Validation of neural networks using hybrid resampling methods

Investigates hybrid approaches of statistical resampling methodologies to neural network validation when the data is sparse. Specifically, cross validation, grouped cross validation, bootstrapping and resubstitution are considered. Resampling methods have been used occasionally in the literature for neural network validation when the amount of data is limited. This research paper uses simulated data to examine different resampling estimates first, then applies stepwise regression to identify different hybrid estimates. The best hybrid estimate, which combines the resampling estimates obtained from twenty bootstrapped samples, five bootstrapped samples and two-fold cross validation, is nearly unbiased with less variability. This hybrid requires constructing twenty-seven validation networks