Order parameter evolution in a feedforward neural network
K. Y. Michael Wong, C Campbell, David C. Sherrington · Journal of Physics A Mathematical and General · 1995
We consider layered neural networks in which the weights are trained with the pseudoinverse rule to store a set of random patterns. Using many-body diagrammatic techniques, the evolution in the network can be described by the overlap order parameter m and the noise parameter Delta . Looping effects are shown to be significant, in contrast to a previous conjecture. Order parameter pairs corresponding to various input conditions are found to collapse on a universal curve.