ABSTRACT
Ever since the birth of connectionism, it has had to fight for survival. In 1969 Minsky and Papert [1] argued that connectionist models suffered from a severe limitation in their mathematical foundation, which was not likely to be overcome, even if several layers of network-units were used. As we all know, this limitation was overcome by the introduction of the backpropagation algorithm and connectionism was revitalised. In 1988 a new attack was launched against connectionism (cf. Fodor & Pylyshyn [2]; Fodor & McLaughlin [3]). This time the claim was that connectionist models could not be regarded as models explaining cognition. They could, at best, be implementations of classical models. The connectionist defence has basically taken two forms; theoretical argumentation (cf. Smolensky [4]; Chalmers [5]; van Gelder [6], [21]; van Gelder & Niklasson (forthcoming)) and practical modelling (cf. Pollack [7]; Chalmers [8]; Smolensky [9]; Niklasson & Sharkey [10], [11], [12]; Blank, Meeden and Marshall [13]).
