A Novel Sparse Extreme Learning Machine based Classifier

Abstract

Compared to traditional classifiers, such as SVM, the extreme learning machine (ELM) achieves similar performance for classification and runs at a much faster learning speed. However, the solution of ELM is dense and plenty of storage space and training time are required for large-scale applications. Traditional ELM method learns the output weights through the calculation of matrix inverse. In this paper, we propose a sparse ELM (SELM) method and the sparsity of output weights can reduce the storage space and training time. Furthermore, SELM updates the output weights through the proximal gradient descent method, which runs faster than the calculation of matrix inverse. Compared with ELM and SVM, SELM obtains better performance with much faster training speed and higher testing accuracy.