Neural networks with long-range feedback: design for stable dynamics

Abstract

Feedback in neural networks is essential. Without it, true dynamics would be lacking. For this reason, many well known models include feedback connections (e.g. Hopfield, ART, neocognitron). Neural networks with feedback are, however, likely to be unstable if not carefully designed. In this paper, we show how to incorporate long-range feedback in a class of dynamically stable nonlinear neural networks.