Sticky Reasoning within Learning Representations