Similarity and Categorization

AI Magazine Volume 23 Number 2 (2002) (© AAAI)

Book Reviews

gories we have are natural partitions of the world. This position is support- Book Reviews ed by data collected by Eleanor Rosch and her colleagues in the 1970s. In other words, we have the categories we do because they preserve existing similarities among objects and are therefore informative. For example, Similarity and our categories tend to be inductively powerful (for example, If I know an object is a bird, I can assume it has Categorization feathers and probably flies). Critics of this view contend that the notion of similarity is vague and ill de- A Review fined and carries no explanatory force on its own. As the philosopher Nelson Goodman (1972) noted, objects can be similar in an unlimited number of Bradley C. Love ways (for example, weigh an odd number of grams, are less than 100 meters long, are less than 101 meters long). According to Goodman, one hat is the nature of human researchers are adopting richer and must specify in what respect two ob- categories? How do we form more well-developed similarity-based jects are similar, or the statement is Wcategories? What is the role approaches. These second-wave simi- empty. If two objects are similar only of similarity in categorization? Can we larity-based approaches address some because they are in the same category, formalize the answers to these ques- of the shortcomings of previous ap- then similarity is a vacuous notion, tions to derive further insights and de- proaches. The 13 chapters that form and any account of categorization velop useful software systems? These this volume are broad in scope but can based on similarity is circular. A sec- were the questions addressed at an in- be characterized as couching catego- ond line of attack on similarity-based terdisciplinary meeting attended by rization in terms of more sophisticat- accounts is that similarity is not a psychologists, computer scientists, an- ed and precise notions of similarity. powerful enough construct to account This work shows promise in elucidat- thropologists, statisticians, and for human categorization. For exam- philosophers held at the University of ple, a man who jumps into the swim- Edinburgh. The edited volume Simi- ming pool at a party while he is wear- larity and Categorization arises from ing a tie and sports coat is classified as this meeting. Similarity and Categorization, Ul- drunk not because he is similar (at The publication of Similarity and rike Hahn and Michael Ramscar, least in a straightforward sense) to editors, Oxford University Press, Categorization is timely because the other examples of drunk people. New York, 279 pp., 2001, ISBN 0- study of categorization is at a theoret- 19-850628-7. Rather, he is classified as drunk be- ical crossroads. In the 1970s, similarity cause this behavior is in accord with was thought to be the basis of catego- our theories of the kinds of things that rization—an object was assigned to drunk people do. This theory-based the category to which it was most sim- view of categorization has been popu- ing the nature of human categoriza- ilar. In the 1980s, a new view emerged larized by Murphy and Medin (1985) tion. that held that similarity was too weak and is reflected in AI techniques such The editors, Ulrike Hahn and and vague a construct to ground hu- as explanation-based learning (DeJong Michael Ramscar, do an excellent job man categorization. On this view, our and Mooney 1986). categories cohere by virtue of being in the introductory and concluding Although the theory-based view of embedded in explanatory systems chapters of explicating the historical categories does touch on some short- (that is, our knowledge of the world is shifts in the perceived relation be- comings of the earlier similarity-based organized around theories akin to sci- tween similarity and categorization. I approaches, the theory-based view has entific theories). According to the the- briefly overview these shifts and how itself proven vague and has not made ory-based view, judgments of similari- the contributed chapters fit into the large headway in understanding hu- ty are largely governed by theories latest movement. man categorization. In contrast, the that determine the relevant properties The idea that similarity is the basis contributors to this volume have for evaluation. Recently, dissatisfac- for categorization is intuitive given made headway by considering richer tion has emerged with the theory- that similar objects tend to be in the accounts of similarity. based account of categorization, and same category. On this view, the cate- For example, Arthur Markman con-

FLAIRS-2002—Pensacola, Florida Proceedings of the Fifteenth International Florida Artiﬁcial Intelligence Research Society Conference

Edited by Susan Haller and Gene Simmons

ISBN 1-57735-141-X 500 pp., illus., index

The topics covered in this year’s proceedings include interactive tutoring, machine learning, constraint satisfaction and theorem proving, robotics, genetic algorithms, agent architectures, automated reasoning, and natural language processing. The proceedings also features papers from fourteen special tracks on artificial intelligence in educational information technology, spatiotem- poral reasoning, semantic web, neural networks, machine learning, evaluation of intelligent systems, categorization and concept representation models and implications, imprecise and indeterminate probabilities in artificial intelligence, uncertain reasoning, integrated intelligent systems, knowledge management, case-based reasoning, artificial intelligence in areospace, and behavioral characteristics and AI systems.

siders analogical forms of similarity ties is. He suggests (as do some theo- similarity than previous similarity- computed over structured representa- ries of memory) that the “feeling” of based accounts. tions, as opposed to previous ap- similarity plays a role in coordinating Many of the contributors stress the proaches in which representations are retrieval strategies. importance of developing computa- assumed to be spatial or featural. Other contributors consider the tional models. Rodriguez offers a re- Markman distinguishes between dif- importance of causal relations in the view of case-based reasoning ap- ferences that are related to commonal- perception of similarity. Keane, proaches, and Pothos and Chater de- ities (alignable differences) and differ- Smyth, and O’Sullivan go further and velop a clustering approach to ences that are not (nonalignable differ- consider how the perceived similarity category formation. Their approach ences). For example, the number of of two objects is affected by their pro- discovers the best partitioning of the wheels is an alignable difference for cessing histories. Items sharing simi- data (in terms of preserving the pair- cars and motorcycles that arises out of lar processing contexts tend to be wise similarity relations in the raw da- the commonality of having wheels. viewed as more similar. These find- ta). Pothos and Chater’s method is An example of a nonalignable differ- ings parallel work demonstrating based on finding the minimum–de- ence for this pair is a car’s seat belt. that thematic relations influence scription-length solution for a set of Psychologically, these two types of dif- similarity. For example, a hammer pairwise similarity ratings that yields ferences are distinct. For example, and a nail are perceived as more sim- the greatest compression (in terms of high-similarity object pairs actually ilar because of their interaction. bits). Gosselin, Archambault, and have more alignable differences than Work in latent semantic analysis that Schyns describe a model of categoriza- low-similarity pairs. Alignable differ- derives semantic representations of tion that explains the relative speed of ences are also remembered better than words through examination of usage classifying objects at different levels in nonalignable differences and play a patterns in large text corporal is a re- a category hierarchy. Palmeri’s chapter larger role in preference formation. lated approach. The aforementioned explores the time course of category Markman ponders what the metacog- contributors all consider a wider learning. He evaluates how well two nitive function of detecting similari- range of input in the calculation of exemplar models (that is, models in

104 AI MAGAZINE Book Reviews which each learning episode leaves an Is similarity so unconstrained that it independent trace in memory, and subsequent items are classified by cal- will never provide a full account of culating pairwise similarities with ev- categorization? ery item stored in memory) predict the time course of classification. One … model, EGCM, captures how stimulus Similarity itself will not explain properties become sequentially available during perception, affecting the categorization, but understanding activation of stored exemplars in how similarities are processed and memory (for example, form might be represented will go a long way perceived before texture). The second model considered by Palmeri is the ex- toward understanding categorization. emplar-based random walk model (EBRW). EBRW works by sequentially re- trieving exemplars from memory that are fed into a diffusion decision proce- dure that settles on a classification when evidence for category member- with similarity-based accounts). For goes far in integrating theory and sim- ship crosses a decision boundary. The example, people judge canonical ilarity-based accounts and demon- mathematics behind the model are members of the category of odd num- strates that similarity-based accounts well motivated, and the predictions bers, such as the number three, as are not as limited as some might are verified through behavioral exper- more typical of odd numbers than think. iments. other odd numbers, such as the num- Still, the question remains, is simi- Other researchers strive to clarify ber 687, despite the fact that a clear larity an imposter as Goodman ar- what is meant by the terms similarity criterion for category membership is gues? Is similarity so unconstrained and category. Hampton’s chapter dis- available (Gleitman et al. 1996). that it will never provide a full ac- tinguishes between concepts that are The concluding chapter offers an count of categorization? The answer is social constructs and those that are informative discussion of the relation both yes and no. Similarity itself will not. He argues that similarity-based between similarity and theory-based not explain categorization, but under- accounts of categorization fair better views. The editors note that the no- standing how similarities are pro- when applied to concepts that are not tion of theory is at least as unwieldy cessed and represented will go a long culturally defined. For example, a and vague as the notion of similarity. way toward understanding categoriza- bank note is not categorized as legal They also note that many of our suc- tion. The current volume makes sub- tender because it is similar to other ac- cessful theories of the world arise stantial progress on this front. ceptable notes. Rather, the note’s sta- from, or conform to, similarity-based tus is tied to its origin. The properties accounts. For example, biological tax- References that need to be evaluated to determine onomies of animals ordered by DNA DeJong, G., and Mooney, R. 1986. Explana- origin are not accessible to nonex- roughly conform to previous tax- tion-Based Learning: An Alternative View. perts. Hampton argues that similarity- onomies that are ordered by the Machine Learning 1(2): 145–176. based accounts, although viable for salient properties of the animals. The Gleitman, L. R.; Gleitman, H.; Miller, C.; many routine categorization tasks, are editors also note that the application and Ostrin, R. 1996. Similar, and Similar not appropriate for such situations. and interpretation of a theory relies on Concepts. Cognition 58(3): 321–376. Along these lines, Sloman, Malt, and similarity-based operations. For exam- Goodman, N. 1972. Seven Strictures on Friedman show that certain measures ple, there is no definitional sequence Similarity. In Problems and Projects, 22–32. Indianapolis, Ind.: Bobs-Merrill. of similarity do not predict object of DNA that specifies what a tiger is. naming (a categorization task) and Evan Heit’s chapter further blurs the Murphy, G. L., and Medin, D. L. 1985. The Role of Theories in Conceptual Coherence. that differences occur across cultures. theory-similarity dichotomy. Heit’s Psychological Review 92(3): 289–316. They argue that social convention model is exemplar based (a similarity- plays a role in naming, which leads to driven model) but captures the effects divergences between similarity and of prior knowledge and theories by Brad Love is an assistant professor in cog- naming. Certainly, there are some cat- seeding the model with appropriate nitive psychology at the University of egories in which similarity does not exemplars. The model is successful in Texas at Austin. He received his Ph.D. in cognitive psychology from Northwestern provide the definition. Some (very few addressing a range of human data that University and his B.S. in cognitive and lin- in reality) categories have tractable until now were out of the reach of sim- guistic sciences from Brown University. His definitions, such as the concept of tri- ilarity-based models (although there main research interest is empirical and angle. Interestingly, even when a defi- are some unresolved theoretical issues computational explorations of human cat- nition is available, all category mem- surrounding how the set of initial ex- egory learning. His e-mail address is bers are not treated equally (consistent emplars is constructed). Heit’s model [email protected].

SUMMER 2002 105