MEMORIZING AND REGENERATING SPATIOTEMPORAL PATTERNS WITH A STRUCTUREDRECURRENT NEURAL-NETWORK

A local connected recurrent neural network with the ability to memoriz e and regenerate some nonlinear complex dynamics is proposed in this p aper and a new learning algorithm for this network is also developed. The network, constructed of adaptive oscillating modules, is easily ap plied to the realization of nonlinear dynamics. The module consists of two simple neuron nodes with recurrent connections. The new learning algorithm can independently train each module with suitable speed for given input data. The network size is also adaptively determined durin g the learning process. This network also has a suitable structure of a parallel VLSI system. Finally, some simulation results are given to verify the effectiveness on the proposed network structure and the lea rning algorithm.