Robust Classification with Context and Rejection
In some real-world applications of machine learning, despite the investments in the training of classification systems and in feature selection, misclassifications occur and their effects are critical. This is common in ill-posed classification problems, where overlapping classes, small or incomplete training sets, and unknown classes are prevalent. We can mitigate misclassifications and their effects by adapting the behavior of the classifier on samples with high potential for misclassification through the use of context and rejection. This combines the advantages brought by use of contextual priors in classification with the advantages of classification with rejection. In classification with rejection, we are able to increase classification performance at the expense of not classifying the entire data set. In this talk we explore the design of robust classifiers using context and rejection and the evaluation of performance of classifiers with rejection. We illustrate the results of robust classification on natural image segmentation, hyperspectral image classification, and automated digital histopathology.