Robust extreme learning machine in the presence of outliers by iterative reweighted algorithm

Abstract

Extreme learning machine (ELM) is widely used to derive the single-hidden layer feedforward neural networks. However, ELM faces a great challenge in the presence of outliers, which can result in the sensitivity and poor robustness. To overcome this dilemma, a non-convex 2-norm loss function is developed to reduce these negative influences by setting a fixed penalty on any potential outliers. A novel robust ELM is proposed in this paper, and the resultant optimization can be implemented by an iterative reweighted algorithm, called IRRELM. In each iteration, IRRELM solves a weighted ELM. Several artificial datasets, real-world datasets and financial time series datasets are employed in numerical experiments, which demonstrate that IRRELM has superior generalization performance and robustness for modeling datasets in the presence of outliers, especially at the higher outlier levels.