Increasing the robustness of DNNs against image corruptions by playing the Game of Noise

By Evgenia Rusak, Lukas Schott, Roland S. Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, Wieland Brendel. 2020

66 citations recorded by the source corpus.

The human visual system is remarkably robust against a wide range of naturally occurring variations and corruptions like rain or snow. In contrast, the performance of modern image recognition models strongly degrades when evaluated on previously unseen corruptions. Here, we demonstrate that a simple but properly tuned training with additive Gaussian and Speckle noise generalizes surprisingly well to unseen corruptions, easily reaching the previous state of the art on the corruption benchmark ImageNet-C (with ResNet50) and on MNIST-C. We build on top of these strong baseline results and show that an adversarial training of the recognition model against uncorrelated worst-case noise distributions leads to an additional increase in performance. This regularization can be combined with previously proposed defense methods for further improvement.

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging, Safety, Robustness & Interpretability

How researchers see this paper

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

In Artificial Intelligence

1. Dawn Song
2. Dan Hendrycks
3. Quoc V. Le
4. Jacob Steinhardt
5. Wieland Brendel
6. Matthias Bethge
7. Nicholas Carlini
8. Percy Liang

In Radiology, Nuclear Medicine and Imaging

1. Dawn Song
2. Ting-Li Chen
3. Justin Gilmer
4. Simon Kornblith
5. Quoc V. Le

In Safety, Robustness & Interpretability

1. Dan Hendrycks
2. Dawn Song
3. Jacob Steinhardt
4. Wieland Brendel
5. Nicholas Carlini
6. Matthias Bethge
7. Matthias Hein
8. Oliver Bringmann

Citations in the corpus

• The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization (2020, background): “Speckle Noise data augmentation muliplies each pixel by (1 + x) with x sampled from a normal distribution [27, 12].”
• The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (2019, background): “Recent work has shown that data augmentation can also improve out-of-distribution robustness [33, 102, 65, 45, 78].”
• Improving robustness against common corruptions by covariate shift adaptation (2020, methodology): “Model base adapt ∆ ResNet50 63.8 59.9 -3.9 SIN 58.6 54.2 -4.4 ANT 61.0 58.0 -3.0 ANT+SIN 53.8 52.0 -1.8 AugMix (AM) 59.0 55.8 -3.2 DeepAug (DAug) 57.8 52.5 -5.3 DAug+AM 53.2 48.9 -4.3 DAug+AM+RNXt101 47.9 44.0 -3.9 First, using Fixup initialization [39] alleviates the need for BN layers.”