Self-organization in stochastic neural networks

The maximization of the mutual information between the stochastic outputs neurons and the clamped inputs is used as an unsupervised criterion for training a Boltzmann machine. The resulting learning rule contains two terms corresponding to the Hebbian and anti-Hebbian learning. The two terms are weighted by the amount of transmitted information in the learning synapse, giving an information-theoretic interpretation to the proportionality constant given in the biological rule of Hebb. The anti-Hebbian term causes the convergence of weights. Simulation for the encoder problem demonstrates optimal performance of this method.