Sparse Representation of Data and Support Vector Machines (in: Proceedings)

Abstract

We apply a new Blind Source Separation method (BSS), using sparseness, for identification of overdetermined linear mixing models, as we impose sparseness assumptions on the mixing matrix and no assumptions on the sources like independence or sparseness. We describe a suitable application of our method, for identification of kernel matrices in Support Vector Machines, under assumptions of sparseness of the kernel and existence of several learning processes with the same initial source data and different target ones. We present two examples confirming the good performance of our overdetermied BSS algorithms.