Choosing the optimal neural network based on Bayesian considerations

Abstract

In this work, we aim to apply the Bayesian estimate to neural networks in order to select the model that differs from the a posteriori estimates. best suited to the teaching data. All of this requires the computation of a multidimensional integral, which is a difficult task even with traditional Monte Carlo methods. For this purpose, the Nested Sampling algorithm is used, and the products of the calculations are obtained by wandering in the field of hyperparameters. We further show how gradient propagation and accidental wandering can be obtained by obtaining a hybrid teaching technique.