Research map / Label-efficient and robust learning / Noisy labels
Wasserstein Adversarial Regularization for Learning With Label Noise
Kilian Fatras, Bharath Bhushan Damodaran, Sylvain Lobry, Rémi Flamary, Devis Tuia, Nicolas Courty
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use word2vec-derived Wasserstein ground cost instead of KL divergence (AR) (ablation) | adversarial regularization · CIFAR-10 (40% noise) | test accuracy: 81.98 -> 84.76 | Improved |
| use word2vec-derived Wasserstein ground cost instead of KL divergence (AR) (ablation) | adversarial regularization · CIFAR-100 (40% noise) | test accuracy: 55.75 -> 58.86 | Improved |
| use word2vec-derived Wasserstein ground cost instead of KL divergence (AR) (ablation) | adversarial regularization · Fashion-MNIST (40% noise) | test accuracy: 89.74 -> 90.41 | Improved |
| use word2vec ground cost instead of uninformative 0-1 ground cost (ablation) | WAR · CIFAR-10 (40% noise) | test accuracy: 74.15 -> 84.76 | Improved |
| use word2vec ground cost instead of uninformative 0-1 ground cost (ablation) | WAR · CIFAR-100 (40% noise) | test accuracy: 51.00 -> 58.86 | Improved |
| use word2vec ground cost instead of pretrained-CNN centroid ground cost (ablation) | WAR · CIFAR-10 (40% noise) | test accuracy: 85.08 -> 84.76 | Hurt |
| use word2vec ground cost instead of pretrained-CNN centroid ground cost (ablation) | WAR · CIFAR-100 (40% noise) | test accuracy: 55.79 -> 58.86 | Improved |
| raise beta from 5 to 20 in adversarial regularization (KL) (ablation) | AR · CIFAR-10 (40% noise) | test accuracy: 81.37 -> 57.36 | Hurt |
| raise beta from 5 to 20 (ablation) | WAR_w2v · CIFAR-10 (40% noise) | test accuracy: 80.16 -> 86.73 | Improved |
| raise beta from 10 to 20 (ablation) | WAR_embed · CIFAR-10 (40% noise) | test accuracy: 85.08 -> 80.36 | Hurt |
| use WAR_w2v instead of CCE under 40% label noise | 9-layer CNN · CIFAR-100 | test accuracy: 42.45 -> 58.86 | Improved |
| use WAR_w2v instead of Pencil with clean labels (0% noise) | 9-layer CNN · CIFAR-100 | test accuracy: 63.32 -> 68.16 | Improved |
Closest papers by meaning
- Understanding Entropic Regularization in GANs (JMLR 2024 · similarity 0.88)
- Wasserstein Adversarial Examples via Projected Sinkhorn Iterations (ICML 2019 · similarity 0.88)
- Label Noise in Adversarial Training: A Novel Perspective to Study Robust Overfitting (NeurIPS 2022 · similarity 0.87)
- Wasserstein of Wasserstein Loss for Learning Generative Models (ICML 2019 · similarity 0.87)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.87)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
