Generalization in an analog neural network
Daniel A. Stariolo, Francisco A. Tamarit · Physical Review A · 1992
We analyze the generalization ability of an iterated-map neural network when an extensive number of patterns is stored through a Hebbian learning mechanism. We show that the model is able to create a concept representative of a set of correlated patterns if a critical minimum number of patterns is presented. This critical number depends on the correlation among the patterns, the slope of the transfer function at the origin, and the ratio between the number of memories and the total number of neurons.