Research map / Label-efficient and robust learning / Domain adaptation and generalization
MLNet: Mutual Learning Network with Neighborhood Invariance for Universal 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 MLNet, a universal domain adaptation method combining neighborhood-invariance feature learning, cross-domain manifold mixup for unknown-class simulation, and a closed-set/open-set consistency constraint, achieving state-of-the-art H-scores on Office, OfficeHome and VisDA in the OPDA setting.
Problem
Universal domain adaptation, where the label-set relation between source and target domains is unknown, requiring target samples to be classified as a known class or unknown; existing methods overlook intra-domain variation in the target and struggle to separate similar known and unknown classes.
Approach
Builds on OVANet by adding (i) neighborhood invariance learning in the target domain with self-adaptive nearest-neighbor selection and Jaccard-based confidence weights, (ii) cross-domain manifold mixup between source and target features to simulate unknown-class samples for open-set classifiers, and (iii) a consistency constraint tying closed-set probabilities to open-set positive scores so that m
Key results
- In the OPDA setting, MLNet reaches an average H-score of 92.8 on Office, 77.4 on OfficeHome and 69.9 on VisDA, the best among compared methods.
- Compared with the OVANet-style baseline, MLNet improves average CDA/PDA accuracy by 4.3 on Office, 9.2 on OfficeHome and 17.0 on VisDA.
- Against UniOT in OPDA, MLNet improves by about 1% on Office and OfficeHome and 12.6% on VisDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MLNet to the OVANet baseline | OVANet / ResNet-50 · Office, OfficeHome, and VisDA; CDA, PDA, ODA, and OPDA averages | accuracy or H-score: +4.3; +9.2; +17.0 (accuracy averages) and +3.7; +4.0; +11.3 (H-score averages) | Helped |
| Add neighborhood invariance learning (NIL) (ablation) | MLNet · Office ODA | H-score: 17.1 -> 92.8 | Helped |
| Add neighborhood invariance learning (NIL) (ablation) | MLNet · Office OPDA | H-score: 35.3 -> 92.8 | Helped |
| Add cross-domain manifold mixup (CMM) (ablation) | MLNet · Office ODA | H-score: 41.3 -> 92.8 | Helped |
| Add cross-domain manifold mixup (CMM) (ablation) | MLNet · Office OPDA | H-score: 50.2 -> 92.8 | Helped |
| Add consistency constraint (CC) (ablation) | MLNet · Office ODA | H-score: 6.9 -> 92.8 | Helped |
| Add consistency constraint (CC) (ablation) | MLNet · Office OPDA | H-score: 75.7 -> 92.8 | Helped |
| Add neighbor confidence weighting (ablation) | MLNet · Office ODA | H-score: 90.8 -> 92.8 | Helped |
| Replace K-nearest neighbors with self-adaptive neighbor search (ablation) | MLNet · Office OPDA | H-score: 77.0 -> 92.8 | Helped |
| Replace source-domain manifold mixup (SMM) with cross-domain manifold mixup (CMM) (ablation) | MLNet · Office OPDA | H-score: 72.7 -> 92.8 | Helped |
| Use the proposed consistency constraint instead of cosine, JS-divergence, or KL-divergence constraint (ablation) | MLNet · Office OPDA | H-score: 84.9; 90.4; 90.8; 89.2 -> 92.8 | Helped |
| Use consistency constraint rather than no CC (ablation) | MLNet on VisDA OPDA · VisDA OPDA | known-class recall: 9.64% to 62.36%(+52.72) | Helped |
Surprising findings
- Removing the consistency constraint drops ODA H-score on Office from 92.8 to 6.9, showing the loss is critical for balancing known and unknown identification.
- Using K-nearest-neighbor selection instead of self-adaptive neighbor search degrades H-score on Office ODA from 92.8 to 62.1.
Limitations
- Hyperparameters beta2 and eta are grid-searched on Office and the weight eta must be doubled for VisDA, so tuning per dataset is needed.
- Gains in CDA and some ODA settings are only competitive rather than best; MLNet is slightly inferior to GATE in CDA.
- The mixup probability argument assumes a simplified class-overlap model and the decision threshold of 0.5 is fixed.
Experimental setup
Datasets: Office OfficeHome VisDA
Benchmarks: Office OfficeHome VisDA
Metrics: accuracy H-score
Baselines: UAN CMU DANCE DCC OVANet GATE MATHS TNT UniOT CPR NCAL
Builds on: OVANet GATE manifold mixup DANCE
Closest papers by meaning
- Mutual Nearest Neighbor Contrast and Hybrid Prototype Self-Training for Universal Domain Adaptation (AAAI 2022 · similarity 0.90)
- Divergence Optimization for Noisy Universal Domain Adaptation (CVPR 2021 · similarity 0.89)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.87)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.86)
- Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.86)
- Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation (ICML 2025 · 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.