Training Algorithm For Radial Basis Function Classifier

Abstract

The computational complexity of kernel machines and their poor performance in the multi-label classification case is a major bottleneck in their success. In this thesis we present a systematic two step batch approach for constructing and training a new multiclass kernel machine (MKM). Unlike other kernel learning algorithms, the proposed paradigm prunes the kernels, and uses Newton’s method to improve the kernel parameters. In each iteration, output weights are found using orthogonal least squares. The proposed hybrid training algorithm is compared with those least square support vector machines(LS-SVM) and support vector machines(SVM). Simulations results on many benchmark and real life datasets show that the proposed algorithm has significantly improved convergence speed, small network size and better generalization than conventional kernel machine training algorithms.