Incorporation of Long-Range Feedback in Neural Networks Under Stability Conditions.

Feedback endows neural networks with several interesting properties. It is thus not surprising that several well-known models (e.g. ART, Hopfield, neocognitron) include feedback connections. However, neural networks with feedback may possess unstable dynamics and should be carefully designed. In this paper we show how to incorporate long-range feedback in a broad class of dynamically stable neural networks using the basic idea of symmetric connections. The case of networks with binary inputs and binary outputs is treated first. Then, as the main contribution of this paper, the analysis is extended to networks with analog (continuous-time continuous-output) neurons.