Research map / Label-efficient and robust learning / Test-time adaptation
TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 TeSLA instead of the strongest listed baseline | test-time adaptation · VisDA-S N-O | mIoU: +5.7% | Helped |
| use TeSLA instead of the strongest listed baseline | test-time adaptation · VisDA-S N-M | mIoU: +6.1% | Helped |
| use flipped cross-entropy instead of basic cross-entropy (ablation) | VisDA-C self-learning · VisDA-C N-M | class-average accuracy improvement over Source: +16.6% -> +23.1% | Helped |
| add the marginal prediction entropy term to flipped cross-entropy (ablation) | VisDA-C self-learning · VisDA-C N-M | class-average accuracy improvement over Source: +23.1% -> +23.8% | Helped |
| add pseudo-label refinement to L_pl (ablation) | TeSLA · VisDA-C | class-average accuracy: 72.3% -> 81.6% (+9.3) | Helped |
| add adversarial-augmentation knowledge distillation to L_pl and PLR (ablation) | TeSLA · VisDA-C | class-average accuracy: +4.7 | Helped |
| combine weak-view ensembling and nearest-neighbor pseudo-label averaging (ablation) | TeSLA pseudo-label refinement · VisDA-C | class-average accuracy | Helped |
| replace learned adversarial augmentation with RandAugment (ablation) | TeSLA · VisDA-C N-O | class-average accuracy: 80.2 -> 81.5 | Hurt |
| increase adversarial augmentation sub-policy dimension from N=2 to N=3 (ablation) | TeSLA · VisDA-C N-O | class-average accuracy: 81.5 -> 82.2 | Helped |
| add TeSLA adversarial augmentation to TENT (ablation) | TENT · VisDA-C N-O | class-average accuracy: 66.5 -> 72.0 | Helped |
| add TeSLA adversarial augmentation to SHOT (ablation) | SHOT · VisDA-C N-O | class-average accuracy: 70.6 -> 73.1 | Helped |
| add TeSLA adversarial augmentation to AdaContrast (ablation) | AdaContrast · VisDA-C N-O | class-average accuracy: 76.9 -> 79.1 | Helped |
Closest papers by meaning
- Learning Loss for Test-Time Augmentation (NeurIPS 2020 · similarity 0.87)
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.85)
- Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach (NeurIPS 2024 · similarity 0.85)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.84)
- Annealing Self-Distillation Rectification Improves Adversarial Training (ICLR 2024 · similarity 0.83)
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double Perturbation (NeurIPS 2024 · similarity 0.83)
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.