ABSTRACT
This chapter focuses on G. F. Marcus et al ’s experiments, showing how an associationist device, a particular neural network architecture, can learn the patterns in the experiments, generalizing to novel sequences, and how the account, rather than being simply an uninteresting implementation of a symbolic model, makes novel predictions about the learning of sequences. A neural network needs two features to generalize over “grammatical” patterns like those in Marcus et al ’s experiments. Playpen is a neural network architecture which is designed to represent and learn relational knowledge and to deal with simple sequential patterns. Micw-relation units (MRU) in the network represented possible binary relations between the syllables in a sequence, and connections between the MRUs represented correlations between the simple binary relations. In the network the similarity of the test sequences to the training sequences had an effect on generalization.
