Why the mind is (still) not a network

Abstract

The Algebraic Mind: Integrating Connectionism and Cognitive Scienceby Gary F. Marcus, MIT Press, 2001. $27.95/£19.95 (xii + 224 pages)ISBN 0 262 13379 2This short book provides one of the best available expositions of the nature and shortcomings of the connectionist approach to cognitive modeling. It is a thorough, clearly written, and generally accurate review of the nature of this popular movement.Gary Marcus has been an effective critic of the connectionist approach to modeling the mind, especially as it applies to language acquisition. In this book he covers the range of published arguments concerning the shortcomings of connectionism. This is its purpose and its strength. It is not, however, a novel way of casting the debate between the symbolic and the connectionist approaches, as implied in the introductory chapter, nor a proposal for reconciling the two approaches (as suggested by the book's subtitle). The book will provide little solace for supporters of connectionism as a way of understanding mental processes.Marcus claims (p. 3) to shun the distinction between ‘eliminative connectionism’ (the claim that connectionist systems will eventually replace the systems of rules and of symbol-processing posited by classical cognitive science models) and ‘implementational connectionism’ (the view that connectionist systems can provide biologically plausible models of the mechanisms used in carrying out symbol processing). Nonetheless, he goes on to make a strong case for the implausibility of eliminativist connectionism. His analysis of the systematic failure of connectionist models to represent individuals (tokens), to provide a compositional system for building complex concepts out of simpler ones, to provide for rule schemas with variables, to deal with what he calls the ‘ubiquitous Universally Quantified One-To-One Mappings’, and so on, along with his detailed focus on language learning (where Marcus has personally contributed empirical evidence that is damaging to the connectionist proposals) show some of the reasons why eliminative connectionism fails.When it comes to implementational connectionism, the initial chapter claims to be much more concessive than the rest of the book actually warrants. For example, Marcus says, ‘If it turns out that the brain does in fact implement symbol-manipulation, implementational connectionism would…provide an enormous advance, tantamount to figuring out how an important part of the brain really works.’ (p. 3). This might appear to give comfort to connectionists who believe their approach to be in harmony with biological bases of cognition, a belief based on a number of misunderstandings and on the illusion that, because they bear a superficial similarity to systems of neurons, connectionist networks are brain-like. But this book gives such connectionists little reason to rejoice, notwithstanding the above quote. If existing connectionist systems are to be viewed as proposals for how symbolic systems might be implemented, then they are not models of biological implementation, but of implementations in some more abstract cellular networks that owe little to the data of neuroscience. Few connectionist systems concern themselves directly with biological evidence, beyond such general claims as that such systems are built from simple elements, are distributed, and are robust in the face of physical damage, none of which are inherent properties of connectionist systems any more than of symbolic systems 1xConnectionism and cognitive architecture: a critical analysis. Fodor, J.A. and Pylyshyn, Z.W. Cognition. 1988; 28: 3–71Crossref | PubMed | Scopus (1117)See all