Research map / Label-efficient and robust learning / Domain adaptation and generalization
Mixup-Induced Domain Extrapolation 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 EDM, a two-stage Mixup-based domain augmentation (reverse-Mixup extrapolation then Dir-Mixup interpolation) that generates extra domains, motivated by a PAC-Bayes domain generalization bound involving the number of source domains and a Wasserstein radius.
Problem
Domain generalization with only a few source domains, which cannot reflect the whole domain population, yields biased domain-invariant representations and poor generalization to unseen targets.
Approach
Generate extrapolated domains by reversing the multi-domain Mixup scheme on Gaussian domain statistics (mean/variance of feature maps), then mix them with source domains to create interpolated and extrapolated domains, expanding the space spanned by sources; AdaIN is used to synthesize samples in the new domains. The method is plug-and-play onto existing DG methods.
Key results
- EDM plugged into DANN gives up to 5.73% accuracy improvement on the sketch domain of PACS.
- DAML + EDM reaches 70.78% accuracy on PACS and 59.58% on Office-Home in the open-set setting, the best among compared methods.
- Mixup + EDM achieves 73.15% average accuracy on Office-Home versus 72.20% for Mixup alone (per Table 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add interpolation and extrapolation domain augmentation | DAML · PACS and Office-Home (open-set) | accuracy: 65.49 -> 70.78; 56.45 -> 59.58 | Helped |
| Add interpolation and extrapolation domain augmentation | DAML · PACS and Office-Home (open-set) | H-score: 51.88 -> 54.12; 53.34 -> 54.19 | Helped |
| Add interpolation and extrapolation domain augmentation | ERM · PACS | accuracy (Avg): 82.75 -> 84.84 | Helped |
| Add interpolation and extrapolation domain augmentation | DANN · PACS | accuracy (Avg): 83.57 -> 85.06 | Helped |
| Add interpolation and extrapolation domain augmentation | Mixup · PACS | accuracy (Avg): 82.26 -> 85.07 | Helped |
| Add interpolation and extrapolation domain augmentation | SAGM · Office-Home | accuracy (Avg): 64.73 -> 65.04 | Helped |
| Add interpolation and extrapolation domain augmentation | ERM · DomainNet | accuracy (Avg): 40.10 -> 40.34 | Helped |
| Add interpolation and extrapolation domain augmentation | DANN · DomainNet | accuracy (Avg): 40.20 -> 40.40 | Helped |
| Add interpolation and extrapolation domain augmentation | Mixup · Office-Home | accuracy (Avg): 64.33 -> 65.10 | Helped |
| Add interpolation and extrapolation domain augmentation | SAGM · PACS | accuracy (Avg): 84.20 -> 85.08 | Helped |
| Add interpolated domains without extrapolated domains (ablation) | ERM · PACS | accuracy (Avg): 82.75 -> 84.48 | Helped |
| Add interpolated domains without extrapolated domains (ablation) | ERM · Office-Home | accuracy (Avg): 63.54 -> 63.83 | Helped |
Surprising findings
- Adding EDM to SAGM gives no significant improvement, which the authors attribute to SAGM's model perturbation indirectly simulating domain drift.
Limitations
- Theoretical bound relies on assumptions such as light-tailed sampling and relaxing the environment gap to a Wasserstein radius.
- Does not directly solve unknown-class detection in open-set settings, so H-score gains are smaller than accuracy gains.
Experimental setup
Datasets: PACS Office-Home DomainNet
Benchmarks: PACS Office-Home DomainNet
Metrics: accuracy H-score
Baselines: ERM DANN Mixup RSC MMD CORAL GroupDRO CrossGrad MixStyle ANDMask VREx DIFEX
Builds on: Mixup Dir-Mixup MixStyle PAC-Bayes DIFEX
Closest papers by meaning
- Adaptive Methods for Real-World Domain Generalization (CVPR 2021 · similarity 0.89)
- Failure Modes of Domain Generalization Algorithms (CVPR 2022 · similarity 0.89)
- Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · similarity 0.87)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.87)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.86)
- Weight Diffusion for Future: Learn to Generalize in Non-Stationary Environments (NeurIPS 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.