Training Spatially Homogeneous Fully Recurrent Neural Networks in Eigenvalue Space.

Abstract

A new design method for spatially-homogeneous, fully recurrent neural networks is presented. In our approach the eigenvalues of the synaptic matrix, rather than the weights, are learned from the examples. When the learning process is carried out, the connection weights are easily computed from the eigenvalues by inverse discrete Fourier transform. The adaptation is performed in the eigenvalue space in order to simply incorporate in the training algorithm the conditions for the uniqueness of the steady-state. As a consequence, the trained networks are insensitive to initial conditions. The method is illustrated by computer simulations concerning two specific feature extraction examples. Copyright 1996 Elsevier Science Ltd.