The Characteristics of the Convergence Time of Associative Neural Networks

Abstract

The authors have analyzed the dynamics of associative neural networks based on macroscopic state equations and have shown that both a layered associative net and an autocorrelation type net have the same convergence property: If a recalling process succeeds, the network converges very fast to one of the memorized patterns. But if a recalling process fails, it converges very slowly to a spurious state or does not converge. This property was also checked by computer simulations on a large scale (N = 1000) neural network. Moreover, it is shown that the convergence time for a successful recall is of order log(N). If this convergence time difference is used, execution time and memory can be saved and it can be determined whether a recalling process succeeds or fails without any additional procedure.