Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
97 tokens/sec
GPT-4o
53 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
5 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Boosting Model Resilience via Implicit Adversarial Data Augmentation (2404.16307v2)

Published 25 Apr 2024 in cs.LG and cs.CV

Abstract: Data augmentation plays a pivotal role in enhancing and diversifying training data. Nonetheless, consistently improving model performance in varied learning scenarios, especially those with inherent data biases, remains challenging. To address this, we propose to augment the deep features of samples by incorporating their adversarial and anti-adversarial perturbation distributions, enabling adaptive adjustment in the learning difficulty tailored to each sample's specific characteristics. We then theoretically reveal that our augmentation process approximates the optimization of a surrogate loss function as the number of augmented copies increases indefinitely. This insight leads us to develop a meta-learning-based framework for optimizing classifiers with this novel loss, introducing the effects of augmentation while bypassing the explicit augmentation process. We conduct extensive experiments across four common biased learning scenarios: long-tail learning, generalized long-tail learning, noisy label learning, and subpopulation shift learning. The empirical results demonstrate that our method consistently achieves state-of-the-art performance, highlighting its broad adaptability.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (56)
  1. Invariance principle meets information bottleneck for out-of-distribution generalization. In NeurIPS, pages 3438–3450, 2021.
  2. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019.
  3. Nuanced metrics for measuring unintended bias with real data for text classification. In WWW, pages 491–500, 2019.
  4. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, pages 1567–1578, 2019.
  5. Imagine by reasoning: A reasoning-based implicit semantic data augmentation for long-tailed classification. In AAAI, pages 356–364, 2022.
  6. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, pages 3008–3017, 2020.
  7. Class-balanced loss based on effective number of samples. In CVPR, pages 9268–9277, 2019.
  8. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8536–8546, 2018.
  9. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016.
  10. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, pages 10477–10486, 2018.
  11. Disentangling label distribution for long-tailed visual recognition. In CVPR, pages 6626–6636, 2021.
  12. Safa: Sample-adaptive feature augmentation for long-tailed image classification. In ECCV, pages 587–603, 2022.
  13. The inaturalist species classification and detection dataset. In CVPR, pages 8769–8778, 2018.
  14. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In CVPR, pages 7610–7619, 2020.
  15. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, pages 2304–2313, 2018.
  16. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020.
  17. Learning multiple layers of features from tiny images. Technical report, 2009.
  18. Out-of-distribution generalization via risk extrapolation (rex). In ICML, pages 5815–5826, 2021.
  19. Graddiv: Adversarial robustness of randomized neural networks via gradient diversity regularization. IEEE TPAMI, 45(2):2645–2651, 2023.
  20. Domain generalization with adversarial feature learning. In CVPR, pages 5400–5409, 2018.
  21. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, pages 5208–5217, 2021.
  22. Logit perturbation. In AAAI, pages 1359–1366, 2022.
  23. Deep learning face attributes in the wild. In ICCV, pages 3730–3738, 2016.
  24. Large-scale long-tailed recognition in an open world. In CVPR, pages 2537–2546, 2019.
  25. Dimensionality-driven learning with noisy labels. In ICML, pages 3355–3364, 2018.
  26. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543–6553, 2020.
  27. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018.
  28. A review: Data pre-processing and data augmentation techniques. Global Transitions Proceedings, 3(1):91–99, 2022.
  29. Long-tail learning via logit adjustment. In ICLR, 2021.
  30. Learning to reweight examples for robust deep learning. In ICML, pages 4334–4343, 2018.
  31. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, pages 4175–4186, 2020.
  32. Meta-learning advisor networks for long-tail and noisy labels in social image classification. ACM TOMM, 19(5s):1–23, 2023.
  33. Imagenet large scale visual recognition challenge. IJCV, 115(3):211–252, 2015.
  34. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In ICLR, 2020.
  35. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.
  36. Gradient matching for domain generalization. In ICLR, 2022.
  37. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pages 1919–1930, 2019.
  38. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In NeurIPS, pages 1513–1524, 2020.
  39. Invariant feature learning for generalized long-tailed classification. In ECCV, pages 709–726, 2022.
  40. Improving deep learning with generic data augmentation. In SSCI, pages 1542–1547, 2018.
  41. Implicit semantic data augmentation for deep networks. In NeurIPS, pages 12635–12644, 2019.
  42. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, pages 13723–13732, 2020.
  43. Aggregated residual transformations for deep neural networks. In CVPR, pages 5987–5995, 2017.
  44. Universal adaptive data augmentation. In IJCAI, pages 1596–1603, 2023.
  45. L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. In NeurIPS, pages 6225–6236, 2019.
  46. Adversarial domain adaptation with domain mixup. In AAAI, pages 6502–6509, 2020.
  47. To be robust or to be fair: Towards fairness in adversarial training. In ICML, pages 11492–11501, 2021.
  48. Improving out-of-distribution robustness via selective augmentation. In ICML, pages 25407–25437, 2022.
  49. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.
  50. Mixup: Beyond empirical risk minimization. In ICLR, 2018.
  51. Balanced knowledge distillation for long-tailed learning. Neurocomputing, 527:36–46, 2023.
  52. Meta label correction for noisy label learning. In AAAI, pages 11053–11061, 2021.
  53. Improving calibration for long-tailed recognition. In CVPR, pages 16489–16498, 2021.
  54. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, pages 9719–9728, 2020.
  55. Combining adversaries with anti-adversaries in training. In AAAI, pages 11435–11442, 2023.
  56. Understanding the interaction of adversarial training with noisy labels. arXiv preprint arXiv:2102.03482, 2021.
User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Xiaoling Zhou (9 papers)
  2. Wei Ye (110 papers)
  3. Zhemg Lee (2 papers)
  4. Rui Xie (59 papers)
  5. Shikun Zhang (82 papers)

Summary

We haven't generated a summary for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Summarize Now
X Twitter Logo Streamline Icon: https://streamlinehq.com

Tweets