- Selected Letters From Readers
Connectionism and Its Consequences
Kip Canfield’s article in the last issue of Postmodern Culture is one of the first pieces of critical theory to discuss the implications of the revolutionary paradigm shift now taking place in the cognitive sciences. The movement inspiring that shift, generally called connectionism, offers a powerful and still controversial alternative to the standard model of mental representation which has more or less dominated western philosophy at least since the Enlightenment (Bechtel). As Canfield notes in his comparison of Paul Smolensky, one of the leading connectionists, with Jacques Derrida, the connectionist critique parallels in many ways the deconstruction of traditional semiotics and structuralism. But, as Canfield also notes, connectionist theory does more than deconstruct the old paradigm: it also purports to offer an alternative account of representation, a genuine Copernican revolution which changes our view of mental life from a symbol centered token/type model to a network based vector/matrix account (Churchland).
Since Canfield focuses almost exclusively on Smolensky and Derrida, however, readers of his essay not already familiar with the paradigm wars in cognitive science may not see just how profound the connectionist revolution could be. For that reason, I want to offer here a brief note on the consequences of connectionism which follows up on a point I have made elsewhere regarding the implications of connectionism for critical theory (Miers).
The attraction of connectionism for cognitive science is its potential for providing a “natural” theory of information processing in the brain. All the evidence indicates that the brain itself is organized as a massively distributed parallel processor (Edelman); what the great debate is about is how to reconcile the neural evidence with the standard, classical theory of mental representation. Classical theory claims that mental representations are the product of arbitrary atomic symbols (i.e., signs rather than symbols in the Coleridgean sense) operated on by formal syntactic rules. Connectionist theory, on the other hand, focuses on the fact that neural networks have no one- to-one mappings, and that they respond to input vectors not by invoking rules, but by dynamic transformations of the network (Churchland). The real issue then is not a simplistic opposition between atomic and distributed elements or between arbitrary and motivated symbols; it is, rather, the problem of explaining how the brain produces the apparent formalism of symbolic representation from the non- classical structure of neural networks.
In broad terms, there are two explanations for this puzzle, the first of which I call weak connectionism or neo- symbolism and the other I term strong or pan-connectionism (Miers; see also Bechtel). In the weak version, the logic of neural networks serves simply to implement some version of the traditional token/type symbol processing which the classical cognitivists like Fodor see as essential for mental representation (Fodor, “Connectionism and Cognitive Architecture”). In the weak case, the classical cognitivists can, with some modifications, “save the appearance” of the symbolic paradigm. Even though connectionist processes may infiltrate and shape whole aspects of mental life, there remains a unique domain of rational or propositional thought governed by the formalisms of symbolic structures. As Fodor has consistently argued for years, not all of the messy flux of what passes through our heads needs be structured like a formal language, and Fodor is quite willing to concede much of mental life to Freudian and even Skinnerian accounts (The Language of Thought 200).
Strong connectionism, on the other hand, is much more radical and leads to a rather uncanny picture of the apparent symbol processing capabilities of the mind. Strong connectionism claims that symbolic representation is, in fact, a rather shallow illusion which is being approximated or mimed by a wholly connectionist strategy that evolved in the brains of mammals long before the appearance of humans. In this account, our sense that there are atomic tokens is being created by a series of vector/matrix interactions, and our notion that tokens belong to some double system of message and code or object language/meta-language is a belated allegory...