Training neural net classifier to improve generalization capability
Masahiro Kayama, Shigeo Abe · Systems and Computers in Japan · 1994
Abstract A training method for neural net classifiers is discussed from the viewpoint of improving their generalization capability. First, the conventional training method which minimizes the square sum of network output errors from training outputs is shown to be inappropriate. This is because category boundaries may be close to some specific clusters which decreases the generalization capability of the network. Then to obtain impartial boundaries to all clusters, a new method is proposed which adds appropriate random numbers to the training inputs and decreases their amplitude to zero as training proceeds. The effectiveness of this method is demonstrated by simulations of alphabet and practical number recognition systems. The proposed method is useful, especially when the quality and quantity of training data are not sufficient. This is particular for cases where a classification system should be constructed quickly with a few representative training data or considerable time or money is required to obtain voluminous training data.