Regression Neural Networks with a Highly Robust Loss Function

Abstract

Artificial neural networks represent an important class of methods for fitting nonlinear regression to data with an unknown regression function. However, usual ways of training of the most common types of neural networks applied to nonlinear regression tasks suffer from the presence of outlying measurements (outliers) in the data. So far, only a few robust alternatives for training common forms of neural networks have been proposed. In this work, we robustify two common types of neural networks by considering robust versions of their loss functions, which have turned out to be successful in linear regression. Particularly, we extend the idea of using the loss of the least trimmed squares estimator to radial basis function networks. We also propose multilayer perceptrons and radial basis function networks based on the loss of the least weighted squares estimator. The performance of these novel methods is compared with that of standard neural networks on 4 datasets. The results bring arguments in favor of the novel robust approach based on the least weighted squares estimator with trimmed linear weights in terms of yielding the smallest robust prediction error in a variety of situations. Robust neural networks are even able to outperform the prediction ability of support vector regression.