Fast regularized kernel function approximation

Abstract

We propose a simple formulation for regularized kernel function approximation. The regression function is obtained by minimizing an unconstrained quadratic function. Further reduced kernel technique is also employed in its formulation which enables it to handle large database. The solution of this objective function is obtained by solving a system of linear equations and thus no need for any quadratic programming problem solver. Its simple MATLAB implementation is also very fast. Computational results on artificial and real world data sets and comparisons are given to demonstrate the fast training time and low prediction error of the proposed formulation.