GENERALIZATION OF BACKPROPAGATION TO RECURRENT AND HIGH-ORDER NETWORKS.

Summary form only given. A general method for deriving back-propagation algorithms for networks with feedback and higher order connectivity is introduced. As an example, this method is used to derive an algorithm for adaptively modifying the weights of a recurrent generalization of the network introduced by Rumelhart et al. The back-propagation involves the integration of an associated differential equation. The weight matrix is updated only after the forward-propagation, back-propagation, and differential equations have reached steady state.