Learning Human Motion Representations: A Unified Perspective