Research map / Label-efficient and robust learning / Domain adaptation and generalization
Advancing Open-Set Domain Generalization Using Evidential Bi-Level Hardest Domain Scheduler
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 EBiL-HaDS, an adaptive domain scheduler for meta-learning-based open-set domain generalization that picks the hardest domain each step using a bi-level follower network trained with max-rebiased evidential learning.
Problem
Open-set domain generalization (OSDG) requires generalizing to unseen domains while recognizing unseen categories; existing meta-learning OSDG methods use a predefined sequential domain scheduler, whose effect on training is unexplored.
Approach
Replace the fixed sequential domain scheduler with an adaptive one: a follower network (trained bi-level with confidence targets from max rebiased discrepancy evidential learning) estimates per-domain reliability, and the least reliable (hardest) domain is chosen to form meta-train/meta-test partitions.
Key results
- On PACS with ResNet18, EBiL-HaDS-bcls reaches 80.34% average OSCR versus 77.89% for MEDIC-bcls.
- On Digits-DG with ConvNet, EBiL-HaDS-bcls reaches 75.87% average OSCR, and beats a random domain scheduler by 4.63 points in OSCR.
- On PACS with ResNet50, EBiL-HaDS-bcls reaches 86.12% average OSCR versus 84.85% for MEDIC-bcls.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add EBiL-HaDS with a binary classification head | PACS, ResNet18 · PACS avg | OSCR: 2.45% | Helped |
| add EBiL-HaDS with a binary classification head | PACS, ResNet50 · PACS avg | OSCR: 1.27% | Helped |
| add EBiL-HaDS | Digits-DG, ConvNet, binary classification head · Digits-DG avg | OSCR: 5.64% | Helped |
| add EBiL-HaDS | OfficeHome, ResNet18 · OfficeHome avg | OSCR: null | Helped |
| add max rebiased discrepancy evidential learning (ablation) | EBiL-HaDS, Digits-DG, conventional classification head · Digits-DG avg | OSCR: 5.07% | Helped |
| add max rebiased discrepancy evidential learning (ablation) | EBiL-HaDS, Digits-DG, binary classification head · Digits-DG avg | OSCR: 3.89% | Helped |
| add deep evidential learning (ablation) | EBiL-HaDS without RBE, Digits-DG, conventional classification head · Digits-DG avg | OSCR: 2.47% | Helped |
| add rebiased discrepancy regularization term (ablation) | EBiL-HaDS without RB, Digits-DG, conventional classification head · Digits-DG avg | OSCR: 2.60% | Helped |
| use adaptive hardest-domain scheduling instead of sequential scheduling (ablation) | Digits-DG, ConvNet, binary classification head · Digits-DG avg | OSCR: 3.89% | Helped |
| use adaptive hardest-domain scheduling instead of random scheduling (ablation) | Digits-DG, ConvNet, binary classification head · Digits-DG avg | OSCR: 5.77% | Helped |
| use follower-network reliability instead of main-network confidence for scheduling (ablation) | EBiL-HaDS, Digits-DG, binary classification head · Digits-DG avg | OSCR: 3.29% | Helped |
| select the hardest rather than the easiest domain (ablation) | EBiL-HaDS, Digits-DG · Digits-DG avg | OSCR | Helped |
Surprising findings
- Some baselines trained with a ResNet152 backbone show OSCR decay relative to ResNet50, which the authors attribute to overfitting in OSDG.
Limitations
- Relies on source domains with unified category sets.
- Only tested on image classification.
- May still produce misclassification and biased predictions.
Experimental setup
Datasets: PACS Digits-DG OfficeHome
Benchmarks: PACS Digits-DG OfficeHome
Metrics: accuracy H-score OSCR
Baselines: ERM ARPL MLDG MEDIC ODG-Net SWAD MixStyle MVDG OpenMax RSC DAML SelfReg
Builds on: MLDG MEDIC ARPL evidential deep learning
Closest papers by meaning
- Set Valued Predictions For Robust Domain Generalization (ICML 2025 · similarity 0.87)
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.87)
- Towards Generalizing to Unseen Domains with Few Labels (CVPR 2024 · similarity 0.87)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.87)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.86)
- Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 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.