Convergence rates of active learning for maximum likelihood estimation

摘要: An active learner is given a class of models, large set unlabeled examples, and the ability to interactively query labels subset these examples; goal learn model in that fits data well. Previous theoretical work has rigorously characterized label complexity learning, but most this focused on PAC or agnostic model. In paper, we shift our attention more general setting - maximum likelihood estimation. Provided certain conditions hold class, provide two-stage learning algorithm for problem. The require are fairly general, cover widely popular Generalized Linear Models, which turn, include models binary multi-class classification, regression, conditional random fields. We an upper bound requirement algorithm, lower matches it up order terms. Our analysis shows unlike classification realizable case, just single extra round interaction sufficient achieve near-optimal performance On empirical side, recent [12] [13] (on linear logistic regression) promise approach.