Toward optimal probabilistic active learning using a Bayesian approach (Q2051315)

Active learning methods select the next input to be learned in order to make the learning more efficient under the supervised learning, that is, the learning of input-output relationships. This paper considers the active learning framework in a classification setting. Using the Dirichlet conjugate prior for the categorical model, the authors define the estimated risk difference and propose a new active learning strategy, expected probabilistic gain for active learning (xPAL). With some modifications, xPAL reduces to existing active learning methods, expected error reduction (EER), probabilistic active learning (PAL), and uncertainty sampling (US). The authors show these relationships theoretically and compare the performance of xPAL with the existing methods numerically. Although little is discussed on the connections to active learning frameworks based on statistical experimental design, xPAL provides a unifying framework for the recent active learning methods.