Incremental Extreme Learning Machine via Fast Random Search Method

摘要

Since extreme learning machine (ELM) was proposed, it has been found that some hidden nodes in ELM may play a very minor role in the network output. To avoid this problem, enhanced random search based incremental extreme learning machine (EI-ELM) is proposed. However, we find that the EI-ELM's training time is too long. In addition, EI-ELM can only add hidden nodes one by one. This paper proposes a fast method for EI-ELM (referred to as FI-ELM). At each learning step, several hidden nodes are randomly generated and the hidden nodes selected by the multiresponse sparse regression (MRSR) are added to the existing network. The output weights of the network are updated by a fast iterative method. The experimental results show that compared with EI-ELM, FI-ELM spends less time on training. Taking this advantage, FI-ELM can generate more hidden nodes to find the hidden node leading to larger residual error decreasing.