Noise Resistant Training for Incremental Extreme Learning Machine

Abstract

For the single hidden layer feedforward network (SHLFN) model, an incremental extreme learning machine (I-ELM) algorithm was proposed to provide a faster learning speed. In I-ELM algorithm, hidden nodes are randomly generated and only the output weights for the newly inserted hidden nodes are estimated and adjusted. It has been proved in theory that the network constructed by I-ELM can used to approximate all continues function perfectly under a noiseless situation. However, node noise is unavoidable in real applications and the existence of node noise will poorly affect the performance of the training algorithms. This paper proposes a noise resistant algorithm for I-ELM named noise resistant group I-ELM (NRGI-ELM). In NRGI-ELM algorithm, a number of randomly generated hidden nodes are inserted to the network at every training iteration and a noise resistant method is used to determine the output weights of the newly inserted hidden nodes. The stimulation results show that NRGI-ELM has a better mean squared error performance when more hidden nodes are inserted to the training SHLFN at every training iteration and NRGI-ELM algorithm is more resistant to node noise when comparing with the original GI-ELM algorithm.