Extreme support vector machine classifier

Abstract

Instead of previous SVM algorithms that utilize a kernel to evaluate the dot products of data points in a feature space, here points are explicitly mapped into a feature space by a Single hidden Layer Feedforward Network (SLFN) with its input weights randomly generated. In theory this formulation, which can be interpreted as a special form of Regularization Network (RN), tends to provide better generalization performance than the algorithm for SLFNs-Extreme Learning Machine (ELM) and leads to a extremely simple and fast nonlinear SVM algorithm that requires only the inversion of a potentially small matrix with the order independent of the size of the training dataset. The experimental results show that the proposed Extreme SVM can produce better generalization performance than ELM almost all of the time and can run much faster than other nonlinear SVM algorithms with comparable accuracy. © 2008 Springer-Verlag Berlin Heidelberg.