A Fast Algorithm for Sparse Extreme Learning Machine

As a more compact network, sparse ELM is an alternative model of extreme learning machine (ELM) for classification, which requires less memory space and testing time than conventional ELMs. In this paper, a fast training algorithm (FTA-sELM) specially developed for sparse ELM is proposed, which improves training speed, achieves better generalization performance and further promotes the application of sparse ELM in large data problem. The proposed algorithm breaks the large quadratic programming (QP) problem of sparse ELM into a series of two-dimensional sub-QP problems, specifically. In every iteration, Newton’s method is employed to solve the optimal solution for each sub-QP problem. Moreover, a new clipping scheme for Lagrange multipliers is presented, which improves convergence performance.