CONVERGENCE-RATES IN SYMMETRICAL NEURAL NETWORKS WITH GLAUBER DYNAMICS

In numerical experiments with randomly-generated, symmetrically-interconnected networks of N McCulloch-Pitts neurons, the median convergence time (t50) is found to increase with size roughly according to the relation [GRAPHICS] of the asynchronous (Glauber) dynamics. An exact treatment of the stochastic model has yet to appear, though some reports seem to confirm Hopfield's casual observation, that the convergence time is roughly four sweeps (irrespective of N).