Imagenet-trained Cnns Are Biased Towards Texture; Increasing Shape Bias Improves Accuracy and Robustness January 01, 1000 https://arxiv.org/pdf/1811.12231 Fullscreen Dark Mode