Research map / Label-efficient and robust learning / Noisy labels
Self-Adaptive Training: beyond Empirical Risk Minimization
Lang Huang, Chao Zhang, Hongyang Zhang
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 self-adaptive training instead of ERM | ResNet-50 · ImageNet validation, 40% label noise | Top1 accuracy: 69.5 -> 71.5 | Improved |
| Use self-adaptive training instead of ERM | CIFAR10 · CIFAR10-C corruption levels 1–5 | average accuracy: 88.44 -> 89.41; 83.22 -> 84.53; 77.26 -> 78.83; 70.40 -> 71.90; 58.91 -> 60.77 | Improved |
| Replace TRADES cross-entropy term with self-adaptive training | TRADES · CIFAR10 test under white-box PGD-20 | robust accuracy: 1%~3% | Improved |
| Remove sample re-weighting (ablation) | Self-Adaptive Training · CIFAR10 and CIFAR100, 40% and 80% label noise | accuracy: 92.64 -> 92.49; 78.58 -> 78.10; 71.38 -> 69.52; 38.72 -> 36.78 | Hurt |
| Remove exponential moving average (ablation) | Self-Adaptive Training · CIFAR10 and CIFAR100, 40% and 80% label noise | accuracy: 92.64 -> 72.00; 78.58 -> 28.17; 71.38 -> 50.93; 38.72 -> 11.57 | Hurt |
| Set α to 0.6 instead of 0.9 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 90.17 | Hurt |
| Set α to 0.8 instead of 0.9 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 91.91 | Hurt |
| Set α to 0.95 instead of 0.9 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 92.54 | No clear effect |
| Set α to 0.99 instead of 0.9 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 84.38 | Hurt |
| Set E_s to 20 instead of 60 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 89.58 | Hurt |
| Set E_s to 40 instead of 60 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 91.89 | Hurt |
| Set E_s to 80 instead of 60 (ablation) | Self-Adaptive Training · CIFAR10, 40% label noise | accuracy: 92.64 -> 92.26 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SELF — SELF: Learning to Filter Noisy Labels with Self-Ensembling (ICLR 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Adaptive Training: Bridging Supervised and Self-Supervised Learning (TPAMI 2022 · similarity 0.94)
- The Marginal Value of Adaptive Gradient Methods in Machine Learning (NeurIPS 2017 · similarity 0.87)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.85)
- Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise (AAAI 2024 · similarity 0.85)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · similarity 0.85)
- Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR 2024 · similarity 0.84)
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.
