Speed-up of kernel based nonlinear subspace method by sparsification of basis vectors

摘要

A speed-up method is presented for the kernel based nonlinear subspace method by Maeda et al. and Tsuda on the basis of sparsification of basis vectors of each subspace. Computational cost of projection of data onto subspaces is reduced by using sparse basis vectors. The presented method is applied to face detection from images. A method is also presented for revising sparse basis vectors on the basis of a power method for eigenvalue problems.