Divide and Ensemble: Progressively Learning for the Unknown