Avoiding Weight-Illgrowth to Improve Generalization : Cascade Correlation Algorithm with Local Regularization

概要

論文の詳細を見る
Aiming at the problem of Neural Networks (NNs) design that if only adopting static learning algorithm (e.g., BP algorithm), one must determine the optimal NN toplogy prior to training process by the method of trial-and-error. Dynamic NNs have been attractive, which incorporate the determination of topologies into the training process, mainly in two ways: destructive (pruning) technique and constructive (growing) technique. By comparing them, we cann't assert rashly which is better. In fact, combining these two methods will be more appropriate. For example, it is nature to incorporate the pruning technology into the growing method in order to overcome overgrowth. Accordingly, in this paper, we will concretely analyze the weight-illgrowth action of cascade correlation algorithm (CCA), which is one of representative constructive algorithms, and use regularization to avoid this phenomenon for improving generalization.
社団法人電子情報通信学会の論文
1997-03-06