Critical storage capacity of the J = ± 1 neural network

Abstract

For neural networks in which the coupling Jij are allowed to take on the values Jij = 1 or Jij = -1, the authors determine numerically the critical storage capacity for random unbiased patterns as a function of the stability. They use an exact enumeration scheme based on the Gray code and a continuous distribution for the patterns to control finite-size effects. Results are presented for N ≤ 25; they indicate an optimal storage capacity of αc ≈ 0.82 (N → ∞).