Active Learning with a Drifting Distribution