Towards Effective Visual Representations for Partial-Label Learning