Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robustmix: Improving Robustness by Regularizing the Frequency Bias of Deep Nets

Jonas Ngnawe, Marianne Abemgnigni Njifon, Jonathan Heek, Yann Dauphin

TL;DR

This work proposes a novel extension of Mixup called Robustmix that regularizes networks to classify based on lower-frequency spatial features and shows that this type of regularization improves robustness on a range of benchmarks such as Imagenet-C and Stylized Imagenets.

Abstract

Deep networks have achieved impressive results on a range of well-curated benchmark datasets. Surprisingly, their performance remains sensitive to perturbations that have little effect on human performance. In this work, we propose a novel extension of Mixup called Robustmix that regularizes networks to classify based on lower-frequency spatial features. We show that this type of regularization improves robustness on a range of benchmarks such as Imagenet-C and Stylized Imagenet. It adds little computational overhead and, furthermore, does not require a priori knowledge of a large set of image transformations. We find that this approach further complements recent advances in model architecture and data augmentation, attaining a state-of-the-art mCE of 44.8 with an EfficientNet-B8 model and RandAugment, which is a reduction of 16 mCE compared to the baseline.

Robustmix: Improving Robustness by Regularizing the Frequency Bias of Deep Nets

TL;DR

This work proposes a novel extension of Mixup called Robustmix that regularizes networks to classify based on lower-frequency spatial features and shows that this type of regularization improves robustness on a range of benchmarks such as Imagenet-C and Stylized Imagenets.

Abstract

Deep networks have achieved impressive results on a range of well-curated benchmark datasets. Surprisingly, their performance remains sensitive to perturbations that have little effect on human performance. In this work, we propose a novel extension of Mixup called Robustmix that regularizes networks to classify based on lower-frequency spatial features. We show that this type of regularization improves robustness on a range of benchmarks such as Imagenet-C and Stylized Imagenet. It adds little computational overhead and, furthermore, does not require a priori knowledge of a large set of image transformations. We find that this approach further complements recent advances in model architecture and data augmentation, attaining a state-of-the-art mCE of 44.8 with an EfficientNet-B8 model and RandAugment, which is a reduction of 16 mCE compared to the baseline.
Paper Structure (10 sections, 7 equations, 4 figures, 3 tables, 1 algorithm)

This paper contains 10 sections, 7 equations, 4 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Results
  5. Datasets and Metrics
  6. Experimental Setup
  7. Robustness Results
  8. Ablation Study
  9. Analysis and Discussion
  10. Conclusion

Figures (4)

  • Figure 1: Illustration of the method. To better illustrate the method, we display the Fourier spectrum of the images next to them. We can see that even though 90% of the higher frequencies belong to the image of a dog, Robustmix assigns more weight to the boathouse label because it assigns more weight to the lower frequencies.
  • Figure 2: Plot of the cumulative energy in ImageNet images as a function of the frequency cutoff.
  • Figure 3: Highlighting the tradeoff between mCE and Clean Error for various models.
  • Figure 4: Test accuracy on ImageNet samples passed through a low-pass filter with increasing cut-off. As expected, we observe that Robustmix is more robust to the removal of high frequencies than Mixup. The comparison is done here on ResNet-50 models.