Research map / Label-efficient and robust learning / Domain adaptation and generalization
Sharpness-Aware Model-Agnostic Long-Tailed 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.
Adds SAM-style sharpness-aware minimization and a maximum squares loss to domain generalization methods (DG via ER, EISNet, FACT) to improve accuracy on unseen domains and long-tailed class distributions.
Problem
Domain generalization methods converge to sharp minima and suffer from class imbalance/long-tailed label distributions, hurting performance on unseen target domains.
Approach
Plug-and-play combination of sharpness-aware minimization applied to the classification objective with a maximum squares loss (replacing entropy-style losses) that yields more balanced gradients across easy and hard classes, added on top of existing DG algorithms.
Key results
- Adding the method to FACT with ResNet50 raises average accuracy across the four benchmarks from 67.96% to 69.88%.
- DG via ER + Ours reaches 83.58% average on PACS with ResNet18, versus 81.32% for DG via ER alone.
- Ablation on PACS with ResNet18: SAM alone adds 0.99 points and max squares loss alone adds 1.38 points over DG via ER; combined gives 2.26.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SAM and maximum square loss to FACT | FACT with ResNet18 · PACS | accuracy: 84.29 -> 85.36 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet18 · OfficeHome | accuracy: 61.89 -> 64.53 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet18 · DigitDG | accuracy: 80.38 -> 83.24 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet18 · DomainNet | accuracy: 43.63 -> 45.22 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet50 · PACS | accuracy: 87.94 -> 90.08 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet50 · OfficeHome | accuracy: 66.92 -> 69.29 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet50 · DigitDG | accuracy: 81.34 -> 83.11 | Helped |
| Add SAM and maximum square loss to FACT | FACT with ResNet50 · DomainNet | accuracy: 49.53 -> 50.97 | Helped |
| Add SAM (ablation) | DG via ER with ResNet18 · PACS | accuracy: 81.32 -> 82.31 | Helped |
| Add maximum square loss (ablation) | DG via ER with ResNet18 · PACS | accuracy: 81.32 -> 82.70 | Helped |
| Add SAM and maximum square loss (ablation) | DG via ER with ResNet18 · PACS | accuracy: 81.32 -> 83.58 | Helped |
Limitations
- Long-tailed aspect is motivated by dataset distributions but not evaluated with long-tailed-specific benchmarks or metrics.
- Theoretical analysis is only sketched; the maximum squares loss is derived for binary classification.
- Gains are modest on some settings (e.g., DigitDG and EISNet on DomainNet) and hyperparameter sensitivity is shown only on PACS with ResNet18.
Experimental setup
Datasets: PACS OfficeHome DigitDG DomainNet
Benchmarks: PACS OfficeHome DigitDG DomainNet
Metrics: accuracy
Baselines: DG via ER EISNet FACT DeepAll
Builds on: Sharpness-Aware Minimization DG via ER MaxSquare EISNet FACT
Closest papers by meaning
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.89)
- Set Valued Predictions For Robust Domain Generalization (ICML 2025 · similarity 0.89)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.88)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.88)
- Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · similarity 0.88)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · 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.