Properties and learning algorithm of discrete neural network with time delay

A mutually connected neural network of discrete type with time delay is considered. First, the convergent states of delay networks are analyzed. It is found that a normal network with only one step delay can be constructed so as to have the same fixed-point states as a given delay network, and that a binary (0/1) neural network with m and n step delays has oscillation of period (m+n ) under the condition that all the connections are inhibitory and symmetrical. Next, the learning problem of the delay network is considered. A learning algorithm based on a modified Hebb´s rule is proposed for making a delay network adapt to periodic external input patterns. Simulation results show that a delay network can memorize a periodic sequence of 2D patterns using this rule