Research map / Label-efficient and robust learning / Domain adaptation and generalization
Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer
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.
Introduces Multi-source Active Domain Adaptation (MADA) and the Detective framework, which combines a hypernetwork-based universal dynamic classifier, evidential domain/predictive uncertainty selection, and contextual diversity to pick target samples to annotate.
Problem
Active domain adaptation has only been studied with a single source domain; when labeled data come from multiple source domains with disparate distributions, existing ADA selection criteria fail due to multi-grained domain shift and miscalibrated predictive uncertainty.
Approach
A hypernetwork generates sample-conditioned classifier parameters so that multiple source domains are modeled as one approximate single source, and an evidential deep learning measure of domain uncertainty is integrated with predictive entropy and a feature-density diversity filter to select target samples for labeling.
Key results
- On Office-Home under MADA, Detective reaches 89.99% mean accuracy versus 83.03% for the best ADA baseline EADA.
- Ablation: removing the universal dynamic network drops mean accuracy by 6.13 points (89.99 to 83.86).
- Integrated uncertainty selector adds 3.71 points and contextual diversity calculator adds 1.19 points on Office-Home.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Detective instead of the strongest reported baseline | MADA with ResNet50 · Office-Home mean | accuracy: 83.03 -> 89.99 | Helped |
| add UDN (ablation) | Detective · Office-Home mean | accuracy: 6.13% | Helped |
| add IUS (ablation) | Detective · Office-Home mean | accuracy: 3.71% | Helped |
| add CDC (ablation) | Detective · Office-Home mean | accuracy: 1.19% | Helped |
| use GPG instead of LPS (ablation) | Detective dynamic layers · Office-Home ∪→A | accuracy: 85.60 -> 86.03 | Helped |
| use GPG instead of LPS (ablation) | Detective dynamic layers · Office-Home mean | accuracy: 88.86 -> 89.99 | Helped |
| use domain uncertainty only, without predictive uncertainty (ablation) | Detective IUS · Office-Home mean | accuracy: 86.93 vs. 89.99 | Hurt |
| use predictive uncertainty only, without domain uncertainty (ablation) | Detective IUS · Office-Home mean | accuracy: 88.96 vs. 89.99 | Hurt |
| combine domain and predictive uncertainty (ablation) | Detective IUS · Office-Home mean | accuracy: 89.99 | Helped |
Surprising findings
- Predictive uncertainty alone performs worse than domain uncertainty alone for selection, and the best domain-to-predictive uncertainty weighting favors domain uncertainty heavily (lambda_dom about 7.5 vs lambda_pre about 0.5).
Limitations
- Reported hyperparameters such as lambda_dom and lambda_pre are tuned on the Office-Home benchmark.
- Evaluation is limited to three image classification benchmarks with ResNet backbones.
Experimental setup
Datasets: Office-Home miniDomainNet Digits-five
Benchmarks: Office-Home miniDomainNet Digits-five
Metrics: accuracy mean accuracy
Baselines: MCD DCTN M3SDA MME DEAL MFSAN MDDA SImpAI MADAN CLUE TQS DUC
Builds on: Evidential Deep Learning HyperNetwork CLUE EADA Dynamic transfer
Closest papers by meaning
- Divide and Adapt: Active Domain Adaptation via Customized Learning (CVPR 2023 · similarity 0.93)
- Diffusion-Based Probabilistic Uncertainty Estimation for Active Domain Adaptation (NeurIPS 2023 · similarity 0.92)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.89)
- Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.88)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.88)
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.