New Microsoft + Cornell Univ paper gives a method that matched a standard model trained on 50% more …