Incremental Feature Extraction Based on Empirical Kernel Map

Abstract

A new incremental kernel principal component analysis is proposed for the nonlinear feature extraction from the data. The problem of batch kernel principal component analysis is that the computation becomes prohibitive when the data set is large . Another problem is that, in order to update the eigenvectors with another data, the whole decomposition from scratch should be recomputed. The proposed method overcomes these problems by incrementally update eigenspace and using empirical kernel map as kernel function. The proposed method is more efficient in memory requirement than a batch kernel principal component and can be easily improved by re-learning the data. In our experiments we show that proposed method is comparable in performance to a batch kernel principal component for the classification problem on nonlinear data set.