Research map / Label-efficient and robust learning / Domain adaptation and generalization
Proxy Denoising for Source-Free 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 ProDe, which treats CLIP predictions as a noisy proxy for a domain-invariant space, denoises them with a proxy confidence-based logit correction, and distills them via a mutual-information objective for source-free domain adaptation.
Problem
Vision-language (CLIP) pseudo-supervision used in source-free domain adaptation is noisy and inaccurate at unknown rates, which harms adaptation.
Approach
Model the reliability of CLIP predictions via a proxy confidence theory tied to the in-training model's divergence from the source model, then correct CLIP logits by subtracting the source-vs-target logit difference, and train with a mutual-information distillation loss plus pseudo-label refinement.
Key results
- ProDe-V reaches 92.6% average accuracy on Office-31 and 84.5% on Office-Home, versus 92.5% for DIFO-V on Office-31 and 83.1% on Office-Home.
- ProDe-V achieves 85.0% average accuracy on DomainNet-126, versus 80.0% for DIFO-V.
- Removing proxy denoising reduces ProDe-V average accuracy by 1.1% (89.3 to 88.2) in the ablation study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProDe-V instead of DIFO-V | CLIP ViT-B/32 SFDA · Office-31 closed-set (Avg.) | accuracy: 92.5 -> 92.6 | Helped |
| Use ProDe-V instead of DIFO-V | CLIP ViT-B/32 SFDA · Office-Home closed-set (Avg.) | accuracy: 83.1 -> 84.5 | Helped |
| Use ProDe-V instead of DIFO-V | CLIP ViT-B/32 SFDA · VisDA closed-set (Sy→Re) | per-class accuracy: 90.3 -> 91.0 | Helped |
| Use ProDe-V instead of DIFO-V | CLIP ViT-B/32 SFDA · DomainNet-126 closed-set (Avg.) | accuracy: 80.0 -> 85.0 | Helped |
| Add L_Apt and L_Ref jointly instead of using either loss alone (ablation) | ProDe-V · Office-31 (Avg. across benchmarks) | accuracy: 86.5 -> 89.3 | Helped |
| Replace mutual information with KL divergence (ablation) | ProDe-V · Office-31, Office-Home, VisDA (Avg.) | accuracy: 89.3 -> 82.1 | Hurt |
| Add proxy denoising (ablation) | ProDe-V · Office-31, Office-Home, VisDA (Avg.) | accuracy: 88.2 -> 89.3 | Helped |
| Retain source-model logits in proxy denoising (ablation) | ProDe-V · Office-31, Office-Home, VisDA (Avg.) | accuracy: 88.6 -> 89.3 | Helped |
| Retain target-model logits in proxy denoising (ablation) | ProDe-V · Office-31, Office-Home, VisDA (Avg.) | accuracy: 87.7 -> 89.3 | Helped |
| Correct proxy predictions at the logit level rather than probability level (ablation) | ProDe-V · Office-31, Office-Home, VisDA (Avg.) | accuracy: 86.2 -> 89.3 | Helped |
Surprising findings
- Zero-shot CLIP is substantially inferior to the proposed method, supporting the claim that CLIP predictions are noisy supervision.
- Replacing mutual information with KL divergence in the adaptation loss causes a 7.2% average drop.
Limitations
- Gamma must be tuned per dataset scale, which the paper notes is sensitive.
- Theoretical justification relies on approximations (Gaussian approximation, KL-based space distances, and P(I) assumed to equal 1), which are loosely grounded.
- ViT-B/16 results on Office-31 trail the DPC baseline, which the authors attribute to possible overfitting.
Experimental setup
Datasets: Office-31 Office-Home VisDA DomainNet-126
Benchmarks: Office-31 Office-Home VisDA DomainNet-126
Metrics: accuracy average accuracy harmonic mean accuracy
Baselines: Source CLIP zero-shot DIFO-R DIFO-V DAPL-R PADCLIP-R ADCLIP-R PDA-R DAMP-R SHOT NRC GKD
Builds on: SHOT CLIP DIFO DAPL
Closest papers by meaning
- Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (CVPR 2024 · similarity 0.90)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.90)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.90)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.88)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · similarity 0.87)
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.