An active learner usually assumes there are some labeled data available based on which a moderate classifier is learned and then examines unlabeled data to manually label the most informative examples. However, for some application domains there are only extremely sparse labeled examples, such as one labeled example per category, attainable. In this case, existing active learning methods can not successfully apply, or the inefficient way of random selection for labeling will be first implemented. In this paper, a method seeking more high-informative examples for labeling based on very limited labeled data is proposed. By investigating the correlation between different views through canonical correlation analysis, our method can launch active learning using only one labeled example from each class. Promising experimental results are presented on several applications.
Shiliang Sun, David R. Hardoon