Non-myopic active learning with mutual information

Abstract

Active learning methods seek to reduce the number of labeled instances needed to train an effective classifier. Most current methods are myopic, i.e. select a single unlabelled sample to label at a time. The batch-mode active learning methods, on the other hand, typically select top N unlabeled samples with maximum score. Such selected samples often cannot guarantee the learner's performance. In this paper, a non-myopic active learning algorithm is presented based on mutual information. Our algorithm selects a set of samples at each iteration, and the objective function of the algorithm is proved to be submodular, which guarantees to find the near-optimal solution. Our experimental results on UCI data sets show that the proposed algorithm outperforms myopic active learning.