Representational Capabilities of Multilayer Feedforward Networks with Time-Delay Synapses

Modelling time-dependent nonlinear systems is a topic of growing interest in neural networks. Promising results have been obtained for the capabilities of recurrent networks, and time-delay networks, but few results have been obtained for the theoretical capabilities of these structures. A global-feedforward local-recurrent network architecture was proposed recently which was demonstrated to have better modelling performance than a global-feedforward local-feedforward network. In this paper the global-feedforward local-recurrent network is analysed, and theoretical proofs are given for its representational capabilities.