Current deep models have achieved human-like accuracy in many computer vision tasks, even defeating humans sometimes. However, these deep models still suffer from significant weaknesses. To name a few, it is hard to interpret ...