Dynamics of associative memory with a self-consistent noise

Abstract

The Glauber dynamics of magnetic systems has been extended to the case of neural networks with a general odd response function. The author derives a set of recursion relations for the overlap parameter, noise average and noise variance taken as macrovariables of the process describing the dynamics of associative memory. The retrieval process is studied then for a hyperbolic tangent transfer function by the self-consistent signal to noise ratio method. It takes into account the fatigue effect of the real neuron. The phase diagrams of the retrieval process reveals an enhanced storage capacity for a certain set of parameter values.