Active Learning with Task-Driven Representations for Messy Pools