Noise-tolerant incremental learning for extreme learning machine

Abstract

The studies in extreme learning machine (ELM) allow us to build a simple single-hidden-layer feedforward network (SLFN) in an efficient way. However, the existing incremental ELM algorithms in those studies are proposed based on the noiseless situation, in which assumes that the outputs of the hidden nodes are not affected by any noise. It is inevitable that node noise exists in real implementation. In this project, therefore, I am going to propose noise-tolerance incremental ELM (NT-IELM) and noise-tolerance convex incremental ELM (NT-CIELM) for SLFNs. In incremental learning, the hidden nodes are added into a SLFN in the one-by-one manner. For the NT-IELM algorithm, it estimates the output weight of newly-inserted node and keeping the previously-learned output weights unchanged. It has better noise-tolerant ability than the two existing incremental ELM algorithms. To step up the noise-tolerant ability, the NT-CIELM algorithm is then proposed by estimating the output weight of newly-added hidden node as well as using a simple rule to update the previously-learned output weights. I experimentally prove that the two proposed algorithms converge. Simulation shows that my two proposed noise-tolerant algorithms have better noise-tolerant ability, compared to the existing incremental ELM algorithms.