Relationships between internal representation and generalization ability in multi layered neural network for binary pattern classification problem

This paper studies relationships between the internal representation and the generalization ability in multi layered neural networks for binary pattern classification problems. Three indices are newly introduced to characterize the internal representation, which represent the derivatives of hidden units, the similarity of learning patterns, and the activity of hidden units. These indices are applied to analyze the internal representation in neural networks for binary pattern classification problems, and it is shown that they are closely connected with the generalization ability.