Generalization of Kernel PCA and Automatic Parameter Tuning

摘要

We propose a generalized kernel PCA which provides much more accuracy information of kernel space. Calculating partial derivatives of eigenvalues with kernel parameters, we can obtain the optimal kernel parameters. The criterion for optimal parameters are given by a quadratic cost function with respect to eigenvalues. We compared our method with SVM for face recognition, and showed that our method works efficiently as expected.