Assessing and Improving Neural Network Predictions by the Bootstrap Algorithm

Abstract

The bootstrap algorithm is a computational intensive procedure to derive nonparametric confidence intervals of statistical estimators in situations where an analytic solution is intractable. It is ap-plied to neural networks to estimate the predictive distribution for unseen inputs. The consistency of different bootstrap procedures and their convergence speed is discussed. A small scale simulation experiment shows the applicability of the bootstrap to practical problems and its potential use. 1