One of the remarkable properties of robust computer vision models is that their input-gradients are often aligned with human perception, referred to in the literature