Maximum Entropy Approximation for Kernel Machines

Kernel machines are widely used in pattern recognition, exploratory data analysis, and statistical signal processing, due to their effectiveness of modeling nonlinear dependencies in the data. The computational burden in evaluating forward functions in testing is the main drawback for kernel machines, especially in high dimensional large training set situations. We present a separable maximum entropy approximation for kernel machines that reduce the computational load for forward function evaluation. The performance of the approximation is demonstrated on kernel-based discriminative nonlinear projections on benchmark datasets.