Research map / Label-efficient and robust learning / Domain adaptation and generalization
Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised 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 LUHP, an instance contrastive learning method for unsupervised domain adaptation that filters target anchors and samples by a category uncertainty weight and trains on high-training-potential positive/negative pairs, reporting SOTA-level accuracy on four UDA benchmarks.
Problem
Instance contrastive learning for unsupervised domain adaptation selects positive and negative samples only by pseudo-label agreement, so uncertain near-boundary samples with noisy pseudo-labels and low-training-potential samples harm training.
Approach
A category uncertainty weight (ratio of distances to different-class vs same-class kNN neighbors) sets per-class thresholds to filter uncertain target samples; a weighted multi-sample triplet loss uses only the K hardest high-potential positives and negatives; low-uncertainty anchors are reused via mixup.
Key results
- Achieves 75.4% average accuracy on Office-Home, about 2.8% above the prior SOTA reported in the paper.
- Reaches 84.6% average accuracy on VisDA-2017, about 3% above CaCo.
- Averages 82.0% on DomainNet (40-class subset, 12 tasks) and 90.7% on Office-31, essentially tied with MSGD (90.8%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LUHP instead of MSGD | UDA with ResNet-50 · Office-31 | accuracy: 90.8 -> 90.7 | Hurt |
| Use LUHP instead of MSGD | UDA with ResNet-50 · Office-Home | accuracy: 72.4 -> 75.4 | Helped |
| Use LUHP instead of MSGD | UDA with ResNet-101 · VisDA-2017 | accuracy: 84.5 -> 84.6 | Helped |
| Use LUHP instead of KUDA | UDA with ResNet-50 · DomainNet | accuracy: 81.0 -> 82.0 | Helped |
| Add LK to the configuration with no listed components (ablation) | LUHP · Office-Home | accuracy: 57.6 -> 73.7 | Helped |
| Add LT to LK (ablation) | LUHP · Office-Home | accuracy: 73.7 -> 74.1 | Helped |
| Add category uncertainty weights to LK + LT + sample reuse (ablation) | LUHP · Office-Home | accuracy: 74.0 -> 75.4 | Helped |
| Add sample reuse to LK + LT + category uncertainty weights (ablation) | LUHP · Office-Home | accuracy: 74.4 -> 75.4 | Helped |
| Add LT to LK + category uncertainty weights + sample reuse (ablation) | LUHP · DomainNet | accuracy: 81.1 -> 82.0 | Helped |
| Add category uncertainty weights to LK + LT + sample reuse (ablation) | LUHP · DomainNet | accuracy: 81.1 -> 82.0 | Helped |
| Add sample reuse to LK + LT + category uncertainty weights (ablation) | LUHP · DomainNet | accuracy: 81.1 -> 82.0 | Helped |
Surprising findings
- On Office-31 the average accuracy is 0.1% below MSGD, so the method is not uniformly best across benchmarks.
Limitations
- Relies on a noisy pseudo-label estimate from kNN at the start of training, with hyperparameters (K, alpha) tuned on a subset of tasks.
- Office-31 results are not best overall; gains are concentrated on Office-Home, VisDA and DomainNet.
Experimental setup
Datasets: Office-31 Office-Home VisDA-2017 DomainNet
Benchmarks: Office-31 Office-Home VisDA-2017 DomainNet
Metrics: classification accuracy (%)
Baselines: DANN BNM BCDM ATDOC ILA-DA CaCo DALN UTEP ELS BIWAA-I MSGD COAL
Builds on: ATDOC ILA-DA CaCo DANN Triplet Loss mixup
Closest papers by meaning
- Unsupervised Domain Adaptation With Similarity Learning (CVPR 2018 · similarity 0.90)
- Mutual Nearest Neighbor Contrast and Hybrid Prototype Self-Training for Universal Domain Adaptation (AAAI 2022 · similarity 0.89)
- Guiding Pseudo-Labels With Uncertainty Estimation for Source-Free Unsupervised Domain Adaptation (CVPR 2023 · similarity 0.89)
- Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments (AAAI 2024 · similarity 0.88)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.88)
- Category-Aware Active Domain Adaptation (ICML 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.