Research map / Label-efficient and robust learning / Domain adaptation and generalization
Prompt-Based Distribution Alignment for Unsupervised Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PDA, a two-branch prompt-tuning method for CLIP-based unsupervised domain adaptation that adds a feature-bank attention module (IFT) to reduce source-target distribution discrepancy.
Problem
Unsupervised domain adaptation with large vision-language models, where prompt tuning alone does not explicitly address domain shift between source and target domains.
Approach
A base branch learns class-discriminative multi-modal prompts with supervised and pseudo-label contrastive losses, while an alignment branch uses image-guided feature tuning (IFT) to attend from image features to source and target feature banks built from zero-shot CLIP confident samples, integrating cross-domain features to minimize domain discrepancy.
Key results
- On Office-Home with ViT-B/16, PDA reaches 85.7% average accuracy, versus 84.2% for MaPLe and 82.1% for zero-shot CLIP.
- On Office-31 with ViT-B/16, PDA reaches 91.2% average accuracy, improving 13.7 points over zero-shot CLIP (77.5%).
- On VisDA-2017 with ViT-B/16, PDA reaches 89.7% average accuracy, outperforming SSRT (88.8% with ResNet101 backbone) and CDTrans-Deit (88.4%) in the paper's comparison tables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the L_x source-classification loss (ablation) | PDA · Office-Home | accuracy: +2.1 | Helped |
| Add the L_x source-classification loss (ablation) | PDA · Office-31 | accuracy: +12.1 | Helped |
| Add the L_x source-classification loss (ablation) | PDA · VisDA-2017 | accuracy: -5.4 | Hurt |
| Add the L_xa alignment loss to L_x (ablation) | PDA · Office-Home | accuracy: +2.5 | Helped |
| Add the L_xa alignment loss to L_x (ablation) | PDA · Office-31 | accuracy: +12.3 | Helped |
| Add the L_xa alignment loss to L_x (ablation) | PDA · VisDA-2017 | accuracy: -3.7 | Hurt |
| Add the L_u target pseudo-label loss (ablation) | PDA with L_x and L_xa · VisDA-2017 | accuracy: -3.7 -> +0.1 | Helped |
| Add the L_ua alignment pseudo-label loss (ablation) | PDA with L_x, L_xa, and L_u · Office-Home | accuracy: +3.1 -> +3.6 | Helped |
| Add the L_ua alignment pseudo-label loss (ablation) | PDA with L_x, L_xa, and L_u · Office-31 | accuracy: +13.0 -> +13.7 | Helped |
| Add the L_ua alignment pseudo-label loss (ablation) | PDA with L_x, L_xa, and L_u · VisDA-2017 | accuracy: +0.1 -> +0.8 | Helped |
| Use beta_s=0.01 and beta_t=0.1 for augmented image features (ablation) | PDA · Office-Home | accuracy: 85.4 -> 85.8 | Helped |
| Increase augmented-feature weights to beta_s=10 and beta_t=10 (ablation) | PDA · Office-Home | accuracy: 79.7 | Hurt |
Surprising findings
- Most prompt tuning methods (e.g., CoOp, MaPLe) underperform zero-shot CLIP on VisDA-2017 with some backbones, attributed to overfitting.
- Ablation shows L_x alone lowers VisDA-2017 accuracy (88.9 to 83.5) and only later losses recover it.
Limitations
- Pseudo-label threshold and warm-up strategy differ by dataset (e.g., Office-31 needs warm-up because zero-shot CLIP accuracy is low), requiring dataset-specific choices.
- Tuning prompts can overfit on some datasets (e.g., VisDA-2017), making the method sensitive to training strategy.
Experimental setup
Datasets: Office-31 Office-Home VisDA-2017
Benchmarks: Office-Home Office-31 VisDA-2017
Metrics: top-1 accuracy average accuracy MMD KL divergence inter-class/inner-class L2 distance ratio r
Baselines: zero-shot CLIP linear probe CLIP CoOp CoCoOp VP VPT IVLP MaPLe DAPL DANN JAN MDD
Builds on: CLIP MaPLe CoOp VPT
Closest papers by meaning
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.95)
- Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation (CVPR 2024 · similarity 0.92)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.91)
- Amend to Alignment: Decoupled Prompt Tuning for Mitigating Spurious Correlation in Vision-Language Models (ICML 2024 · similarity 0.89)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.89)
- Adapting to Distribution Shift by Visual Domain Prompt Generation (ICLR 2024 · similarity 0.89)
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.