Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unknown Domain Inconsistency Minimization for 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 UDIM, which adds a data-space and parameter-space perturbation term that minimizes loss-landscape inconsistency between the source domain and emulated unknown domains, combined with SAM-family optimizers for domain generalization.
Problem
Domain generalization needs models that transfer from source domains to unseen domains; existing SAM variants only seek flat minima on the source domain and leave the unknown-domain risk term unoptimized.
Approach
Emulate unknown domains by adversarially perturbing source inputs in a closed form, then minimize the loss-landscape inconsistency (gradient-variance matching of per-sample gradients) between the perturbed and original source data, added to SAM-style objectives.
Key results
- On CIFAR-10-C, UDIM w/ GAM reaches 75.3% average accuracy versus 72.8% for GAM alone, and UDIM w/ SAM reaches 74.2% versus 71.8% for SAM.
- On PACS single-source DG, UDIM w/ SAM averages 74.7% versus 64.8% for SAM, and UDIM w/ GAM averages 74.9% versus 65.4% for GAM.
- Authors report accuracy gains up to 9.9% in some settings when adding UDIM to sharpness-based approaches.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UDIM to SAM | SAM · CIFAR-10-C | average accuracy: 71.8 -> 74.2 | Helped |
| add UDIM to SAGM | SAGM · CIFAR-10-C | average accuracy: 72.0 -> 73.9 | Helped |
| add UDIM to GAM | GAM · CIFAR-10-C | average accuracy: 72.8 -> 75.3 | Helped |
| add UDIM to SAM | SAM · PACS, LOODG | average accuracy: 85.9 -> 88.7 | Helped |
| add UDIM to SAGM | SAGM · PACS, LOODG | average accuracy: 85.3 -> 88.0 | Helped |
| add UDIM to GAM | GAM · PACS, LOODG | average accuracy: 86.0 -> 88.1 | Helped |
| add UDIM to SAM | SAM · PACS, SDG | average accuracy: 64.8 -> 74.7 | Helped |
| add UDIM to SAGM | SAGM · PACS, SDG | average accuracy: 64.8 -> 74.5 | Helped |
| add UDIM to GAM | GAM · PACS, SDG | average accuracy: 65.4 -> 74.9 | Helped |
Surprising findings
- Gains are largest in the single-source setting, where the emulated-domain term helps most when domain information is most restricted.
Limitations
- Per-sample gradient computation is costly; the method restricts variance matching to classifier parameters for efficiency.
- Hyperparameters (lambda1, rho_x, rho') were selected among reported ranges, with best performances reported.
Experimental setup
Datasets: CIFAR-10 CIFAR-10-C PACS OfficeHome DomainNet
Benchmarks: CIFAR-10-C PACS OfficeHome DomainNet
Metrics: test accuracy sharpness
Baselines: ERM SAM SAGM GAM M-ADA LTD Fishr CutMix
Builds on: Sharpness-Aware Minimization GAM SAGM Fishr
Closest papers by meaning
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.91)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 2025 · similarity 0.90)
- One-Step Generalization Ratio Guided Optimization for Domain Generalization (ICML 2025 · similarity 0.89)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.88)
- Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (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.