Improving generalization ability of multilayer networks by excluding irrelevant input components

We propose a learning method to improve generalization ability of neural networks for pattern recognition in the case that a priori knowledge about training targets is obtained. As a priori knowledge, we use a linear subspace in pattern space that can be regarded as irrelevant to recognition. By reflecting such knowledge on weight representation, we try to improve the generalization ability. The knowledge about the subspace is introduced as linear constraints on weight representation. Finally, we verify the effectiveness of our method by experiments. In the experiments, the subspace that can be regarded as irrelevant to recognition is determined statistically by using discriminant analysis