What Is Computational Psychology?

Abstract

ly-defined information-processing tasks. With respect to virtually all matters normally studied by psychologists, we cannot identify the basic task. The one possible exception is language-or, rather, parsing. The task of a parsing system is to map the input word-string onto structural descriptions couched in syntactic terms. Syntactic structure enters into our definition of what language is. So whether chimps can learn language is now taken to depend, not on their producing humanoid speech-sounds or new meaningful symbol-strings, but on their ability to make syntactic disThis content downloaded from 157.55.39.254 on Sun, 04 Sep 2016 04:48:29 UTC All use subject to http://about.jstor.org/terms WHAT IS COMPUTATIONAL PSYCHOLOGY? 27 criminations. We already have some candidate syntactic descriptions, and (controversial) abstract arguments about the relative power of transformational versus context-free grammars to effect the mapping. We even have one example of an abstract task-identification and a computerized mapping-algorithm based on it, where these are clearly conceived by their authors at distinct explanatory levels. Noam Chomsky's formal grammatical rules do not describe actual psychological processes, but abstract structural relationships. It is often said that Chomsky ignored process because he is a linguist, not a psychologist: psychologists are interested in how language happens in the mind, whereas linguists are concerned only with language in itself. This is true, so far as it goes. But it is sometimes taken to imply that if Chomsky had been really interested in psychology, he would have concentrated on specifying the mental processes involved. This is the typical reaction of computational psychologists (in the broad sense). In Marr's sense, however, Chomsky was doing psychology all along. Indeed, he was attempting the mathematical groundwork on which any scientific process-oriented (second-level) explanation must be based. He was right to ignore process, leaving it to others (such as Mitchell Marcus3) to think about the specific algorithms that could exploit the structural constraints discovered by him. Syntax is a relatively simple case (which is not to deny that it is both difficult and controversial). What of semantics? Attempts to formalise semantics have been made, but are also controversial. We cannot, in general, specify the task involved in mapping a word-string onto a semantic representation. As for mapping an extended text onto its thematic representation, we are even more in the dark. And what structural regularities underlie language-use in conversations? In all these cases, there are scattered insights and a variety of computer-models supposedly directed to the problem. But, in the strong sense discussed within this section, we do not have even the beginnings of a computational understanding of these matters. We are in the same position with respect to most topics S M. P. Marcus, A Theory of Syntactic Recognition for Natural Language (Cambridge, Mass.: MIT Press, 1980). This content downloaded from 157.55.39.254 on Sun, 04 Sep 2016 04:48:29 UTC All use subject to http://about.jstor.org/terms 28 I-MARGARET A. BODEN studied by psychologists. Even where it is intuitively plausible that a psychological category rests on an identifiable task, we have little idea of what this might be. As Marr put it, when criticizing explanations of problem-solving based on computer simulations of people doing cryptarithmetic: no doubt something computational is going on when we do mental arithmetic, but it almost certainly is not mental arithmetic. Moreover, there may not be one task for each category. Social psychologists offer 'theories' of attitude-change, but it is not obvious that attitudechange is a natural kind. And what of learning, social interaction, or personality development: is it even plausible that an abstract information-processing definition could be given for these phenomena? General arguments of a 'transcendental' character are needed in each case, specifying a generic information-processing task and identifying computational constraints to be exploited by creatures enjoying psychological powers of the relevant class. Philosophical study of the necessary conditions of (for example) conversation, personal interaction, and intentional action in general might be useful, and comparable discussions are to be found in the literature of psychology and AI. But even where such insights exist, they do not have the degree of precision which enabled Marr to prove that his representational system is in principle capable of computing various sorts of 2D-to-3D mapping whereas others are not. He had physics to help him, and physics might allow abstract computational analysis of other low-level sensory systems. It is not obvious, however, that high-level psychological abilities (even including high-level vision) can helpfully be thought of in this way. If they can, then we may hope for a computational psychology that will explain why mental abilities are as they are, by grounding them in general conditions of experience which constrain all psychological beings. But if they cannot, then (nonsensory) computational psychology is possible-if at all-only in a sense weaker than that defined by Marr.