Bootstrap re-sampling and cross-validation for neural network learning

A technical framework to assess the impact of re-sampling on the ability of a neural network is presented to correctly learn a classification problem.We use the bootstrap expression of the prediction error to identify the optimal re-sampling proportions in a numerical experiment with binary classes and propose a new,simple method to estimate this optimal proportion.An upper and a lower bounds for the optimal proportion are derived based on Bayes decision rule.The analytical considerations to extend the present method to crossvalidation are also illustrated.Includes bibliographic reference