Delimiting kernel-based learning methods
(this project has ended)

Learning systems based on kernels are a powerful class of algorithms that includes Support Vector Machines and Gaussian Processes. These systems have become a major part of current research into and applications of adaptive systems. Despite this fact very little is known about when we can expect these systems to perform well. There has even been the assumption made that they provide a universal learning methodology. The proposed project will address this in order to:

• provide theoretical tools that describe when a set of functions can be realised by hyperplanes with non-trivial margins in some feature space;
• describe how the degree of matching between a kernel and a problem domain can be measured;
• develop methods for choosing kernels as attuned as possible to a particular problem/domain;
• develop alternative `luckiness' functions that give rise to efficient generic learning methods for problems that cannot be solved using kernel methods.

Funding

• EPSRC

Principal Investigators

• jst

Other Investigators

• aa