Research map / Label-efficient and robust learning / Domain adaptation and generalization
LFME: A Simple Framework for Learning from Multiple Experts in Domain Generalization
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 LFME, which trains one target model alongside per-domain experts and regularizes the target's logits toward expert probabilities via an L2 term, improving ERM across DomainBed and semantic segmentation at no test-time cost.
Problem
Domain generalization methods often fail to consistently beat ERM across datasets, and naive use of domain experts requires aggregation design and extra inference cost when test domains are unknown.
Approach
Train one expert per source domain and a single target model simultaneously; regularize the target's logits toward the normalized expert probabilities with an MSE term, so the target becomes an expert on all sources and only the target is used at inference. The authors interpret this as a regression-style knowledge distillation that keeps logits in a smaller range (more information used) and re-we
Key results
- On DomainBed with ResNet18, LFME averages 60.8% across five datasets, the best among compared methods (ERM 58.1%), and improves TerraInc to 46.3% vs ERM 38.8%.
- With ResNet50, LFME averages 64.6% vs ERM 62.8% and SD 63.9%; with SWAD it reaches 68.5%.
- On GTAV+Synthia to real segmentation, LFME averages 38.37 mIoU vs baseline 32.38 and PinMem 38.74 mIoU avg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add expert-guided logit regularization to ERM | ResNet18 image classification · DomainBed average across PACS, VLCS, OfficeHome, TerraInc, and DomainNet | accuracy: 58.1 -> 60.8 | Helped |
| add expert-guided logit regularization to ERM | ResNet50 image classification · DomainBed average | accuracy: 62.8 -> 64.6 | Helped |
| add expert-guided logit regularization to the segmentation baseline | DeepLabv3+ with ResNet50 · Segmentation average | mIoU: 32.38 -> 38.37 | Helped |
| replace expert probabilities with one-hot labels in logit regularization (ablation) | ERM · PACS average | accuracy: 79.8 $\pm$ 0.4 -> 81.4 $\pm$ 0.5 | Helped |
| replace expert probabilities with one-hot labels in logit regularization (ablation) | ERM · TerraInc average | accuracy: 38.8 $\pm$ 1.0 -> 42.9 $\pm$ 0.7 | Helped |
| use standard label smoothing instead of LFME (ablation) | ERM · PACS and TerraInc | accuracy | Mixed |
| use expert-logit regression instead of probability-to-logit regularization (ablation) | LFME · PACS average | accuracy: 79.8$\pm$0.4 -> 79.9$\pm$0.2 | No clear effect |
| use probability-to-expert-logit regression instead of LFME regularization (ablation) | LFME · PACS average | accuracy: 79.8$\pm$0.4 -> 79.7$\pm$0.5 | Hurt |
| use probability-to-probability regression instead of LFME regularization (ablation) | LFME · PACS average | accuracy: 79.8$\pm$0.4 -> 81.1$\pm$1.1 | Helped |
| use teacher-probability cross-entropy distillation instead of LFME regularization (ablation) | LFME · PACS average | accuracy: 79.8$\pm$0.4 -> 80.4$\pm$0.4 | Helped |
| aggregate expert predictions at inference instead of distilling them into one target model (ablation) | LFME · PACS average | accuracy: 82.4$\pm$0.1 vs 76.9$\pm$0.1 | Hurt |
| increase alpha/2 from 0 to 0.01 (ablation) | LFME · PACS average | accuracy: 79.8 $\pm$ 0.4 -> 81.8 $\pm$ 0.2 | Helped |
Surprising findings
- Simply adding an MSE term with one-hot labels (ERM+) already improves ERM across all unseen domains, a 'free lunch' for DG.
- Naive aggregation of experts (averaging, model soup, confidence-based selection, learned weighting) fails to improve over the baseline.
Limitations
- Requires domain labels for training data, so it cannot be used when domain labels are unavailable.
- Not applicable to single-source domain generalization.
- Training cost roughly doubles compared with ERM since one expert per domain is trained.
Experimental setup
Datasets: PACS VLCS OfficeHome TerraInc DomainNet GTAV Synthia Cityscapes BDD100K Mapillary
Benchmarks: PACS VLCS OfficeHome TerraInc DomainNet Cityscapes BDD100K Mapillary
Metrics: accuracy mIoU mAcc top5 count score
Baselines: ERM MMD RSC IRM DANN CORAL SD Fish MIRO Meta-DMoE ITTA RIDG
Builds on: Meta-DMoE DAELDG knowledge distillation DomainBed SWAD
Closest papers by meaning
- Is Large-scale Pretraining the Secret to Good Domain Generalization? (ICLR 2025 · similarity 0.87)
- PEER Pressure: Model-to-Model Regularization for Single Source Domain Generalization (CVPR 2025 · similarity 0.86)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.86)
- Improving Domain Generalization with Domain Relations (ICLR 2024 · similarity 0.86)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.86)
- DOGE: Domain Reweighting with Generalization Estimation (ICML 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.