Leveraging Human Reasoning To Understand And Improve Visual Question Answering