Indirect convergence neural network associative memory

Abstract

A simple but robust neural network associative memory that utilizes indirect convergence is described. The definition of indirect convergence is that during the synchronous iterative recall process, every neuron state update must be in the right direction, i.e., no wandering transition is allowed. Since it is based on the right-direction-only convergence mechanism, the number of iterative update steps required to converge to a stored state is decreased at the expense of the storage capacity. The use of a simple parametric method to characterize the indirect convergence net is explored. The tradeoff between the number of stored states and their attraction force is analyzed. The major advantage of such network is its rapidity in seeking for the stored state. The generalized higher-order version of this indirect convergence network is also discussed.< >