Research map / Label-efficient and robust learning / OOD and anomaly detection
On the Connection between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use domain-wise adversarial perturbations instead of ERM | image-classification domain generalization · CMNIST and NICO | test accuracy (%): 58.5 ± 0.3 -> 68.4 ± 2.0 (CMNIST); 72.1 ± 1.6 -> 72.6 ± 1.7 (NICO) | Helped |
| use domain-wise adversarial perturbations instead of ERM | image-classification domain generalization · PACS and TerraInc | test accuracy (%): 81.5\pm 0.0 -> 82.0 \pm 0.1 (PACS); 42.6 \pm 0.9 -> 42.7 \pm 0.7 (TerraInc) | Mixed |
| use DAT instead of ERM | single-source domain generalization · CMNIST, NICO, PACS, TerraInc | average test accuracy (%): 45.8 ± 0.7 -> 46.0\pm 2.0; 63.2\pm 2.9 -> 64.4\pm 0.7; 59.8\pm 1.6 -> 59.9\pm 2.4; 27.3 \pm 4.0 -> 28.1\pm 4.0 | Helped |
| use DAT instead of ERM | single-source Digits generalization · Digits average | accuracy (%): 49.29 -> 52.1 | Helped |
| generate domain-specific perturbations rather than one universal perturbation per model (ablation) | multi-domain adversarial generalization · CMNIST and NICO | test accuracy (%): 58.7 \pm 2.3 -> 68.4 \pm 2.0 (CMNIST); 69.1 \pm 1.2 -> 72.6 \pm 1.7 (NICO) | Helped |
| replace DAT with an ensemble of single-domain UAT models (ablation) | multi-domain adversarial generalization · CMNIST and NICO | test accuracy (%): 68.4 \pm 2.0 -> 58.2 \pm 2.3 (CMNIST); 72.6 \pm 1.7 -> 60.8 \pm 0.2 (NICO) | Hurt |
| increase DAT perturbation radius beyond 10^-1 (ablation) | DAT · NICO | test accuracy (%): 72.9% to 67.9% | Hurt |
| increase the perturbation step size within the tested radius [10^-2, 10^-1] (ablation) | DAT · NICO | test accuracy (%): 72.6 \pm 1.7 -> 72.0 \pm 2.1 -> 68.9 \pm 1.5 | Hurt |
| use ℓ∞ rather than ℓ2 perturbations with radius [10^-1, 10^0] (ablation) | DAT · NICO | test accuracy (%): 72.6 \pm 1.7 -> 72.5 \pm 1.6 | No clear effect |
| increase batch size used to estimate domain-level perturbation updates (ablation) | DAT · PACS | test accuracy | Helped |
| use DAT rather than ERM | background-shift classification · Mixed-Rand test, Mixed-Same training | test accuracy (%): 71.9% -> 72.6% | Helped |
Closest papers by meaning
- Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors (NeurIPS 2022 · similarity 0.90)
- Understanding and Improving Feature Learning for Out-of-Distribution Generalization (NeurIPS 2023 · similarity 0.88)
- Out-of-distribution Generalization for Total Variation based Invariant Risk Minimization (ICLR 2025 · similarity 0.87)
- Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.87)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.86)
- Towards Robust Out-of-Distribution Generalization Bounds via Sharpness (ICLR 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.