Minimum classification error vs. maximum margin: How should we penalize unseen samples?

Abstract

One of the ultimate goals for classifier training is to achieve the classifier parameters that correspond to the minimum classification error probability status that should be derived using a classification error count loss. Recently, to pursue this ideal status, Minimum Classification Error (MCE) training has been successfully revised as Large Geometric Margin MCE training and Kernel MCE training. This paper gives an overview of the recent advancements of the MCE training methodology and discusses related issues.