Kernel Dictionary Learning

Abstract

Dictionary Learning has proven to be a useful tool in many signal processing and machine learning applications, producing compelling results. An important use-case is as an unsupervised method of learning latent representations that can be used as the input features to a larger supervised classification system. The basic algorithm relies on inner-products between the input signals and a set of atomic dictionary elements and as such is amenable to kernel methods, which have been found to be very powerful in techniques such as non-linear Support Vector Machines and Kernel Regression. Based on previous kernel Dictionary Learning approaches, in this work we propose Sparse Kernel Dictionary Learning which provides significant gains in efficiency over its non-sparse counterpart. Additionally, we consider the online setting in which data arrive sequentially and demonstrate how sparse Dictionary Learning with kernels can be scaled up to extremely large datasets.