GLOBAL ANALYSIS OF PARALLEL ANALOG NETWORKS WITH RETARDED FEEDBACK

We analyze the retrieval dynamics of analog "neural" networks with clocked sigmoid elements and multiple signal delays. Proving a conjecture by Marcus and Westervelt, we show that for delay-independent symmetric coupling strengths, the only attractors are fixed points and periodic limit cycles. The same result applies to a larger class of asymmetric networks that may be utilized to store temporal associations with a cyclic structure. We discuss implications for various learning schemes in the space-time domain.