Research map / Label-efficient and robust learning / Domain adaptation and generalization
Style Adaptation and Uncertainty Estimation for Multi-Source Blended-Target 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 SAUE, a method for multi-source blended-target domain adaptation that augments source features with similarity-weighted target style statistics, adds Dirichlet-based uncertainty elimination, and uses a discriminator-free adversarial alignment without domain labels.
Problem
Blended-target domain adaptation with a single source provides insufficient feature information and unknown target domain labels; the paper targets adaptation from multiple labeled source domains to a blended unlabeled target domain.
Approach
Uses Wasserstein-distance-based similarity factors to select and weight blended-target channel-wise style statistics for augmenting normalized source low-level features, adds an evidential (Dirichlet) uncertainty loss with KL divergence for misclassified source samples, and aligns domains with a classifier-based nuclear-norm Wasserstein discrepancy without a discriminator or domain labels.
Key results
- On the DomainNet MBDA setting (ResNet-50), SAUE reaches 70.9% average accuracy vs. 63.8% for MCDA and 63.4% for AMDA.
- On Office-Home, SAUE achieves 73.7% average accuracy, versus 72.1% for DGWA and 71.9% for HTA.
- On ImageCLEF-DA, SAUE achieves 84.3% average accuracy, versus 82.7% for DGWA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAUE instead of DGWA | MBDA with ResNet-50 · DomainNet (Table 2 configuration; average) | accuracy: 65.0 -> 70.9 | Helped |
| Use SAUE instead of DGWA | MBDA with ResNet-50 · Office-Home (average) | accuracy: 72.1 -> 73.7 | Helped |
| Use SAUE instead of DGWA | MBDA with ResNet-50 · ImageCLEF-DA (average) | accuracy: 82.7 -> 84.3 | Helped |
| Use SAUE instead of DGWA | MBDA with ResNet-101 · DomainNet C+P+Q→R+S (average) | accuracy: 47.4 -> 50.2 | Helped |
| Use SAUE instead of DGWA | MBDA with ResNet-101 · DomainNet R+S→C+P+Q (average) | accuracy: 45.8 -> 48.2 | Helped |
| Use SAUE instead of DALN | MBDA with ResNet-101 · VisDA-2017 Syn.→Rel. | accuracy: 80.6 -> 81.5 | Helped |
| Add style adaptation (SA) (ablation) | SAUE on DomainNet · DomainNet (Table 6 configuration; average) | accuracy: 67.6 -> 70.9 | Helped |
| Add uncertainty loss Lunc (ablation) | SAUE on DomainNet · DomainNet (Table 6 configuration; average) | accuracy: 69.9 -> 70.9 | Helped |
| Add both style adaptation and uncertainty loss Lunc (ablation) | SAUE on DomainNet · DomainNet (Table 6 configuration; average) | accuracy: 66.5 -> 70.9 | Helped |
| Select target styles using Wasserstein distance and similarity weights instead of random selection (ablation) | SAUE style adaptation on DomainNet · DomainNet (Table 6 configuration; average) | accuracy: 70.1 -> 70.9 | Helped |
Surprising findings
- Style adaptation gains are larger with ResNet-50 (+2.8%) than with ResNet-101 (+1.6%) on DomainNet.
- Performance is very sensitive to the annealing coefficient lambda_e, with too-small values causing substantial degradation.
Limitations
- Authors note potential failure under larger distribution shifts, such as more source or target domains.
- Sensitive to the annealing hyperparameter lambda_e.
Experimental setup
Datasets: ImageCLEF-DA Office-Home DomainNet VisDA 2017
Benchmarks: ImageCLEF-DA Office-Home DomainNet VisDA 2017
Metrics: accuracy
Baselines: MCD DAN TSA DALN BIWAA SCDA MDAN DCTN DIDA MTDA-ITA MCDA AMDA
Builds on: EDL (evidential deep learning) MCDA DALN DANN AMDA
Closest papers by meaning
- Blending-Target Domain Adaptation by Adversarial Meta-Adaptation Networks (CVPR 2019 · similarity 0.91)
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.90)
- Class Overwhelms: Mutual Conditional Blended-Target Domain Adaptation (AAAI 2023 · similarity 0.88)
- Adversarial Style Mining for One-Shot Unsupervised Domain Adaptation (NeurIPS 2020 · similarity 0.88)
- CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.87)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · similarity 0.86)
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.