ABSTRACT
Introduction
To what extent can the behavior of a neural network performing a classification task, for example the balance scale, be called rule based? On the one hand, obviously, explicit rules are not available, although the responses of a network might be consistent with rules. On the other hand, networks are not thought to learn fully in accordance with behaviorist theories and are believed to have particular cognitive processing properties such as selective encoding of input patterns. The hidden units should function as mediating concepts. However, it is impossible to study the internal state of the human brain in the same detail as the internal structure of a neural network. We propose a simple empirical test for neural networks that discriminates between the formation of stimulus-driven associations and the formation of cognitive concepts: the discrimination-shift task (Kendler, 1995).
