Improvement algorithm for approximate incremental learning

Abstract

This paper presents an improved algorithm of Incremental Simple-PCA. The Incremental Simple-PCA is a fast incremental learning algorithm based on Simple-PCA. This algorithm need not hold all training samples because it enables update of an eigenvector according to incremental samples. Moreover, this algorithm has an advantage that it can calculate the eigenvector at high-speed because matrix calculation is not needed. However, it had a problem in convergence performance of the eigenvector. Thus, in this paper, we try the improvement of this algorithm from the aspect of convergence performance. We performed computer simulations using UCI datasets to verify the effectiveness of the proposed algorithm. As a result, its availability was confirmed from the standpoint of recognition accuracy and convergence performance of the eigenvector compared with the Incremental Simple-PCA. © 2009 Springer-Verlag Berlin Heidelberg.