THE EFFECTS OF LESIONS ON THE GENERALIZATION ABILITY OF A PERCEPTRON

We investigate the effects of dilution (lesions) on the memorization a nd generalization abilities of a single-layer perceptron whose non-zer o weights are constrained to take on binary (+/-1) values only. The di luted perceptron is trained to realize a Boolean linearly separable ma pping generated by a fully connected perceptron. In the case where the training process is disturbed by noise and the vanishing weights are chosen so as to minimize the training error, we find that the dilution can improve the storage capacity and the generalization ability of th e network. If the weights are cut randomly, however, the dilution will always degrade the network's performance. In this case we show that t he main effect of dilution is to introduce an effective noise in the t raining examples.