Browse Books

Reviewer: Philip L. Phipps

Once in a while I read about an insight that I feel is right on. In my opinion, Clark's book contains such an insight. His purpose is to demonstrate and describe an architecture of our minds that more accurately explains how our minds are organized and operate (think) on different levels of processing detail. He does this fairly successfully by analyzing the arguments for connectionist models (neural nets) and those for classical AI models (symbolic computations). Many of the proponents of connectionism are cognitive scientists, but some are well-known AI people. He then proposes a composite description of the mind, using aspects of both paradigms. The book is a challenge to read because of both the many big words and the implied technical background. It will be especially difficult for someone unfamiliar with concepts and issues in artificial intelligence and distributed processing. Apparently the book is meant for cognitive scientists, AI theorists, and philosophers like the author. A good preliminary would be the two volumes of Rumelhart and McClelland [1]. Clark's subtitle “Philosophy, Cognitive Science, and Parallel Distributed Processing” is, by his own admission, rather intimidating. The author is a philosopher first, with AI and biology his associated interests. He thus gives a very interesting view of issues in these secondary domains. Clark's style is entertaining because of his descriptive language, subtitles, and the monologue arguments he gets into. The discourse is further enhanced by many quotes and a number of interesting thought experiments that illustrate the concepts, arguments, and claims. In the introduction, “What the Brain's-Eye View Tells the Mind's-Eye View,” Clark gives his impression of what is to come and states the goals of his work. The goals are to show that “(1) folk psychology does not seek to model computational processes . . . and (2) parallel distributed processing and conventional approaches to mental modeling need not be uniformly regarded as competing paradigms of cognitive architecture. . . .” Part 1, “The Mind's-Eye View,” consists of four chapters: “Classical Cognitivism,” “Situation and Substance,” “Folk Psychology, Thought, and Context,” and “Biological Constraints.” In chapter 1, Clark gives an outline of how AI came about (Turing made AI conceivable, McCarthy et al. made it possible, and we are “still waiting to make it actual”). In the next chapter, he summarizes arguments by Dreyfus and Searle that attack classical cognitivism and AI as inadequate for use in cognitive science. Folk psychology is used by people to track mental states in other people and is a poor tool for use in modeling and explaining the substructure of people's minds. It does provide the observational data that the more formal systems must account for. “Biological Constraints” is an interesting discussion of why and how our brains and minds might have evolved as the human race developed. The purpose of this chapter, especially, is to identify those biological constraints that any model or theory must satisfy if it is to represent a human mind. Part 2, “The Brain's-Eye View,” contains the remaining six chapters: “Parallel Distributed Processing” (PDP), “Informational Holism,” “The Multiplicity of Mind: A Limited Defence of Classical Cognitivism,” “Structured Thought, Part 1,” “Structured Thought, Part 2,” and “Reassembling the Jigsaw.” The book is completed by an epilogue (which could have been left out), an appendix on further arguments for connectionist systems and their limitations, notes and side comments, a very good bibliography, and an index. The fifth chapter briefly summarizes PDP or connectionism, with its distributed representations and superpositional storage, and evaluates its applicability over symbolic processing. Chapter 6 discusses informational holism, an attribute of PDP that has the ability to shade meanings, perform abstraction, and exhibit symbolic flexibility. In the next chapter, however, we learn that PDP has failings at tasks in which classical AI does well. Hence the mind is probably made up of multiple computational architectures. The two chapters on structured thought further analyze and illustrate the limits of PDP and classical AI. These illustrations include representations of the rules of grammar, explanations for the errors in past-tense recognition in children and machines, dysphasia, and naive physics (common sense about our world). Chapter 9 ends with a discussion of what it means for a system to have thoughts. Finally, chapter 10 puts it all together with an instantiated “thinker” built of microfunctional elements in a PDP-like architecture but augmented with algorithmic elements to do symbolic type processing. The latter could well be implemented as PDP substructures. This book is quite thought-provoking as it objectively reviews the pros and cons of various theories of a number of authors about what is required to model the human mind and what it means to think (for a human or machine).