Instance-selecting regularization penalty for supervised image classifcation

Abstract

We propose a new approach to perform classification in the presence of large amounts of irrelevant or noisy instances. This is achieved by designing a convex objective function whose optimization finds the most relevant samples for classification, and uses them to discriminate between classes. In particular, we combine the logistic loss with a combination of L2 and L2,1-type regularization penalties that enforce within-group sparsity. We study the properties of the L2,1 regularization, compare it to standard regularization penalties, and propose the final algorithm with its optimization. Experimental results in real world applications show improvement over unstructured regularization penalties.