Multi-Task Active Learning with Output Constraints