Cognitive Modeling (Bradford Books)

Features

• Cover Type: Paperback with 1291 pages
• Published by: The MIT Press August 15, 2002
• Written in: English
• ISBN 10 Number: 0262661160
• ISBN 13 Number: 978-0262661164
• Book Dimensions: 9 x 7.1 x 2 inches
• Weighs: 3.9 pounds

Product Description
Computational modeling plays a central role in cognitive science. This book provides a comprehensive introduction to computational models of human cognition. It covers major approaches and architectures, both neural network and symbolic; major theoretical issues; and specific computational models of a variety of cognitive processes, ranging from low-level (e.g., attention and memory) to higher-level (e.g., language and reasoning). The articles included in the book provide original descriptions of developments in the field. The emphasis is on implemented computational models rather than on mathematical or nonformal approaches, and on modeling empirical data from human subjects.

About The Author
Thad A. Polk is Assistant Professor in the Department of Psychology at the University of Michigan.

Colleen M. Seifert is Associate Professor in the Department of Psychology at the University of Michigan.

Reader Reviews
This book could be considered to be a collection of articles on the `computational theory of mind.' Although the articles are somewhat out of date, due to the advances in neuroscience and cognitive science that have occurred since the time of publication of the book, it does serve as a good motivation for the understanding of more recent developments. I did not read all of the articles in the book, and so my review will be confined to the ones that I did. The article on ACT in chapter 2 is basically a theory of cognition that is based on recursion. Referring to ACT as a "simple theory of complex cognition", John Anderson, the author of the article, wants to simulate the manner in which humans develop recursive programs. The machine that is to simulate this makes use of `production rules,' in its knowledge base, which the author claims is exhaustive enough to produce complex cognition. To produce true machine intelligence, all one has to do is to tune these production rules and make use of them as needed. As the author describes it, the original ACT theory was based on human associative memory, but the one described in this article is called ACT-R, and can simulate adaptive behavior in the presence of a noisy environment. The author describes various simulations using ACT-R, and concludes that it is sensitive to prior information and to information about what is appropriate response to the situation it finds itself in. The author stresses more than once the simplicity of the ACT-R system: it is able to encode data from the environment as declarative knowledge, encode the changes in the environment as procedural knowledge, and encode the statistics of this knowledge use in the environment. Another highly interesting article is the one by Alan Prince and Paul Smolensky on the application of optimization theory to linguistics. Called `optimality theory' by the authors in their extensive research on the topic, in the article they discuss the relations between optimality in grammar and optimization in neural networks. The authors discuss with great clarity the role that constraints play in the construction of linguistic structures, and the fact that these constraints typically conflict with each other. This conflict between grammatical constraints must thus be managed by a successful grammatical architecture. Optimality theory asserts that these constraints are universal in the sense that they are present in every language. The connection of optimality theory with neural networks arises when one is interested in finding out if the properties of optimality theory can be explained in terms of fundamental principles of cognition. The computational theory of neural networks the authors believe holds some clues on these properties. In order to make the connection with grammatical issues, as abstract as they are, and because neural networks are highly nonlinear dynamical systems, one must find a way of encapsulating the complicated behavior of neural networks. The authors accomplish this by the use of Lyapunov functions, which for reasons of consistency of terminology they call `harmony functions.' For those neural networks admitting a harmony function, the initial activation pattern flows through the network to construct a pattern of activity that maximizes "harmony." Most interestingly, the harmony function for a neural network performs the same function as does the mechanisms needed for well-formed grammar. The patterns of activation are thus a mathematical analog of the structure of linguistic representations. However, the authors are careful to note that not every weighting scheme for the neural network will give a possible human language. It is here where the constraints play an essential role in limiting the possible linguistic patterns and relations. The article by Keith Holyoak and Paul Thagard discusses the construction of a correspondence between a source analog and of a target. This is the so-called analogical mapping, which is constructed using a collection of structural, semantic, and pragmatic constraints. In the view of the authors, the concept of analogy can be broken down into four components, namely the selection of a source analog, the actual mapping, an analogical inference (transfer), and the actual learning that takes place. The authors omit discussion of the last component in this article. The finding of the correspondences between the two analogs can result in a combinatorial explosion, and so use is made of appropriate constraints. These constraints consist of those that exemplify structural consistency, those of semantic similarity, and lastly of pragmatic centrality. The theory of analogical mapping that the authors propose is governed by these constraints. They discuss the ACME (Analogical Constraint Mapping Engine) algorithm as one that constructs a network of units representing mapping hypotheses and eventually converges to a state that represents the best mapping. They list several applications of ACME, such as radiation problems, attribute mappings, chemical analogies, and the classical `farmer's dilemma' problem. ACME was also able to simulate a number of empirical results related to human analogical reasoning. The analogical mapping they discuss is most powerful in a specific domain however. This domain-specificity is a typical restriction for most of the efforts in learning theory and artificial intelligence.