Nonlinear mappings for generative kernels on latent variable models

Abstract—Generative kernels have emerged in the last yearsas an effective method for mixing discriminative and generativeapproaches. In particular, in this paper, we focus on kernelsdefined on generative models with latent variables (e.g. thestates in a Hidden Markov Model). The basic idea underlyingthese kernels is to compare objects, via a inner product,in a feature space where the dimensions are related to thelatent variables of the model. Here we propose to enhancethese kernels via a nonlinear normalization of the space,namely a nonlinear mapping of space dimensions able toexploit their discriminative characteristics. In this paper weinvestigate three possible nonlinear mappings, for two HMMbasedgenerative kernels, testing them in different sequenceclassification problems, with really promising results.