ABSTRACT
This chapter reviews the theory, implementation and application of the recurrent back-propagation(RBP) algorithm for convergent dynamical systems. The additive model is used as the canonical example, although general expressions for arbitrary convergent dynamical systems are given in the appendix. Forcing techniques for creating fixed points are discussed from a dynamical point of view. The general features of a physical implementation of a collective dynamical system that performs the RBP algorithm are discussed. The results of three simulations are presented. The first shows how learning curves exhibit discontinuities as fixed points are introduced. The second shows how recurrent networks can be trained to perform error correction and autoassociative recall. The last shows how RBP has been used to solve the correspondence problem in random-dot stereograms.
