Understanding Cognitive Science
Chow, Siu L., Canadian Psychology
MICHAEL R. W. DAWSON Understanding Cognitive Science Oxford, UK Blackwell Publishers Inc., 1998, 351 pages (ISBN 0-631-20895-X, C$52.85, Softcover) Reviewed by SIU L. CHOW
In this book, the mind is an information processor. It is studied as a physical symbol system (which is a special case of the Universal Turing Machine, UTM) in the classical approach, but as a network of connections (a la neural connections in the brain) in the connectionist approach. The first theme of the book is the necessity and sufficiency of the tri-level criteria for assessing the success of cognitive science, the three criteria being (a) a clear delineation of what the cognitive problems are, (b) an account of the transition of knowledge-states in the course of solving cognitive problems, and (c) a description of the physical device that implement the changes in knowledge-states. The second theme is that the differences between the classical and connectionist approaches are more apparent than real because both approaches have the potential to satisfy the trilevel hypothesis.
It is argued that any satisfactory account of the mind as an information processor must give satisfactory descriptions at the computational (semantic), procedural (algorithmic), and physical (implemental) levels. A computational description of the machine serves to delineate the domain pertinent to cognitive science. Both approaches satisfy the first criterion because all cognitive scientists agree on what "information" means.
The steps used in solving cognitive problems (e.g., a chess game) must be made explicit in the procedural description. The classical approach is superior to the connectionist approach because explicit, mechanical, step-by-step problem-solving rules can be identified for the UTM, whereas the connectionist often cannot describe how a successful network achieves its feats.
The implementation description provides details of the medium in which the processing procedures are realized. It is in the form of a serial UTM in which there is a sharp separation of data and processes. In the connectionist approach, processing is implemented in a brain-like network in which data are distributed throughout the network whose multiple processors work in parallel. Moreover, the network can learn to solve problems without applying rules by virtue of its hidden units and back-propagation, error-correction capability. However, a brief assessment of the tri-level hypothesis sets in high relief why cognitive science has little or nothing to do with the human mind.
Consider the necessity of giving a description of the physical properties of the information processor (i.e., the implementation description). There is no self-evident reason why knowing the material of which a gadget is built is essential to understanding how (or why) it works. In other words, the implementation description plays no role in understanding the problem-solving machine.
It is a thesis of the tri-level hypothesis that an understanding of the mind requires knowing the processing steps that solve cognitive problems (i.e., having a computation description). There are many examples that highlight the success of cognitive science in this regard. Be that as it may, the computation description is debatable for different reasons for the classical and the connectionist accounts. Computation is achieved in the UTM by carrying out mechanical algorithms that are devised analytically (hence, called "analytic algorithms"). The issue of validation arises because analytic algorithms are not observable. The validation criterion used is Pylyshyn's (1984) strong equivalence, which subsumes Newell and Simon's (1972) sufficiency criterion. Specifically, an analytic algorithm is validated when it matches the human subject's protocol algorithm (viz., the sufficiency criterion) in exactly the same way (viz., the strong equivalence criterion). This criterion presupposes that the protocol algorithm is veridical. …