ORDER-PARAMETER EVOLUTION IN A FEEDFORWARD NEURAL-NETWORK

We consider layered neural networks in which the weights are trained w ith the pseudoinverse rule to store a set of random patterns. Using ma ny-body diagrammatic techniques, the evolution in the network can be d escribed by the overlap order parameter m and the noise parameter Delt a. Looping effects are shown to be significant, in contrast to a previ ous conjecture. Order parameter pairs corresponding to various input c onditions are found to collapse on a universal curve.