Parallel Distributed Processing: Explorations in the Micro-structure of Cognition

These two volumes, although only in part concerned directly with child language, nevertheless represent an approach which will increasingly impinge on acquisition theory. The two volumes are best described as a connectionist manifesto: wide in scope and full of energy, willingness to attack rather than to side-step problems, invention and promise. James McClelland and David Rumelhart edited the book and they also contributed to a majority of the chapters. The book introduces connectionism (or parallel distributed processing (PDP)) and demonstrates how it can be applied in varied areas of psychology, biology and artificial intelligence. The first volume is a general outline of foundations of the PDP research programme and of basic principles of connectionist models. The second volume presents specific models of various psychological and biological processes. (A third volume which should be available by the time this review is published makes available connectionist tools.) The psychological processes covered are schema-based thinking, speech perception, reading, learning and memory, learning morphology, and sentence processing. This form certainly makes the book a good introduction into connectionism. It allows readers to start in the area of their expertise, to familiarize themselves with the approach and to move to chapters outlining principles, or other models, if they get that far. A developmental psychologist interested in getting the flavour of connectionism should certainly read part 1 in the first volume and part 4 in the second volume. Connectionism is an approach to cognitive science like symbolic approaches, it is concerned with the processes of internal representation and it uses them to explain thinking, perceiving, acting, learning and development. What is new and attractive about connectionism is the kind of representational processes postulated. These are quite different from what researchers in psychology and AI are used to. They are not symbolic; there are no explicit statements of, for example, rules, schemas or prototypes. Connectionist representations are not ' local' single symbols do not represent single entities or their sets. Rather, 'each entity is represented by a pattern of activity distributed over many computing elements and each computing element is involved in representing many entities' (vol. 1, p. 77). So, for example, one network could ' represent' several rules and exceptions to them. We are asked to think about PDP models as a way of IMPLEMENTING symbolic representations in parallel networks: occasionally it is asserted that symbolic models of cognitive processes are only approximations to mechanisms of distributed representations. The problems which exercise con-