Neural networks with biased bipolar synapses and biased patterns
Raju Viswanathan · Journal of Physics A Mathematical and General · 1993
A neural network model with bipolar synaptic couplings is studied with a global constraint on the synapses which earn neuron receives, corresponding to a bias of the couplings. When the network is used to store biased patterns, it is shown that the maximal storage capacity for any non-zero bias of the stored patterns corresponds to a particular type of the connectivity pattern of the network as encoded in the constraint. The 'space of interactions' approach yields the dependence of the storage capacity on the distribution of the couplings for every pattern bias. It turns out that the optimal coupling bias is non-zero and independent of the bias in the patterns.