Bootstrap for neural model selection

  Request Code

  API Access   Call/Text an Expert

Bootstrap techniques (also called resampling computation techniques) have introduced new advances in modeling and model evaluation. Using resampling methods to construct a series of new samples which are based on the original data set, allows to estimate the stability of the parameters. Properties such as convergence and asymptotic normality can be checked for any particular observed data set. In most cases, the statistics computed on the generated data sets give a good idea of the confidence regions of the estimates. In this paper, we debate on the contribution of such methods for model selection, in the case of feedforward neural networks. The method is described and compared with the leave-one-out resampling method. The effectiveness of the bootstrap method, versus the leave-one-out methode, is checked through a number of examples.