Robustness properties of Facebook's ResNeXt WSL models

A. Emin OrhanarXiv 2019 · OpenReview

TL;DR

This paper investigates the robustness of Facebook's ResNeXt WSL models, trained on a massive dataset of ~1 billion Instagram images. These models exhibit remarkable resilience to common image corruptions and perturbations, significantly outperforming previous benchmarks. While they show some limited adversarial robustness, they don't achieve true adversarial resilience and retain a strong texture bias, suggesting that simply scaling up data isn't a silver bullet for all robustness challenges.

What Problem Does This Paper Address?

The field of computer vision has made incredible strides in image recognition, largely driven by models trained on datasets like ImageNet. However, these models often exhibit brittle behavior: they can be easily fooled by minor image corruptions, adversarial perturbations, or even subtle shifts in visual cues like shape versus texture. A key question is whether simply increasing the scale of training data, particularly with weakly supervised, real-world data, can inherently improve these robustness properties. This paper aims to answer this by evaluating Facebook's recently released ResNeXt WSL models, trained on an unprecedented ~1 billion Instagram images, against various robustness benchmarks.

Methods

The researchers evaluated five ResNeXt WSL models (with varying architectural depths) and compared them to a standard ImageNet-trained ResNeXt model. Their evaluation focused on several key areas:

  • Common Corruptions and Perturbations: They used the ImageNet-C and ImageNet-P benchmarks, which systematically introduce various corruptions (e.g., noise, blur) and transformations to images. Robustness is measured by metrics like mean corruption error (mCE) and mean flip rate (mFR).
  • Natural Adversarial Examples: The ImageNet-A benchmark, consisting of natural images that are difficult for standard classifiers, was used to assess out-of-sample generalization.
  • Adversarial Robustness: They employed both black-box and white-box Projected Gradient Descent (PGD) attacks to test the models' resilience to adversarial perturbations. They specifically investigated the impact of varying the number of PGD steps.
  • Feature Visualization: To understand the internal representations, they generated "maximizing images" by finding images that maximally activate specific neurons in the network. This technique, inspired by prior work, can reveal signatures of robustness or non-robustness.
  • Shape Bias: Using shape-texture cue conflict stimuli, they tested whether the models rely more on global shape information or local texture information, comparing them to human performance.

Datasets

  • Training Data: ~1 billion weakly supervised images from Instagram, fine-tuned on ImageNet.
  • Evaluation Benchmarks:
    • ImageNet-C: A benchmark for common image corruptions.
  • ImageNet-P: A benchmark for image perturbations.
  • ImageNet-A: A benchmark for "natural adversarial examples" (images that are hard for standard classifiers).
  • Standard ImageNet validation set: Used for baseline accuracy and comparison.
  • Shape-texture cue conflict stimuli: Created by Geirhos et al. (2019) to assess shape vs. texture bias.

Key Insights and Ideas

  • "Unreasonable Effectiveness of Data": The paper strongly supports the idea that massive-scale training data can lead to significant improvements in robustness, even without explicit adversarial training.
  • ImageNet-C/P as Better Benchmarks: The robustness gains on ImageNet-C and ImageNet-P far exceeded gains in standard ImageNet validation accuracy, suggesting these benchmarks are more indicative of true progress in robust image recognition.
  • Limited Adversarial Robustness: While WSL models show some resilience to 10-step PGD attacks, their robustness rapidly degrades with more steps, indicating they don't achieve genuine adversarial robustness like models specifically trained for it.
  • Feature Visualization as a Diagnostic Tool: The "maximizing images" generated for WSL models were perceptually similar to seed images and to each other, mirroring the behavior of non-robust models, thus supporting the conclusion of limited adversarial robustness.
  • Persistent Texture Bias: Despite increased scale, the WSL models remain significantly more texture-biased than humans, suggesting that simply more data doesn't automatically solve fundamental biases in learned representations.

Key Results

  • The largest ResNeXt WSL model (resnext101_32x48d_wsl) achieved state-of-the-art results on ImageNet-C, ImageNet-P, and ImageNet-A by a significant margin.
  • On ImageNet-C, the best WSL model achieved an mCE of 45.7, compared to 66.6 for the ImageNet-trained model.
  • On ImageNet-A, the best WSL model achieved a top-1 accuracy of 61.0%, a substantial improvement over the ImageNet-trained model's 10.2%.
  • Against 10-step white-box PGD attacks (with $\epsilon=0.06$), the best WSL model achieved 40.7% accuracy, whereas the standard ImageNet model dropped to 0.0%.
  • However, increasing PGD steps to 50 reduced WSL model accuracy to near zero, while a dedicated adversarially trained model maintained higher accuracy.
  • Shape bias increased in WSL models (e.g., 42.8% for the largest WSL model vs. 25.9% for the standard model), but remained far below human levels (95.9%).

Limitations and Future Work

  • Genuine Adversarial Robustness: The paper concludes that ResNeXt WSL models do not achieve true adversarial robustness, as their performance degrades rapidly with increased PGD steps. This suggests that while large-scale data helps, it's not a complete solution for adversarial attacks.
  • Persistent Texture Bias: The models still exhibit a strong texture bias compared to humans, indicating that more data alone doesn't resolve all learned biases.
  • Future Work: The authors pose the open question of what tasks and model biases are necessary to achieve genuinely human-like, general-purpose visual representations that are robust, shape-based, and handle out-of-sample generalization effectively. They suggest that further research is needed to understand how to achieve these more desirable properties.

Summary updated

Abstract

We investigate the robustness properties of ResNeXt class image recognition models trained with billion scale weakly supervised data (ResNeXt WSL models). These models, recently made public by Facebook AI, were trained with ~1B images from Instagram and fine-tuned on ImageNet. We show that these models display an unprecedented degree of robustness against common image corruptions and perturbations, as measured by the ImageNet-C and ImageNet-P benchmarks. They also achieve substantially improved accuracies on the recently introduced "natural adversarial examples" benchmark (ImageNet-A). The largest of the released models, in particular, achieves state-of-the-art results on ImageNet-C, ImageNet-P, and ImageNet-A by a large margin. The gains on ImageNet-C, ImageNet-P, and ImageNet-A far outpace the gains on ImageNet validation accuracy, suggesting the former as more useful benchmarks to measure further progress in image recognition. Remarkably, the ResNeXt WSL models even achieve a limited degree of adversarial robustness against state-of-the-art white-box attacks (10-step PGD attacks). However, in contrast to adversarially trained models, the robustness of the ResNeXt WSL models rapidly declines with the number of PGD steps, suggesting that these models do not achieve genuine adversarial robustness. Visualization of the learned features also confirms this conclusion. Finally, we show that although the ResNeXt WSL models are more shape-biased than comparable ImageNet-trained models in a shape-texture cue conflict experiment, they still remain much more texture-biased than humans, suggesting that they share some of the underlying characteristics of ImageNet-trained models that make this benchmark challenging.
BibTeX
@misc{orhan2019robustness,
  author = {A. Emin Orhan},
  title = {{Robustness properties of Facebook's ResNeXt WSL models}},
  year = {2019},
  eprint = {1907.07640},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.07640},
}
Figure