Training, dynamics, and complexity of architecture-specific recurrent neural networks
Jacques Ludik · SUNScholar (Stellenbosch University) · 1994
ENGLISH ABSTRACT: This dissertation describes the main results of a pioneering effort to develop novel architectures, training strategies, dynamics analysis techniques and theoretical complexity results for architecure-specific reccurent neural network (ASRNNs). To put the study of ASRNNs into an appropriate perspective, a temporal processing framework that describes the different neural network approaches taken, was constructed. ASRNNs are more powerful than non-recurrent networks and computationally less expensive, more stable, and easier to study than general-purpose recurrent networks. The focus was on Elman, Jordan, and Temporal Autoassociation ASRNNs using discrete-time backpropagation.