Active learning by learning

摘要: Pool-based active learning is an important technique that helps reduce labeling efforts within a pool of unlabeled instances. Currently, most pool-based strategies are constructed based on some human-designed philosophy; is, they reflect what human beings assume to be "good questions." However, while such philosophies can useful specific data sets, it often difficult establish the theoretical connection those true performance interest. In addition, given single philosophy unlikely work all scenarios, choosing and blending under different scenarios but challenging practical task. This paper tackles this task by letting machines adaptively "learn" from set particular set. More specifically, we design algorithm connects with well-known multi-armed bandit problem. Further, postulate that, appropriate choice for learner, possible estimate fly. Extensive empirical studies resulting ALBL confirm performs better than state-of-the-art leading learning, which philosophy.