ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness.