Research map / Label-efficient and robust learning / Domain adaptation and generalization
AdvST: Revisiting Data Augmentations for Single 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 AdvST, which treats compositions of standard augmentations with learnable parameters as semantics transformations and trains a robust model via adversarial min-max learning, with a DRO interpretation, achieving top average single-domain generalization on Digits, PACS, and DomainNet.
Problem
Single domain generalization: training a classifier on one source domain so it robustly generalizes to unknown target domain shifts, where standard augmentations are underused.
Approach
Make standard augmentations (rotate, translate, HSV, etc.) learnable semantics transformations, adversarially optimize their parameters to maximize loss with a feature-space distance regularizer, and train the model on the worst-case augmented samples; theoretically connected to Wasserstein-ball distributionally robust optimization.
Key results
- On Digits (MNIST source), AdvST-ME averages 80.0% over four target domains vs 74.8% for PDEN, the prior best.
- On PACS (Photo source), AdvST averages 64.1% vs 52.5% for L2D and 48.2% for ACVC.
- On DomainNet (Real source), AdvST averages 27.1% vs 26.9% for ACVC; the gain over the best baseline is small.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use semantics transformations instead of none of the three components (ablation) | AdvST-ME · Digits (MNIST source; four target domains, Avg.) | accuracy (%): 59.3±1.5 -> 77.0±0.4 | Helped |
| add contrastive regularization to semantics transformations (ablation) | AdvST-ME · Digits (MNIST source; four target domains, Avg.) | accuracy (%): 77.0±0.4 -> 77.8±0.2 | Helped |
| add entropy regularization to semantics transformations (ablation) | AdvST-ME · Digits (MNIST source; four target domains, Avg.) | accuracy (%): 77.0±0.4 -> 78.8±0.2 | Helped |
| add both contrastive and entropy regularization to semantics transformations (ablation) | AdvST-ME · Digits (MNIST source; four target domains, Avg.) | accuracy (%): 77.0±0.4 -> 80.0±0.4 | Helped |
| add maximum-entropy regularization | AdvST · PACS (Photo source; three target domains, Avg.) | accuracy (%): 64.1±0.4 -> 62.5±0.8 | Hurt |
| add maximum-entropy regularization | AdvST · DomainNet (Real source; five target domains, Avg.) | accuracy (%): 27.1±0.2 -> 27.2±0.1 | Helped |
| add maximum-entropy regularization | AdvST · OfficeHome (Avg.) | accuracy (%): 53.5±0.1 -> 53.2±0.2 | Hurt |
| remove each standard transformation in turn from the augmentation pool (ablation) | AdvST-ME · Digits (four target domains, Avg.) | accuracy | Hurt |
| remove Translate from the augmentation pool (ablation) | AdvST-ME · Digits (SYN) | accuracy | Hurt |
| remove Translate from the augmentation pool (ablation) | AdvST-ME · Digits (MNIST-M) | accuracy | Helped |
| remove Scale from the augmentation pool (ablation) | AdvST-ME · Digits (SVHN and USPS) | accuracy | Mixed |
| use AdvST rather than pixel-level adversarial augmentation | SDG methods · Digits (four target domains, Avg.) | accuracy (%): 54.6 -> 80.1 | Helped |
Surprising findings
- Maximizing output entropy (AdvST-ME) did not help on PACS, performing slightly below AdvST when generalizing from natural to artistic domains.
- Translate helps SYN generalization but hurts MNIST-M, showing augmentation utility is domain-dependent.
Limitations
- Maximization over T_max steps per sample increases training cost relative to standard augmentation.
- Performance trade-offs across target domains; some augmentations help one target and hurt another.
- Gains on DomainNet are marginal over the strongest baseline.
Experimental setup
Datasets: MNIST PACS DomainNet OfficeHome
Benchmarks: Digits PACS DomainNet OfficeHome
Metrics: classification accuracy average accuracy over target domains
Baselines: ERM CCSA d-SNE JiGen ADA ME-ADA M-ADA AutoAug RandAug RSDA PDEN L2D
Builds on: ADA ME-ADA AutoAugment RandAugment
Closest papers by meaning
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.89)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.88)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.87)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.87)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 2025 · similarity 0.87)
- Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.87)
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.