Active learning with optimal distribution for image classification

In this paper, we focus on the issue of building up a training set for the task of image classification at minimal labeling costs. It is a topic that has attracted the considerable attention in the recent years. We propose a novel active learning algorithm with optimal distribution. In order to solve the problems of the noisy distribution and the sampling bias in the actively sampling process, the empirical risk on the selected examples is weighted by density ratio, and then the risk on the test examples is estimated using only unlabeled examples and the marginal label distribution. Finally, the optimal training distribution is derived by minimizing the expected error of the risk. Our approach has been demonstrated on the task of image classification on the difficult benchmark PASCAL VOC 2007 dataset.