Adaptive Kernel Matching Pursuit for Pattern Classification

摘要

A sparse classifier is guaranteed to generalize better than a denser one, given they perform identical on the training set. However, methods like Support Vector Machine, even if they produce relatively sparse models, are known to scale linearly as the number of training examples increases. A recent proposed method, the Kernel Matching Pursuit, presents a number of advantages over the SVM, like sparser solutions and faster training. In this paper we present an extension of the KMP in which we prove that adapting the dictionary to the data results in improved performances. We discuss different techniques for dictionary adaptation and present some results on standard datasets.