Collaborative Refining for Learning from Inaccurate Labels