Active Learning from Multiple Knowledge Sources

Abstract

Some supervised learning tasks do not fit the usual single annotator scenario. In these problems, ground-truth may not exist and multiple annotators are generally available. A few approaches have been proposed to address this learning problem. In this setting active learning (AL), the problem of optimally selecting unlabeled samples for labeling, offers new challenges and has received little attention. In multiple annotator AL, it is not sufficient to select a sample for labeling since, in addition, an optimal annotator must also be selected. This setting is of great interest as annotators’ expertise generally varies and could depend on the given sample itself; additionally, some annotators may be adversarial. Thus, clearly the information provided by some annotators should be more valuable than that provided by others and it could vary across data points. We propose an AL approach for this new scenario motivated by information theoretic principles. Specifically, we focus on maximizing the information that an annotator label provides about the true (but unknown) label of the data point. We develop this concept, propose an algorithm for active learning, and experimentally validate the proposed approach.