関数近似問題に対する階層型ニューラルネットワークの内部表現と汎化能力の関係

概要

論文の詳細を見る
This paper discusses the relation between internal representations and generalization ability of multi-layered neural networks for function approximation problems. Here, the over learning problem, which makes the generalization ability poor, is divided into the two problems; the excessive degrees of freedom and the non-uniqueness of weights. These problems are carefully discussed from the viewpoint of internal representation; weights and outputs of hidden units. First, the excessive degrees of freedom problem is discussed by introducing the entropy for weights. Next, the non-uniqueness of weights problem is also discussed by introducing the principal components analysis method. Finally, the relations between internal representations and generalization ability of multi-layered neural networks are concretely discussed by numerical results.
福山大学の論文