A new method in obtaining a better generalization in artificial neural networks

Abstract

Overtraining is a serious problem in the neural network algorithms, including the backpropagation algorithm. In order to measure the performance of a neural network, ordinarily some of the data is sacrificed and used as a test set (cross-validation method). When the data is very scarce or is expensive, e.g. medical applications such as computer aided diagnosis, this waste of the data becomes intolerable. A new technique is introduced which uses the shape of the training mean squared error graph versus number of epochs and predicts when is the best time (epoch number) to discontinue the training.