Finding the bias-variance tradeoff during neural network training and its implication on structure selection

Neural network overtraining in training is a common problem still requiring a lot of attention In order to solve the problem of overtraining this paper proposes a new method to find the bias-variance tradeoff using a bootstrap estimate of the expected prediction risk that is calculated during training. The relation of this method with regularization and its implication on model structure selection is discussed. Finally, some experiments are discussed which show the applicability of the proposed method to the model structure selection problem.