CNNs trained on ImageNet are biased

Geirhos et al. (2018) published a paper on openreview on how training CNNs on ImageNet are heavily biased by texture and neglect shapes.