Reviewed by
Benoit Lemaire
"Anatomy of the mind", Ron Sun’s new book, is about computational cognitive modeling. More precisely, the book describes the generic architecture CLARION that offers a framework to implement all sorts of cognitive models, from immediate memory to social behavior, through categorization, decision making, and motivational processes. The idea behind Sun’s proposal is that understanding human cognition cannot be achieved by simply aggregating small specific models, tailored to specific tasks. Instead there should be a unification process that constrains the specific models by the fact that they are all working in the same brain. CLARION proposes generic mechanisms to implement knowledge and procedures that can mimic human behavior, either implicit or explicit.The book is composed of about 400 pages: 135 pages for describing the basic components of CLARION, 135 pages for presenting how CLARION can account for a large variety of classical results or "psychological laws", 50 pages for the description of social simulation problems and 50 pages for discussing the strengths and scope of the architecture. The described social simulations concern a tribal society survival task, a collective decision making task and the development of academic publications.
Since the goal of the author is to tackle all aspects of cognition ("maximum scope and minimum mechanisms"), the framework sometimes appears to be very general and the choice between all available mechanisms seems to be left to the modeler of a specific task. It is worth noting that the book does not contain a how-to section with guidelines for implementing a particular cognitive or social process. In addition, the particular details of the numerous examples of implementation are often not described; the reader wanting to implement a particular model is required to refer to the author’s publications to find the details of a similar approach. Therefore "Anatomy of the mind" is more of a theoretical than a practical book.
One of the merits of the book is that the rationale behind each component is presented and well-documented. The bibliography contains almost 400 references. The choice of each component is justified with respect to existing behavioral results in the scientific literature. Ron Sun compares his approach to the classical BDI framework which has some similarity with CLARION (Beliefs/Knowledge, Desires/Drives, Intentions/goals) but he says that the main difference is that BDI is rather based on folk psychology and has no demonstrated psychological validity. This raises the difference between artificial intelligence and computational cognitive modeling: AI scientists design artificial intelligent agents whereas cognitive modelers design models of intelligent humans. The latter are in need of psychologically valid formalisms, which may not be the case of the former.
Ron Sun shows how CLARION is able to reproduce many classical behavioral results. Actually, there are two different ways for assessing a computational model: either by using experimental data of a given task and computing the numerical difference with model predictions, or by showing that the model is able to reproduce effects that were found in many experimental measurements. CLARION is generally tested on effects, not magnitude, because the goal is to account for a large variety of phenomenon and not looking for a precise quantitative fitting. This is acceptable because experimental psychology does not go further: researchers test theories by looking for the effects that they predict.
A nice chapter towards the end of the book is devoted to provide answers to general questions that one could ask about CLARION, such as "can CLARION be disproved?", "how does CLARION characterize implicit and explicit processes?", "are there too many free parameters?" or "why has social simulation been emphasized?".
To sum up, it is an interesting book that provides a deep understanding of the CLARION architecture. However, given a model, it is still hard to tell what belongs to the core of CLARION and what was specifically added for that specific model. Still, CLARION is the result of a 20 year long attempt to build a unified theory of cognition, which would interest cognitive and social scientists but also AI researchers.
Return to Contents
of this issue