DETERMINATION OF DOMAIN OF ATTRACTION BY SHORT-TIME OVERLAP DYNAMICS IN NEURAL NETWORKS

Since the overlap dynamics of spin-glass-like neural network is solely governed by such two parameters, the symmetry of the net's connections and the stabilities of the embedded patterns, we propose that the domain of attraction of a network could be determined by a short-time (at least the first two-step) dynamics, in which both parameters begin to be visible. The first step overlap dynamics is calculated by probability theory and the second step overlap dynamics for pseudo-inverse-model is obtained by extended Krauth's theory. The calculated domain of attraction for this model is in good agreement with the results obtained by computer simulation.