Experimental evaluation of Kernel Minimum Classification Error training

Recently, one popular discriminative training method for classifier design, Minimum Classification Error (MCE) training, has been significantly revised. This revision upgraded Large Geometric Margin Minimum Classification Error (LGM-MCE) training by embedding a kernel-based feature space projection mechanism. This latest MCE training is called Kernel Minimum Classification Error (KMCE) training and provides an efficient training procedure that can be performed in a comparatively low dimensional parameter space for a linear discriminant function defined in a kernel-projected high-dimensional feature space. Only KMCE's formalization was reported, but no experimental evaluations were conducted. In this paper, we evaluate KMCE training through systematic experiments and reveal that it achieves high classification rates when a reasonable amount (much less than needed by Support Vector Machines) of classifier parameters, such as weight vectors and prototypes, are available.