Noise Resistant Training for Extreme Learning Machine.

The extreme learning machine (ELM) concept provides some effective training algorithms to construct single hidden layer feedforward networks (SHLFNs). However, the conventional ELM algorithms were designed for the noiseless situation only, in which the outputs of the hidden nodes are not contaminated by noise. This paper presents two noise-resistant training algorithms, namely noise-resistant incremental ELM (NRI-ELM) and noise-resistant convex incremental ELM (NRCI-ELM). For NRI-ELM, its noise-resistant ability is better than that of the conventional incremented ELM algorithms. To further enhance the noise resistant ability, the NRCI-ELM algorithm is proposed. The convergent properties of the two proposed noise resistant algorithms are also presented.