Reduced Computation for Extreme Learning Machine Based on Cholesky Factorization

Abstract

This paper focuses on the implementation of matrix decomposition methods based on Cholesky factorization to reduce the computation time of pseudoinverse matrix solutions in the training process of extreme learning machine (ELM). The direct solution of the pseudoinverse matrix in ELM may result in singularity, and the cost of ELM computation time increases significantly as the number of hidden neurons grows. We proposed an approach for computing the pseudoinverse matrix in ELM based on the Cholesky factorization matrix decomposition method and compared our model to singular value decomposition (SVD), a well-known method for matrix decomposition. The numerical results show that our approach outperforms the SVD in reducing the computation time of the ELM training process without losing its generalization performance.