Potts-glass model of layered feedforward neural networks

The layered feedforward neural network is extended to a q-state Potts-glass model. The Potts-glass version of the network is realized by imposing local inhibition on a group of Ising spins and introducing competitive updating rules on them. The dynamics of such a system is solved exactly, and the storage capacity of the network is found to be proportional to ${\mathit{q}}^{\mathrm{\Delta }}$, with Δ≊1.85 in the case of storing unbiased patterns. For biased patterns, we obtain the phase diagram for q=3 as a function of the storage capacity and the bias parameters, which indicates that the storage capacity decreases with the bias.