On improving learning capability of ELM and an application to brain-computer interface

TL;DR

This paper enhances the learning capability of Extreme Learning Machine (ELM) by replacing SVD with more efficient matrix decomposition methods, improving performance and speed in brain-computer interface applications.

Contribution

It introduces and compares five alternative matrix decomposition techniques to SVD for ELM, demonstrating their effectiveness in large-scale real-world data.

Findings

Hessenberg decomposition speeds up training when speed is prioritized.

Householder reflection improves classification performance.

Different methods suit different application priorities.

Abstract

As a type of pseudoinverse learning, extreme learning machine (ELM) is able to achieve high performances in a rapid pace on benchmark datasets. However, when it is applied to real life large data, decline related to low-convergence of singular value decomposition (SVD) method occurs. Our study aims to resolve this issue via replacing SVD with theoretically and empirically much efficient 5 number of methods: lower upper triangularization, Hessenberg decomposition, Schur decomposition, modified Gram Schmidt algorithm and Householder reflection. Comparisons were made on electroencephalography based brain-computer interface classification problem to decide which method is the most useful. Results of subject-based classifications suggested that if priority was given to training pace, Hessenberg decomposition method, whereas if priority was given to performances Householder reflection method should be preferred.