Research map / Label-efficient and robust learning / Domain adaptation and generalization
Safe Self-Refinement for Transformer-Based Domain Adaptation
Check 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 SSRT self-refinement and Safe Training instead of Baseline-B | ViT-B/16 UDA · Office-Home | average accuracy: 81.05 -> 85.43 | Helped |
| Use SSRT self-refinement and Safe Training instead of Baseline-B | ViT-B/16 UDA · VisDA-2017 | average accuracy: 85.23 -> 88.76 | Helped |
| Use SSRT self-refinement and Safe Training instead of Baseline-B | ViT-B/16 UDA · DomainNet | average accuracy: 38.5 -> 45.2 | Helped |
| Use latent token perturbation rather than raw-input perturbation (ablation) | SSRT-B · Office-Home | average accuracy: 85.0 -> 85.4 | Helped |
| Use latent token perturbation rather than raw-input perturbation (ablation) | SSRT-B · DomainNet | average accuracy: 44.2 -> 45.2 | Helped |
| Choose perturbation layers from multiple depths instead of a single fixed layer (ablation) | SSRT-B · DomainNet | average accuracy: -1.0%, -1.5% and -1.5% | Helped |
| Combine both KL supervision directions rather than use only one direction (ablation) | SSRT-B · DomainNet | average accuracy: 41.1 -> 41.8; 40.1 -> 41.8 | Helped |
| Add Safe Training to bidirectional self-refinement (ablation) | SSRT-B · DomainNet | average accuracy: 41.8 -> 43.4 | Helped |
| Allow gradients through both teacher probabilities and the latent input token sequence (ablation) | SSRT-B · Office-Home, Pr→Cl | accuracy: 74.98 -> 71.32; 72.60 -> 71.32 | Helped |
| Add a confidence filter for target predictions (ablation) | SSRT-B · Office-Home, Pr→Cl | accuracy: 20.30 -> 67.79 | Helped |
| Add Safe Training to restore checkpoints and adjust perturbation and loss weights (ablation) | SSRT-B · Office-Home, Pr→Cl | accuracy: 20.30 -> 67.79 | Helped |
| Add random classifier-head dropout when the perturbation scalar is zero (ablation) | SSRT-B · Office-Home and DomainNet | accuracy | Helped |
Closest papers by meaning
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.87)
- CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation (ICLR 2022 · similarity 0.87)
- Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.86)
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.85)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.85)
- Split Adaptation for Pre-trained Vision Transformers (CVPR 2025 · similarity 0.85)
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.