GENERALIZATION IN A MULTISTATE NEURAL-NETWORK

The generalization ability of an extremely dilute feedback neural netw ork with multistate neurons is studied by means of a deterministic noi seless parallel dynamics. The overlap with any one of a macroscopic nu mber of binary, full activity, concepts is determined when the network is trained with examples of variable activity according to a Hebbian learning algorithm that favours stable symmetric mixture states. Expli cit results about the phase diagram and the generalization error are o btained for a network with three-state neurons which remain inactive b elow a threshold theta. It is shown that the generalization ability ca n be considerably enhanced either by training the network with low-act ivity examples or by means of a moderate increase in theta.