Dynamics of a Neural Network Model of Associative Memory

Introduction The associative memory is one of the fundamental algorithms of information processing of the brain, which is characterized by such features as "learning through examples," "memory formation by Hebbian rule," and "ability of generalization. " Since J. Hopfield proposed the autocorrelation associative memory and applied the analogy to the spin glass model in 1982, theoretical studies and simulations of neural network models of the associative memory revived and it was made clear that the memory capacity, the basin of attraction (the stability) of each memory pattern and the spurious memory are main problems. The Hopfield's original model uses the autocorrelation matrix and neurons with signum output function. The model is simple but has flaws in the above problems. Several models which improves these flaws were proposed. Most of them replace the simple autocorrelation matrix by more sophisticated structures. We have proposed a model which uses neurons with nonmonot