Accelerated Learning with Kernels

Опубликовано 8 августа 2016, 18:14

Kernel learning algorithms occupy a prominent position within machine learning having given state-of-the-art performance in several domains. Much of the power of kernel methods comes from their ability to implicitly represent complex functions in high dimensional spaces. This however, comes at a price of increased hypothesis complexity that causes these algorithms to be slow at prediction time. With an increase in demand for real time applications, this prevents kernel algorithms from being applied to several domains. A second limitation of traditional kernel-based learning methods is their dependence on so-called "Mercer kernels" that prevents them from fully utilizing rich domain-specific knowledge in the learning process. Our work seeks to address both these issues by developing kernel learning algorithms that offer fast prediction routines. We further develop a learning framework that allows efficient use of non-Mercer kernels in addition to offering fast training and testing routines.