Improving Model Training via Self-learned Label Representations