Research map / Label-efficient and robust learning / Domain adaptation and generalization

D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation

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 D-GAP, a dataset-agnostic augmentation that uses task-gradient sensitivity maps in Fourier amplitude space to adaptively mix source and target amplitudes, plus a pixel-space blending branch, to improve OOD robustness under unsupervised domain adaptation.

Problem

Out-of-domain robustness under domain shifts (background, style, instrument) in unsupervised domain adaptation, where generic augmentations are inconsistent and dataset-specific augmentations need expert knowledge; fixed frequency perturbations ignore model response and pixel detail.

Approach

Computes a frequency-wise sensitivity map from the task-loss gradient with respect to the source amplitude spectrum, converts it to a mixing map that controls how much target-domain amplitude is injected at each frequency, and then blends the result with a pixel-space mix to recover spatial detail.

Key results

• Camelyon17 OOD accuracy 96.4% vs 92.2% for best generic augmentation (SAM) and 91.4% for LP-FT.
• BirdCall OOD macro F1 40.7% vs 35.1% for FACT; Galaxy10 OOD accuracy 83.4% vs 74.1% for SAM.
• Average OOD accuracy on PACS 89.03 and OfficeHome 70.22 and Digits-DG 84.5, outperforming FACT, SAM, Su et al. and AGST.

Surprising findings

• Pixel-only mixing performs worse than frequency-only mixing, and the paper reports that fixed low/high/ring frequency masks reduce OOD performance relative to the gradient-guided map.
• Connectivity analysis shows the method's beta/gamma ratio is moderate, suggesting it slightly perturbs class-relevant features while still achieving the best OOD performance.

Limitations

• Requires additional gradient computation with respect to the Fourier amplitude, increasing training time roughly 3.7x over ERM.
• Moderate beta/gamma connectivity suggests some perturbation of class-relevant features.
• Some baseline numbers (FACT, SAM on benchmarks) are copied from source papers rather than re-run under identical settings.

Experimental setup

Datasets: iWildCam Camelyon17 BirdCalls Galaxy10 PACS Office-Home Digits-DG

Benchmarks: iWildCam Camelyon17 BirdCalls Galaxy10 PACS Office-Home Digits-DG

Metrics: accuracy macro F1 OOD accuracy connectivity ratios (alpha/gamma, beta/gamma) Jaccard similarity

Baselines: RandAugment CutMix MixUp Cutout FACT SAM LISA DANN DeepCORAL ERM LP-FT Connect Later

Builds on: FACT Connect Later LP-FT FDA SAM (semantic-aware mixup)

Closest papers by meaning

• Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (CVPR 2024 · similarity 0.88)
• In Search of Forgotten Domain Generalization (ICLR 2025 · similarity 0.86)
• LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.85)
• CutFreq: Cut-and-Swap Frequency Components for Low-Level Vision Augmentation (AAAI 2024 · similarity 0.85)
• D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · similarity 0.84)
• PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.84)