Random Orthogonal Projection Based Enhanced Bidirectional Extreme Learning Machine

Abstract

Bidirectional extreme learning machine (B-ELM) divides the learning process into two parts: At odd learning step, the parameters of the new hidden node are generated randomly, while at even learning step, the parameters of the new hidden node are obtained analytically from the parameters of the former node. However, some of the odd-hidden nodes play a minor role, which will have a negative impact on the even-hidden nodes, and result in a sharp rise in the network complexity. To avoid this issue, we propose a random orthogonal projection based enhanced bidirectional extreme learning machine algorithm (OEB-ELM). In OEB-ELM, several orthogonal candidate nodes are generated randomly at each odd learning step, only the node with the largest residual error reduction will be added to the existing network. Experiments on six real datasets have shown that the OEB-ELM has better generalization performance and stability than B-ELM, EB-ELM, and EI-ELM algorithms.