Robust Neural Networks Learning: New Approaches

摘要

The paper suggests an extended version of principle of empirical risk minimization and principle of smoothly winsorized sums minimization for robust neural networks learning. It's based on using of Maveraging functions instead of the arithmetic mean for empirical risk estimation (M-risk). Theese approaches generalize robust algorithms based on using median and quantiles for estimation of mean losses. An iteratively reweighted schema for minimization of M-risk is proposed. This schema allows to use weighted version of traditional back propagation algorithms for neural networks learning in presence of outliers.