Research map / Label-efficient and robust learning / Domain adaptation and generalization
Invertible Projection and Conditional Alignment 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 IPCA, an invertible-projection plus conditional-alignment method for multi-source blended-target domain adaptation (MBDA), where target domain labels are unavailable.
Problem
Multi-source domain adaptation when the target data is a blended mixture of unlabeled sub-target domains whose domain labels are unknown, so existing MSDA/BTDA methods lack sufficient target knowledge and alignment signals.
Approach
Builds an invertible neural network projection (IPM) between source and blended-target feature spaces, regularizes the original and projected features for consistency via conditional entropy maximization on their variation part, and performs discriminator-free adversarial alignment using the classifier as discriminator with a conditional kernel Bures (CKB) metric and pseudo-labels.
Key results
- Achieves 70.1% average accuracy on DomainNet MBDA tasks, versus 65.0% for DGWA.
- Achieves 73.7% average accuracy on Office-Home, versus 71.3% for DGWA.
- Achieves 83.6% average accuracy on ImageCLEF-DA, versus 81.7% for DGWA; removing the adversarial loss Ld drops DomainNet accuracy by 4.3 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IPCA instead of DGWA | MBDA · DomainNet | accuracy: 65.0 -> 70.1 | Helped |
| use IPCA instead of DGWA | MBDA · Office-Home | accuracy: 71.3 -> 73.7 | Helped |
| use IPCA instead of DIDA | MBDA · ImageCLEF-DA | accuracy: 82.2 -> 83.6 | Helped |
| add IPM projection (ablation) | IPCA · DomainNet MBDA | accuracy: 68.1 -> 70.1 | Helped |
| add conditional entropy loss Lent (ablation) | IPCA · DomainNet MBDA | accuracy: 69.2 -> 70.1 | Helped |
| add conditional adversarial alignment loss Ld (ablation) | IPCA · DomainNet MBDA | accuracy: 65.8 -> 70.1 | Helped |
| replace IPM with Random Augmentation (ablation) | IPCA · ImageCLEF-DA MBDA | accuracy: 83.6 -> 82.5 | Hurt |
| replace IPM with pAdaIN (ablation) | IPCA · ImageCLEF-DA MBDA | accuracy: 83.6 -> 82.9 | Hurt |
| replace IPM with MixStyle (ablation) | IPCA · ImageCLEF-DA MBDA | accuracy: 83.6 -> 82.7 | Hurt |
| vary β (ablation) | IPCA · ImageCLEF-DA b+i→c+p | accuracy | No clear effect |
| vary α (ablation) | IPCA · ImageCLEF-DA b+i→c+p | accuracy | Mixed |
Limitations
- Performance is somewhat sensitive to the adversarial loss weight alpha, per hyper-parameter analysis.
- Evaluation uses reconstructed MBDA protocols on standard datasets, where the target is a mix of two sub-domains.
Experimental setup
Datasets: DomainNet Office-Home ImageCLEF-DA
Benchmarks: DomainNet Office-Home ImageCLEF-DA
Metrics: accuracy
Baselines: DANN DALN SCDA MDAN DIDA MTDA MCDA AMDA DGWA HTA
Builds on: DALN DANN Invertible neural networks (Lugmayr et al. 2020) MixStyle-style feature augmentation consideration Domain Adaptive Neural Networks conditional alignment
Closest papers by meaning
- Style Adaptation and Uncertainty Estimation for Multi-Source Blended-Target Domain Adaptation (NeurIPS 2024 · similarity 0.90)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.89)
- Collaborative Semantic Consistency Alignment for Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.89)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (CVPR 2024 · similarity 0.88)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · 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.