A new associative memory which inhibits a meaningless output

A hierarchial associative memory is important in realizing hierarchical pattern recognition, pattern inference, and information retrieval. The inhibition of a meaningless output plays an essential role in stopping the spread of meaningless activation in such a network. An associative memory with a dynamical threshold to solve this problem is proposed. A macroscopic state equation is obtained, and the dynamics of the network are analyzed based on the equation. Computer simulation shows that the network converges to either a zero pattern (a failed recall) or one of the memorized patterns (a successful recall), and does not generate a meaningless output. The basin of attraction is comparable to that of a conventional model if the threshold is properly scheduled in time. It is shown that the convergence time of failed recall and the basin of attraction is controllable by the threshold schedule. >