Efficient formation of a basis in a kernel induced feature space
Gavin C. Cawley, Nicola L. C. Talbot · UEA Digital Repository (University of East Anglia) · 2002
Baudat and Anouar [1] propose a simple greedy algorithm for estimation of an approximate basis of the subspace spanned by a set of fixed vectors embedded in a kernel induced feature space. The resulting set of basis vectors can then be used to construct sparse kernel expansions for classification and regression tasks. In this paper we describe five algorithmic improvements to the method of Baudat and Anouar, allowing the construction of an approximate basis with a computational complexity that is independent of the number of training patterns, depending only on the number of basis vectors extracted.