"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization" Paper Review
0 likes•531 views
The document presents a critical analysis of out-of-distribution generalization in vision models, emphasizing the importance of robustness in various applications. It introduces new benchmarks, including imagenet-r, streetview storefronts, and deepfashion remixed, along with a novel data augmentation technique called deepaugment. The study evaluates seven hypotheses regarding model robustness and concludes that robustness manifests in multiple ways, suggesting the need for further development of robustness methods in the research community.
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization" Paper Review
- 1. 2020/09/08 Ho Seong Lee (hoya012) Cognex Deep Learning Lab Research Engineer SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 1
- 2. Contents • Introduction • Related Work • New Benchmarks • DeepAugment • Experiments • Conclusion SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 2
- 3. Introduction Human vision system is robust, but existing vision models are not robust. • Humans can deal with many forms of corruption such as blur, pixel noise and abstract change in structure and style. • Achieving robustness is essential in safety-critical and accuracy-critical applications. SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 3 Dog! Dog!Dog! Dog! Dog! starfish! baseball! drumstick!
- 4. Introduction Most work on robustness methods for vision has focused on adversarial examples. • For standardizing and expanding the robustness topic, several studies have begun. • At first, establish benchmark dataset and evaluation metric. SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 4
- 5. Related Works Pioneers of robustness in ML - Dan Hendrycks (My Favorite Researcher..) SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 5 Reference: https://people.eecs.berkeley.edu/~hendrycks/
- 6. Related Works Pioneers of robustness in ML – Madry Lab SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 6 Reference: http://madry-lab.ml/
- 7. Related Works Pioneers of robustness in ML – Bethge Lab SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 7 Reference: http://bethgelab.org/
- 8. Related Works SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 8 Benchmarking Neural Network Robustness to Common Corruptions and Perturbations • Create ImageNet-C and ImageNet-P test set and Benchmarks. ImageNet-C ImageNet-P
- 9. Related Works SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 9 Benchmarking Neural Network Robustness to Common Corruptions and Perturbations • Robustness Metrics = 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 ( ) * each level of severity s (1 ≤ s ≤ 5). ImageNet-C ImageNet-P
- 10. Related Works SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 10 Natural Adversarial Examples • Introduce natural adversarial examples and create 7,500 ImageNet-A test set. (200 class) • Download numerous images related to an ImageNet class from website iNaturalist and Flickr. • Delete the images that ResNet-50 correctly classify. Finally, select a subset of high-quality images. ImageNet-A
- 11. Related Works SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 11 AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty • Propose a technique to improve the robustness and uncertainty estimates of image classifiers. • Use JS Divergence Consistency loss
- 12. New Benchmarks SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 12 Seven robustness hypotheses • Larger Models: increasing model size improves robustness. • Self-Attention: adding self-attention layers to models improves robustness. • Diverse Data Augmentation: robustness can increase through data augmentation. • Pretraining: pretraining on larger and more diverse datasets improves robustness. • Texture Bias: convolutional networks are biased towards texture, which harms robustness. • Only IID Accuracy Matters: accuracy on independent and identically distributed test data entirely determines natural robustness. • Synthetic ≠Natural: synthetic robustness interventions including diverse data augmentations do not help with robustness on naturally occurring distribution shifts.
- 13. New Benchmarks SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 13 Introduce three new robustness benchmarks • It has been difficult to arbitrate these hypotheses because existing robustness datasets preclude the possibility of controlled experiments by varying multiple aspects simultaneously. • To address these issues and test the seven hypotheses outlined above, create new test sets.
- 14. New Benchmarks SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 14 ImageNet-Renditions (ImageNet-R) • 30,000 test set containing various renditions (e.g., paintings, embroidery, etc.) of ImageNet object classes (200). The rendition styles (“Painting”, “Toy”) are not ImageNet-R’s classes. • Original ImageNet dataset discouraged such images since annotators were instructed to collect “photos only, no painting, no drawings, etc.” (Deng, 2012). Authors do the opposite.
- 15. New Benchmarks SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 15 StreetView StoreFronts (SVSF) • Contains business storefront images taken from Google Streetview (20 classes). • Investigate natural shifts in the image capture process using metadata (e.g. location, year, camera type) • Create one training set(200K) and five in-distribution test set(10K) from images taken in US/Mexico/Canada during 2019 using “new” camera. Unfortunately, unleased.. • Make four out-of-distribution test set(10K): “2017”, “2018”, “France”, “old camera”
- 16. New Benchmarks SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 16 DeepFashion Remixed • Changes in camera operation can cause shifts in attributes such as object size, object occlusion, camera viewpoint, and camera zoom. • To measure this, create multi-labeled training set(48K) and 8 out-of-distribution test set (Total 121K). Medium scale, Medium occlusion, side/back viewpoint, no zoom-in Small and large scale, Minimal and heavy occlusion, Frontal and not-worn viewpoints, Medium and large zoom-in
- 17. DeepAugment SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 17 New data augmentation technique: DeepAugment • In order to explore the Diverse Data Augmentation hypothesis, introduce a new data augmentation. • Pass an image through an img-to-img networks (such as autoencoder or SR network) • But rather than processing the image normally, distort the internal weights and activations by applying randomly sampled ops(zeroing, negating, convolving, transposing), applying activation functions on. • This creates diverse but semantically consistent images.
- 18. DeepAugment SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 18
- 19. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 19 Experimental Setup
- 20. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 20 Experimental Results: ImageNet-R • ImageNet-200: The original ImageNet test set restricted to ImageNet-R’s 200 classes. • Pretraining → improves IID/OOD gap with very small portion. • Self-Attention → increase IID/OOD gap • Diverse Data Augmentation, Larger Models → improves IID/OOD gap significantly! OODIID Diff -0.2 0.0 0.2 -12.5 -2.0 -6.4 -7.1 -5.7 -10.8 -4.1 Error Rate (↓)
- 21. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 21 Experimental Results: StreetView StoreFronts • No method helps much on country shift, where error rates roughly double across the board. • Images captured in France contain noticeably different architecture styles and storefront designs. • Unable to find conspicuous and consistent indicators of the camera and year. → insensitive feature. • Data augmentation primarily helps combat texture bias as with ImageNet-R. • But existing augmentations are not diverse enough to capture high-level semantic shifts such as building architecture. Error Rate (↓)
- 22. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 22 Experimental Results: DeepFashion Remixed • All evaluated methods have an average OOD mAP that is close to the baseline. • DFR’s size and occlusion shifts hurt performance the most. • Nothing substantially improved OOD performance beyond what is explained by IID performance, so here it would appear that Only IID Accuracy Matters. mAP scores (↑)
- 23. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 23 Experimental Results: ImageNet-C • DeepAugment + AugMix → attain the SOTA result. • Evidence for Larger Models, Self-Attention, Diverse Data Augmentation, Pretraining, and Texture Bias. • Evidence against the Only IID Accuracy Matters. Error Rate, mCE (↓)
- 24. Experiments SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization 24 Experimental Results: Real Blurry Images • ImageNet-C use various synthetic corruptions that have different from real-world. • Collect a small dataset of 1,000 real-world blurry images and evaluate various models. • Everything that helped in ImageNet-C was also helpful in Real Blurry Images. • Evidence against the Synthetic ≠Natural. Error Rate (↓)
- 25. Conclusion SNUAI 8th | The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization • Introduce three new benchmarks, ImageNet-R, SVSF, and DFR. (+ Real Blurry Images) • Introduce new data augmentation technique DeepAugment. • With these benchmarks, evaluate seven robustness hypotheses. • It seems that robustness has a many faces (multivariate). • If so, research community should prioritize creating new robustness methods. 25