Adaptive learning algorithm for pattern classification

In this paper, a pattern classification task was regarded as a sample selection problem where a sparse subset of sample from the labeled training set was chosen. We proposed an adaptive learning algorithm utilizing the least square function to address this problem. Using these selected samples, which we call informative vectors, a classifier capable of recognizing the test samples was established. This novel algorithm is a combination of searching strategies that, not only based on forward searching steps, but adaptively takes backward steps to correct the errors introduced by earlier forward steps. We experimentally demonstrated on face image and text dataset that classifier using such informative vectors outperformed other methods.