The Origins and Prevalence of Texture Bias in Convolutional Neural Networks

By Katherine L. Hermann, Ting-Li Chen, Simon Kornblith. Neural Information Processing Systems, 2019

356 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

A Scaling Law for Syn2real Transfer: How Much Is Your Pre-training Effective? (2021, background): “Because CNNs prefer to recognize objects by textures [21, 26], the pre-trained models may overﬁt to learn the texture features.”
Partial success in closing the gap between human and machine vision (2021, methodology): “When comparing self-supervised SimCLR models with augmentation-matched baseline models trained in the standard supervised fashion (Figure 15 in the Appendix), we find that the augmentation scheme (rather than the self-supervised objective) indeed made the crucial difference: supervised baselines sho”
Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations (2022, methodology): “[37] showed that shape information is decodable from the features of ImageNet-trained models on the GST dataset to a certain extent.”
The Role of ImageNet Classes in Fréchet Inception Distance (2022, background): “Fourth, ImageNet classifiers are known to base their decisions primarily on textures instead of shapes [13, 15].”
Toward Transformer-Based Object Detection (2020, background): “Other works have explored the relationship between selfattention modules and convolutional layers [8] as well as the limitations of convolutional neural networks [40, 16].”