Estimating regional noise on neural network predictions

摘要

A new method for estimating the variance of noise for nonlinear regression is presented. The noise is modelled to be regional, i.e. its variance depends on the input, and it consists of two sources: measurement errors and inherent noise of the underlying function. Our approach consists of two neural networks using Bayesian methods, which are trained in sequence. It is orientated by the assumption of unbiased predictions of the mean and the confidence of network prognoses, which are used to predict the variance of noise. We demonstrate our approach on two toy and one real data sets.