Research map / Label-efficient and robust learning / Domain adaptation and generalization
Towards Combating Frequency Simplicity-biased Learning 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 two in-batch adversarial Fourier-amplitude augmentation modules (AAUA and AAD) that dynamically perturb over-reliance frequency components to suppress frequency shortcut learning for single-source domain generalization.
Problem
Single-source domain generalization fails because networks learn frequency shortcuts (simplicity-biased, class-discriminative spatial frequency bands), and prior frequency/style augmentations increase these shortcuts while improving accuracy.
Approach
Adversarially perturb the amplitude spectrum of in-batch images: AAUA adversarially resamples low-frequency amplitude statistics (mean/variance) with instance normalization, and AAD drops frequency bands chosen by amplitude-gradient sensitivity with random thresholds, together with a JS consistency loss, to modify the dataset's frequency structure and block shortcut learning.
Key results
- On PACS, the method reaches 67.91% average accuracy, versus 65.71% for SADA and 59.10% for FACT.
- On Market-to-Duke person re-ID, mAP rises from 33.2 (AdvStyle) and 33.1 (SADA) to 33.8 with the proposed method, and on Duke-to-Market mAP reaches 34.2.
- Frequency shortcut quantification on ImageNet-10 DFM-filtered test set lowers avg TPR to 0.212 and avg FPR to 0.030 versus 0.286/0.041 for ResNet-18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AAUA to AAD + JS loss (ablation) | PACS classification · PACS | Avg.: 63.91 -> 67.91 | Helped |
| add AAD to AAUA + JS loss (ablation) | PACS classification · PACS | Avg.: 66.47 -> 67.91 | Helped |
| add the JS consistency loss to AAD + AAUA (ablation) | PACS classification · PACS | Avg.: 66.35 -> 67.91 | Helped |
| add AAD to AAUA without the JS consistency loss (ablation) | PACS classification · PACS | Avg.: 65.53 -> 66.35 | Helped |
| add AAUA to AAD without the JS consistency loss (ablation) | PACS classification · PACS | Avg.: 63.31 -> 66.35 | Helped |
| combine AAUA and AAD | ResNet-18 frequency-shortcut evaluation · ImageNet-10 DFM-filtered test set | Avg TPR: 0.286 -> 0.212 | Helped |
| combine AAUA and AAD | ResNet-18 frequency-shortcut evaluation · ImageNet-10 DFM-filtered test set | Avg FPR: 0.041 -> 0.030 | Helped |
| use AAUA alone (ablation) | ResNet-18 frequency-shortcut evaluation · ImageNet-10 DFM-filtered test set | Avg FPR: 0.041 -> 0.068 | Hurt |
| use AAD alone (ablation) | ResNet-18 frequency-shortcut evaluation · ImageNet-10 DFM-filtered test set | Avg FPR: 0.041 -> 0.083 | Hurt |
| add the proposed augmentations | OSNet person re-identification · Market→Duke | mAP: 29.7 to 33.8 | Helped |
Surprising findings
- Using AAUA or AAD alone increases frequency shortcut application relative to the ResNet-18 baseline, and only their combination suppresses shortcuts.
- Prior successful augmentations such as AugMix and style augmentation apply more frequency shortcuts, suggesting accuracy gains can be hallucinated.
Limitations
- Direct localization of frequency shortcuts is not solved; the method only indirectly prevents them.
- Robustness of the frequency-shortcut evaluation metric is unexplored.
- Ablations show AAUA or AAD alone can increase shortcut usage in high-frequency bands.
Experimental setup
Datasets: PACS Digits CIFAR-10-C Market1501 DukeMTMC ImageNet-10
Benchmarks: PACS Digits CIFAR-10-C Market1501 DukeMTMC ImageNet-10
Metrics: accuracy mAP Rank-1 Rank-5 Rank-10 avg TPR avg FPR
Baselines: ERM GUD M-ADA AugMix RandConv L2D FACT DSU SADA FFM ALT ABA
Builds on: FACT SADA AugMix DSU AdvStyle
Closest papers by meaning
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.87)
- Tuning Frequency Bias of State Space Models (ICLR 2025 · similarity 0.85)
- Understanding Domain Generalization: A Noise Robustness Perspective (ICLR 2024 · similarity 0.84)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.84)
- PEER Pressure: Model-to-Model Regularization for Single Source Domain Generalization (CVPR 2025 · similarity 0.84)
- Data Distributional Properties As Inductive Bias for Systematic Generalization (CVPR 2025 · similarity 0.84)
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.