On the unlearning procedure yielding a high-performance associative memory neural network
J. A. Horas, Pedro Marcelo Pasinetti · Journal of Physics A Mathematical and General · 1998
We consider a fully connected Hopfield-like neural network as a set of N independent perceptrons. We trained these perceptrons using the so-called inverse perceptron rule, obtaining a matrix of synaptic couplings, that make a number of spurious states unstable . We numerically determine the optimum number of spurious states, obtained by random shooting, that must be destabilized in order to obtain an improvement in performance. The unlearning procedure generated, is shown to be able to give a high-performance associative memory characterized by: (1) an enhancement in storaging capacity; (2) an enlargement in the size of attraction basins; (3) a reduction in the number of spurious attractors and (4) a reliable and fast retrieval.