Approximate Consistency: Towards Foundations of Approximate Kernel Selection

Abstract

Kernel selection is critical to kernel methods. Approximate kernel selection is an emerging approach to alleviating the computational burdens of kernel selection by introducing kernel matrix approximation. Theoretical problems faced by approximate kernel selection are how kernel matrix approximation impacts kernel selection and whether this impact can be ignored for large enough examples. In this paper, we introduce the notion of approximate consistency for kernel matrix approximation algorithm to tackle the theoretical problems and establish the preliminary foundations of approximate kernel selection. By analyzing the approximate consistency of kernel matrix approximation algorithms, we can answer the question that, under what conditions, and how, the approximate kernel selection criterion converges to the accurate one. Taking two kernel selection criteria as examples, we analyze the approximate consistency of Nyström approximation and multilevel circulant matrix approximation. Finally, we empirically verify our theoretical findings.