Dropout-based Support Vectors Regularization

Abstract

In this thesis we consider a new regularization technique that exploits the probabilistic Dropout scheme at the sample level. The new regularization approach is incorporated into the Maximum Margin Classification (MMC) framework resulting in a new variant of the Support Vector Machine classifier. We show here that the added regularizer comes with a geometrical interpretation related to the selection of support vectors. In addition, we illustrate that the new formulation is consistent with the guarantee provided in the Statistical Learning Theory. Experimental results from several classification problems show better generalization performance achieved by adding the new regularization as compared to the standard approach.