Extreme Learning Machine Based on Adaptive Matrix Iteration

Yuxiang Li, Weidong Zou, Can Wang, Yuanqing Xia

Abstract: Under the continuous optimization and development of various algorithms in machine learning, the performance of the algorithm model on classification and regression prediction problems has become an important evaluation metric for the quality of algorithms. In order to solve the problems of low testing accuracy and unsatisfactory generalization performance of the models trained by the traditional extreme learning machine, this paper proposes an extreme learning machine algorithm based on adaptive convergence factor matrix iteration. This algorithm optimizes the calculation method of solving the hidden layer output weight matrix, while retaining the network structure model of the traditional extreme learning machine. This algorithm is implemented with a matrix iterative method that includes an adaptive convergence factor to compute the output weight matrix. As a result, it can adaptively select the optimal convergence factor according to the structure of the iterative equations, and thus use iterative method to solve linear equations efficiently and accurately upon ensuring the convergence of the equations. The experiment results show that the proposed algorithm has better performance in model training efficiency and testing accuracy, compared with the traditional extreme learning machine, the support vector machine, and other algorithms for data classification and regression prediction.