Hybrid hierarchical extreme learning machine

Abstract

Restricted by the shallow structure of Extreme Learning Machine(ELM), the ideal fitting effect can not be achieved even if large hidden nodes are set. In order to obtain better feature representation and classification performance, this paper proposes a Hybrid Hierarchical Extreme Learning Machine (HH-ELM) on the hierarchical thought of Hierarchical Extreme Learning Machine(H-ELM). The feature extraction part uses ELM-Based Auto-Encoder(ELM-AE) based on L1-norm regularization to optimize the hidden layer weights, and the classification part adopts Improved Tow-hidden-layer Extreme Learning Machine(ITELM). Experimental results on UCI datasets and Mnist images datasets show that HH-ELM has better classification results and robustness.