We consider problems of sequence processing and propose a solution based on a discrete-
state model in order to represent past context. We introduce a recurrent connectionist …

The extraction of symbolic knowledge from trained neural networks and the direct encoding
of (partial) knowledge into networks prior to training are important issues. They allow the …

Recurrent neural networks that are trained to behave like deterministic finite-state automata
(DFAs) can show deteriorating performance when tested on long strings. This deteriorating …

Recurrent neural networks readily process, recognize and generate temporal sequences. By
encoding grammatical strings as temporal sequences, recurrent neural networks can be …

This dissertation shows that the tractability and efficiency of training particular connectionist
networks to implement certain classes of grammars can be formally determined by applying …