A study on artificial neural network generalization capability

A learning model called the mutual interaction model (MIM) for artificial neural networks is proposed. In the model, the pattern distribution information is taken into account to obtain good generalization. Experiments were carried out for two kinds of artificial pattern distributions. Test data recognition rates were compared with those obtained by backpropagation (BP). On the average, the pattern distributions on the output layer obtained by MIM learning are more compact than those obtained by BP learning. This indicates that the network generalization capability is improved by MIM learning as compared with BP learning, though a problem remains regarding dependence on the initial network parameter assignment