Deep banach space kernels

Abstract

The recent success of deep learning has encouraged many researchers to explore the deep/concatenated variants of classical kernel methods. Some of which includes MLMKL, DGP and DKL. Although, These methods have proven to be quite useful in various real-world settings. They still suffer from the limitations of only utilizing kernels from Hilbert spaces. In this paper, we address these shortcomings by introducing a new class of concatenated kernel learning methods that use the kernels from the reproducing kernel Banach spaces(RKBSs). These spaces turned out to be one of the most general spaces where a reproducing Kernel exists. We propose a framework of construction for these Deep RKBS models and then provide a representer theorem for regularized learning problems. We also describe the relationship with its deep RKHS variant as well as standard Deep Gaussian Processes. In the end, we construct and implement a two-layer deep RKBS model and demonstrate it on a range of machine learning tasks.

Tasks

Reproductions