A comparative study of pseudo-inverse computing for the extreme learning machine classifier

Most feed-forward artificial neural network training algorithms for classification problems are based on an iterative steepest descent technique. Their well-known drawback is slow convergence. A fast solution is an Extreme Learning Machine (ELM) computing the Moore-Penrose inverse using SVD. However, the most significant training time is pseudo-inverse computing. Thus, this paper proposes two fast solutions to pseudo-inverse computing based on QR with pivoting and Fast General Inverse algorithms. They are QR-ELM and GENINV-ELM, respectively. The benchmarks are conducted on 5 standard classification problems, i.e., diabetes, satellite images, image segmentation, forest cover type and sensit vehicle (combined) problems. The experimental results clearly showed that both QR-ELM and GENINV-ELM can speed up the training time of ELM and the quality of their solutions can be compared to that of the original ELM. They also show that QR-ELM is more robust than GENINV-ELM.