Temporal processing with connectionist networks

Abstract

A collection of tasks is proposed for evaluating temporal neural network algorithms. Within this framework two procedures, a novel learning algorithm and an algorithm for generating temporal representations, are considered. The internal target generation learning algorithm for recurrent networks is designed for overcoming the problem of sparse targets in a temporal task. The temporal autoassociation representation of temporal sequences is designed to retain sequential order information in a recurrent network. On a simple benchmark it is shown to significantly improve convergence times over simple recurrent networks. Both the algorithm and the representation help bridge the gap between inputs and delayed targets that makes many temporal problems difficult. >