• ImageNet trained CNNs are biased towards recognising textures rather than shapes https://openreview.net/pdf?id=Bygh9j09KX 5 comments 24/12/2020