1.2.1 The Nature of Thought and Language

Why use connectionist models at all? It is not immediately obvious why lists of ones, zeros and floating point numbers should be of much relevance to understanding human cognition. Especially so if one keeps in mind what Fodor & Pylyshyn (1988) have written about the issue: "The main fruit of "brain style modeling" has been to revive psychological theories whose limitations had previously been pretty widely appreciated" (64) given that connectionism can only rely on the "co-occurrence statistics that were the hallmark of old-fashioned Associationism" (31).

Fodor & Pylyshyn based their harsh conclusion on the fact that connectionist models -- at least as they saw them -- failed to account for four essential properties of (human) cognition:

• productivity: we can have infinitely many thoughts using finite means (i.e. the grey matter of the brain);
• systematicity: it is not possible to be able to think that 'Kim kisses Sandy' without also being able to think that 'Sandy kisses Kim' -- no aRb without bRa;
• compositionality: complex thoughts are the result of combining atomic constituents via syntactic rules -- the semantics of the complex expression can be derived in an unambiguous manner from the meaning of the constituents and the rules;
• inferentiality: it is not possible to think 'Kim kisses Sandy' without also thinking 'Kim', 'kisses' and 'Sandy' -- no P&Q without P and/or without Q.

Because cognition is claimed to have these properties, a successful theory will need to make use of 'classical' means -- i.e. representations and processes that are sensitive to the internal structure of these representations.

That Fodor & Pylyshyn's argumentation drew a lot of animated replies goes almost without saying. Luckily, there is no need here to concern us with the various approaches which were tried to show that they were wrong (see Van Everbroeck 1994 for a summary of this discussion). Generally speaking, however, one of the major boons of the resurrection of neural network research in Cognitive Science has been that the debate about the fundamental nature of human cognition has been reopened. It is no longer self-evident to assume that concepts are stored in neat hierarchical networks with labeled branches, or that thoughts should be represented by means of structured sentential propositions. Nonetheless, connectionists still have to prove that their models can indeed account for precisely those facts about thought which led Fodor & Pylyshyn to believe that only a classical theory could explain them. Though the explanations offered by connectionists need not be fully commensurable with their classical counterparts, they should at least address approximately the same data.

It is at this point that CLASPnet can make a modest contribution to the task connectionists face: Fodor & Pylyshyn repeatedly referred to natural language as being analogous to cognition in that language too was claimed to be fully productive, systematic, and compositional. Hence, language would need a classical theory as well, because a non-classical theory would never be able to account for the linguistic data. One of the relevant aspects of natural language in this respect is how properties of phrases, clauses, sentences, paragraphs, and texts could be deduced from the words which make up these higher-level constructs. In a classical theory, these constructs would need to have explicit access to their constituent parts. But in a distributed connectionist model like CLASPnet, these constituent words are not represented explicitly in the memory of the model (Van Gelder 1990, 1992). As a consequence, the model offers evidence for the position that natural language is not 'classical' (cf. Dascal 1992; Sloman 1993). And if that conclusion is true, finally, the argumentation for a classical theory of cognition is weakened considerably (but see Pinker 1992).