FINITE-SIZE-SCALING IN NEURAL NETWORKS

We demonstrate that the fraction of pattern sets that can be stored in single- and hidden-layer perceptrons exhibits finite size scaling. Th is feature allows one to estimate the critical storage capacity a, fro m simulations of relatively small systems. We illustrate this approach by determining a,, together with the finite size scaling exponent nu, for storing Gaussian patterns in committee and parity machines with b inary couplings and up to K = 5 hidden units.