Improvement of generalization for a perceptron with localized representation

This paper intends to solve problems in the BP learning in the conventional multilayered perceptron, and proposes two methods, the local generalization and the global generalization, improve the generalization ability of the perceptron with the localized representation. In the local generalization, the generalization learning is executed so that the generalization ability function based on the distance between the separating partial hyperplanes and the set of patterns for learning, is maximized under the given order of separation in the learning. A problem then is that the range of adjustment for the separating partial hyperplane is limited depending on the order of separations, which prevents the improvement of the generalization ability. In the global generalization, the generalization ability function is defined considering the order of category separation and the optimal separating partial hyperplanes are determined simultaneously to maximize the function. Simulation experiments are carried out for the learning task on the two-dimensional plane and the recognition task for the handwritten Chinese characters. It is seen that the generalization ability is improved greatly. It is seen that the local generalization can realize nearly the same generalization ability, and the global generalization can realize a better generalization ability, compared to the BP learning.