily on assumptions that aredifficult to test, and thus, they areextremely hard to compare and vali-date. These difficulties lead to thecommon occurrence of rivalresearchers talking past each other.Nonetheless, I cannot agree with hisresponse, which appears tantamountto withdrawal from the argument.Even if only 10 percent is not intu-ition, then this 10 percent should notbe ignored because it might be thebasis for a future foundation. Further,if you want people to build on yourwork, you must convince them thatyour intuitions are correct; in whichcase, you are required to do morethan just tell them that you are cor-rect and make eloquent analogies.

If Hofstadter truly believes that AImodels cannot be defended rational-ly, then it might explain why hisanalysis of other models can appearshallow, but by doing so, he risksirrelevance. This outcome would beextremely unfortunate because Ibelieve his basic ideas might be onthe right track. However, figuring outwhat exactly his ideas mean anddefending them will take more thanHofstadter telling us his intuitions. Irecommend reading this bookbecause it is eloquently written andprovocative. However, at the mo -ment, Hofstadter makes his challengetoo easy to ignore for those whoshould take the most notice of it.

ReferencesFalkenhainer, B.; Forbus, K. D.; and Gent-ner, D. 1990. The Structure MappingEngine. Artificial Intelligence 41: 1–63.

Langley, P.; Simon, H. A.; Bradshaw, G. L.;Zytkow, J. M. 1987. Scientific Discovery:Computational Explorations of the CreativeProcess. Cambridge, Mass.: MIT Press.

Bruce D. Burns is a postdoctoral fellow inthe Department of Psychology at the Uni-versity of California at Los Angeles. Hehas done both experimental and compu-tational modeling work on analogy.

FALL 1995 83

Book Review

Computation &IntelligenceEdited by George F. Luger

Published by the AAAI Press /The MIT Press700 pp., $35.00 paperback ISBN 0-262-62101-0

This comprehensive collection of29 readings covers artificial intelli-gence from its historical roots to

current research directions and practice.With its helpful critique of the selections,extensive bibliography, and clear presen-tation of the material Computation andIntelligence will be a useful adjunct to anycourse in AI as well as a handy referencefor professionals in the field.

The book is divided into five parts.The first part contains papers that pre-sent or discuss foundational ideas link-ing computation and intelligence,typified by A. M. Turing's “ComputingMachinery and Intelligence.” The sec-ond part, Knowledge Representation,presents a sampling of the numerousrepresentational schemes by Newell,Minsky, Collins and Quillian, Winograd,Schank, Hayes, Holland, McClelland,Rumelhart, Hinton, and Brooks. Thethird part, Weak Method Problem Solv-ing, focuses on the research and designof syntax based problem solvers, includ-ing the most famous of these, the LogicTheorist and GPS. The fourth part, Rea-soning in Complex and Dynamic Envi-ronments, presents a broad spectrum ofthe AI communities' research in knowl-edge-intensive problem solving, fromMcCarthy's early design of systems with"common sense" to model based reason-ing. The two concluding selections, byMarvin Minsky and by Herbert Simon,respectively, present the recent thoughtsof two of AI's pioneers who revisit theconcepts and controversies that havedeveloped during the evolution of thetools and techniques that make up thecurrent practice of artificial intelligence.

To order, call toll free 1-800-356-0343(US & Canada) or (617) 625-8569, Mas-terCard & VISA accepted. Prices will behigher outside the US and are subject tochange without notice. Visit the AAAIPress website! http://www.aaai. org/Pub-lications /Press/press.html