Kernel Parameter Selection by Gap Maximization between Intra and Inter-Class Samples

Abstract

By maximizing the gap between classes in the reproducing kernel Hilbert space (RKHS), our method optimizes for the sigma values of radial basis function (RBF) or gaussian kernels. For each sample, we try to ensure the distance gap between intra-class and inter-class in RKHS to be large. Unlike previous methods of multiple kernel learning, our method does not need large amount of computations, which allows us to apply the proposed method to a larger set of data. Our method is compared with the method of kernel target alignment which is one of the most popular methods in multiple kernel learning to prove its efficiency of finding the optimal kernel parameter for the Face vs Non-face dataset.