arxiv: 2604.06349 · v1 · submitted 2026-04-07 · 💻 cs.LG · cs.AI· cs.CV

Recognition: no theorem link

Bi-Level Optimization for Single Domain Generalization

Marzi Heidari , Hanping Zhang , Hao Yan , Yuhong Guo

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:42 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV

keywords single domain generalizationbi-level optimizationdomain prompt encoderfeature-wise linear modulationsurrogate domainsdistribution shiftsmachine learningrobust generalization

0 comments

The pith

BiSDG improves single domain generalization by using bi-level optimization to decouple task learning from domain modeling with surrogate data shifts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses single domain generalization, where a model must perform well on unseen target domains after training on only one labeled source domain. It introduces BiSDG to simulate distribution shifts by building surrogate domains through label-preserving transformations of the source data. A domain prompt encoder produces lightweight modulation signals that adjust features via feature-wise linear modulation. The learning is cast as bi-level optimization, with the inner loop optimizing task performance under fixed prompts and the outer loop updating the encoder to improve generalization across the surrogates. A gradient approximation scheme makes this efficient without second-order derivatives, and experiments show consistent gains over prior methods on standard benchmarks.

Core claim

The central claim is that single domain generalization can be advanced by explicitly separating task optimization from domain modeling in a bi-level framework: surrogate domains are generated from source transformations to mimic shifts, the domain prompt encoder produces modulation signals for feature augmentation, the inner objective minimizes task loss with fixed prompts, and the outer objective maximizes performance across surrogates by updating the encoder, all enabled by a first-order gradient approximation for practical training.

What carries the argument

The domain prompt encoder that generates lightweight modulation signals for feature-wise linear modulation inside a bi-level optimization loop separating task performance from domain generalization.

If this is right

The inner-outer loop separation allows task-specific learning to proceed without interference from domain adaptation signals.
Surrogate domains built from transformations enable training that generalizes without requiring any target data or labels.
The first-order gradient approximation removes the need for expensive second-order computations during bi-level updates.
Consistent benchmark gains indicate the framework can replace or augment existing single-domain generalization techniques.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The modulation approach might transfer to multi-domain settings where multiple source domains are available but still require shift simulation.
Different families of transformations could be tested to determine which shift types are best captured for particular data modalities.
The encoder's lightweight signals suggest potential compatibility with resource-constrained deployment where full fine-tuning is impractical.

Load-bearing premise

Label-preserving transformations applied to the source data produce surrogate domains that sufficiently represent the distribution shifts occurring in real unseen target domains.

What would settle it

A new benchmark dataset where target domains exhibit shifts not reproducible by the paper's label-preserving transformations on the source, and where BiSDG shows no performance advantage over prior single-domain methods.

Figures

Figures reproduced from arXiv: 2604.06349 by Hanping Zhang, Hao Yan, Marzi Heidari, Yuhong Guo.

**Figure 1.** Figure 1: Overview of the BiSDG framework for Single Domain Generalization (SDG). Starting from a labeled source dataset, BiSDG [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Sensitivity analysis for four hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

read the original abstract

Generalizing from a single labeled source domain to unseen target domains, without access to any target data during training, remains a fundamental challenge in robust machine learning. We address this underexplored setting, known as Single Domain Generalization (SDG), by proposing BiSDG, a bi-level optimization framework that explicitly decouples task learning from domain modeling. BiSDG simulates distribution shifts through surrogate domains constructed via label-preserving transformations of the source data. To capture domain-specific context, we propose a domain prompt encoder that generates lightweight modulation signals to produce augmenting features via feature-wise linear modulation. The learning process is formulated as a bi-level optimization problem: the inner objective optimizes task performance under fixed prompts, while the outer objective maximizes generalization across the surrogate domains by updating the domain prompt encoder. We further develop a practical gradient approximation scheme that enables efficient bi-level training without second-order derivatives. Extensive experiments on various SGD benchmarks demonstrate that BiSDG consistently outperforms prior methods, setting new state-of-the-art performance in the SDG setting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BiSDG's bi-level split with domain prompts is a reasonable technical move for single domain generalization, but the surrogate domains built from label-preserving transforms are the part that may not hold up against real target shifts.

read the letter

BiSDG sets up a bi-level optimization where the inner loop handles task learning on fixed domain prompts and the outer loop updates a domain prompt encoder to improve performance across surrogate domains. Those surrogates come from applying label-preserving transformations to the single source. The paper also gives a first-order gradient approximation so the whole thing trains without second derivatives. This decoupling is the main new piece, and it is presented cleanly enough that the formulation itself is easy to follow. The abstract reports consistent gains over prior methods on standard SDG benchmarks, which would be useful if the experiments back it up with proper controls. The prompt encoder for feature modulation is a lightweight addition that fits the setting without adding much overhead. The stress-test concern lands here. Label-preserving operations like rotation or jitter often fail to reproduce the stylistic or semantic shifts in benchmarks such as PACS or Office-Home. When the outer objective only sees those limited surrogates, the learned prompts can overfit to the chosen augmentations rather than to arbitrary unseen targets. The abstract does not detail how the surrogates were chosen, whether ablations tested their coverage, or whether error bars and statistical tests support the SOTA claims. Without those, the generalization story rests on an assumption that may not be true for the test domains. This paper is aimed at researchers working on single-source robustness and optimization-based domain generalization. Readers who want to try bi-level tricks or prompt-style modulation in limited-data settings could extract some ideas from it. The work shows clear problem framing and a workable implementation path, so it is worth sending to referees. I would recommend peer review, with the main request being stronger evidence that the surrogate construction actually approximates the distribution shifts that matter.

Referee Report

2 major / 1 minor

Summary. The paper proposes BiSDG, a bi-level optimization framework for Single Domain Generalization (SDG) from a single labeled source domain. Surrogate domains are constructed via label-preserving transformations (e.g., rotation, color jitter) of the source data to simulate shifts. A domain prompt encoder generates lightweight modulation signals for feature-wise linear modulation (FiLM) to capture domain context. The inner loop optimizes task performance with fixed prompts, while the outer loop updates the prompt encoder to maximize generalization across surrogates; a first-order gradient approximation enables efficient training without second-order derivatives. The abstract claims that extensive experiments on SDG benchmarks show consistent outperformance of prior methods and new state-of-the-art results.

Significance. If the central claim holds and the surrogate-domain construction plus bi-level training demonstrably improves generalization to arbitrary unseen targets (beyond the chosen augmentations), the work would represent a meaningful advance in SDG by decoupling task learning from domain modeling via prompts. This could influence robust ML pipelines where target data is unavailable, particularly if the gradient approximation proves scalable. The approach builds on prompt-based modulation ideas but applies them specifically to bi-level SDG optimization.

major comments (2)

[Abstract] Abstract: The central claim of SOTA SDG performance rests on the assumption that label-preserving transformations of the source data produce surrogate domains whose distribution shifts are representative of those in unseen target domains (e.g., stylistic or semantic shifts in PACS or Office-Home). This assumption is load-bearing for the bi-level objective, yet the provided text offers no justification, sensitivity analysis, or comparison showing that the chosen transformations span real target shifts rather than merely the augmentation distribution; without such evidence the outer-loop optimization may overfit to the surrogates.
[Abstract] Abstract: The assertion that 'extensive experiments on various SGD benchmarks demonstrate that BiSDG consistently outperforms prior methods, setting new state-of-the-art' is presented without any reported baselines, number of runs, statistical tests, ablation results, or error bars. Because the empirical superiority is the primary evidence for the framework's value, the absence of these details in the manuscript text prevents assessment of whether the claimed gains are robust or merely due to implementation differences.

minor comments (1)

[Abstract] The description of the practical gradient approximation scheme for bi-level training is too terse; a short derivation or pseudocode would clarify how first-order updates avoid second-order derivatives while still approximating the outer objective.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and detailed comments on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the paper.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim of SOTA SDG performance rests on the assumption that label-preserving transformations of the source data produce surrogate domains whose distribution shifts are representative of those in unseen target domains (e.g., stylistic or semantic shifts in PACS or Office-Home). This assumption is load-bearing for the bi-level objective, yet the provided text offers no justification, sensitivity analysis, or comparison showing that the chosen transformations span real target shifts rather than merely the augmentation distribution; without such evidence the outer-loop optimization may overfit to the surrogates.

Authors: We acknowledge that the abstract is concise and does not elaborate on this point. In the main text we motivate the choice of label-preserving transformations (rotation, color jitter, etc.) by noting that they are standard in the SDG literature for simulating common real-world shifts such as style, lighting, and orientation changes that appear in target domains of PACS and Office-Home. To directly address the concern about representativeness, we will add a dedicated paragraph in the revised manuscript with (i) a brief literature reference to prior SDG works using similar augmentations, (ii) a sensitivity analysis varying the transformation set and reporting target-domain performance, and (iii) a quantitative comparison (e.g., via MMD or FID) between surrogate-induced shifts and actual target shifts. This will make the load-bearing assumption explicit and testable. revision: yes
Referee: [Abstract] Abstract: The assertion that 'extensive experiments on various SGD benchmarks demonstrate that BiSDG consistently outperforms prior methods, setting new state-of-the-art' is presented without any reported baselines, number of runs, statistical tests, ablation results, or error bars. Because the empirical superiority is the primary evidence for the framework's value, the absence of these details in the manuscript text prevents assessment of whether the claimed gains are robust or merely due to implementation differences.

Authors: We agree the abstract should be more informative. The full manuscript already contains the requested details: tables comparing BiSDG against multiple baselines (ERM, Mixup, RSC, etc.), results averaged over multiple random seeds with standard deviations, ablation studies on the prompt encoder and bi-level components, and paired t-tests for statistical significance. To resolve the referee's concern, we will revise the abstract to include a concise quantitative statement (e.g., 'outperforms prior methods by X% on average across benchmarks, with results averaged over 5 runs and statistical significance reported') while directing readers to the experiments section for full tables, error bars, and ablations. We will also ensure every result table in the revised version explicitly states the number of runs and includes error bars. revision: partial

Circularity Check

0 steps flagged

No circularity: bi-level optimization and surrogate-domain construction remain independent of the target generalization claim

full rationale

The paper defines a standard bi-level objective (inner: task loss on fixed prompts; outer: prompt-encoder update to maximize performance across label-preserving surrogate domains) and reports empirical results on SDG benchmarks. No equation reduces the outer generalization objective to a quantity defined by the inner loop or by the surrogate construction itself. Surrogate domains are an explicit modeling choice whose adequacy is an empirical assumption, not a definitional tautology. No self-citation is load-bearing for the central claim, no uniqueness theorem is invoked, and no fitted parameter is relabeled as a prediction. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that label-preserving transformations create useful surrogate domains and on the introduction of a new domain prompt encoder component whose effectiveness is asserted via experiments.

axioms (1)

domain assumption Label-preserving transformations of source data can simulate realistic unseen domain shifts
Invoked to justify construction of surrogate domains for the outer optimization loop.

invented entities (1)

Domain prompt encoder no independent evidence
purpose: Generates lightweight modulation signals for feature-wise linear modulation to capture domain-specific context
New component introduced to produce augmenting features; no independent evidence provided in abstract.

pith-pipeline@v0.9.0 · 5482 in / 1176 out tokens · 46641 ms · 2026-05-10T18:42:17.753649+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Stochastic gradient descent tricks.Neural Net- works: Tricks of the Trade: Second Edition, page 421–436,

L ´eon Bottou. Stochastic gradient descent tricks.Neural Net- works: Tricks of the Trade: Second Edition, page 421–436,
[2]

Domain generalization by solving jigsaw puzzles

Fabio M Carlucci, Antonio D’Innocente, Silvia Bucci, Bar- bara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7

2019
[3]

Swad: Domain generalization by seeking flat minima

Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. InAd- vances in Neural Information Processing Systems (NeurIPS),
[4]

Meta-causal learning for single domain generalization

Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 7

2023
[5]

Autoaugment: Learning augmentation strategies from data

Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasude- van, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 7

2019
[6]

Randaugment: Practical automated data augmentation with a reduced search space

Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. InAdvances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 7

2020
[7]

Attention consistency on visual corruptions for single-source domain generalization

Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition Work- shops (CVPRW), 2022. 2, 7

2022
[8]

Neural network recog- nizer for hand-written zip code digits

John S Denker, WR Gardner, Hans Peter Graf, Donnie Hen- derson, Richard E Howard, W Hubbard, Lawrence D Jackel, Henry S Baird, and Isabelle Guyon. Neural network recog- nizer for hand-written zip code digits. InAdvances in Neural Information Processing Systems (NeurIPS), 1989. 6

1989
[9]

Improved Regularization of Convolutional Neural Networks with Cutout

Terrance DeVries and Graham W Taylor. Improved regular- ization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017. 1, 2, 7

work page internal anchor Pith review arXiv 2017
[10]

Learning to learn with variational information bottleneck for domain gener- alization

Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain gener- alization. InEuropean Conference on Computer Vision (ECCV), 2020. 1

2020
[11]

Unsupervised domain adaptation by backpropagation

Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. InInternational Conference on Machine Learning (ICML), 2015. 6

2015
[12]

Bi-level optimization for semi-supervised learning with pseudo-labeling

Marzi Heidari and Yuhong Guo. Bi-level optimization for semi-supervised learning with pseudo-labeling. InAAAI conference on artificial intelligence (AAAI), 2025. 2

2025
[13]

Aug- mix: A simple data processing method to improve robustness and uncertainty

Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Aug- mix: A simple data processing method to improve robustness and uncertainty. InInternational Conference on Learning Representations (ICLR), 2019. 2, 7

2019
[14]

Meta-learning the invariant repre- sentation for domain generalization.Machine Learning, 113: 1661–1681, 2024

Chen Jia and Yue Zhang. Meta-learning the invariant repre- sentation for domain generalization.Machine Learning, 113: 1661–1681, 2024. 2

2024
[15]

Springer, 2011

Vladimir Koltchinskii.Oracle inequalities in empirical risk minimization and sparse recovery problems: ´Ecole D’´Et´e de Probabilit´es de Saint-Flour XXXVIII-2008. Springer, 2011. 7

2008
[16]

Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86, no, 11:2278–2324,

Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86, no, 11:2278–2324,
[17]

Set transformer: A frame- work for attention-based permutation-invariant neural net- works

Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Se- ungjin Choi, and Yee Whye Teh. Set transformer: A frame- work for attention-based permutation-invariant neural net- works. InInternational conference on machine learning (ICML). PMLR, 2019. 4

2019
[18]

Deeper, broader and artier domain general- ization

Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain general- ization. InInternational Conference on Computer Vision (ICCV), 2017. 6

2017
[19]

Learning to generalize: Meta-learning for do- main generalization

Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Learning to generalize: Meta-learning for do- main generalization. InAAAI Conference on Artificial Intel- ligence (AAAI), 2018. 1

2018
[20]

Pro- gressive domain expansion network for single domain gen- eralization

Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xi- aoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Pro- gressive domain expansion network for single domain gen- eralization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7

2021
[21]

Deep domain gener- alization via conditional invariant adversarial networks

Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain gener- alization via conditional invariant adversarial networks. In European Conference on Computer Vision (ECCV), 2018. 1

2018
[22]

DARTS: Differentiable architecture search

Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. InInternational Confer- ence on Learning Representations (ICLR), 2019. 2

2019
[23]

Domain generalization using causal matching

Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. InInternational Con- ference on Machine Learning (ICML), 2021. 1

2021
[24]

Unified deep supervised domain adap- tation and generalization

Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gi- anfranco Doretto. Unified deep supervised domain adap- tation and generalization. InInternational Conference on Computer Vision (ICCV), 2017. 7

2017
[25]

Reading digits in natural images with unsupervised feature learning.NeurIPS work- shop on deep learning and unsupervised feature learning,

Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis- sacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning.NeurIPS work- shop on deep learning and unsupervised feature learning,
[26]

Hyperparameter optimization with ap- proximate gradient

Fabian Pedregosa. Hyperparameter optimization with ap- proximate gradient. InInternational conference on machine learning (ICML), 2016. 2

2016
[27]

Advancing open-set domain generalization using ev- idential bi-level hardest domain scheduler

Kunyu Peng, Di Wen, Kailun Yang, Ao Luo, Yufan Chen, Jia Fu, M Saquib Sarfraz, Alina Roitberg, and Rainer Stiefelha- gen. Advancing open-set domain generalization using ev- idential bi-level hardest domain scheduler. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 3

2024
[28]

Moment matching for multi-source domain adaptation

Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. InInternational Conference on Com- puter Vision (ICCV), 2019. 6

2019
[29]

Film: Visual reasoning with a general conditioning layer

Ethan Perez, Florian Strub, Harm De Vries, Vincent Du- moulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InAAAI conference on artificial intelligence (AAAI), 2018. 1, 4

2018
[30]

Learning to learn single domain generalization

Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7

2020
[31]

Bi- level meta-learning for few-shot domain generalization

Xiaorong Qin, Xinhang Song, and Shuqiang Jiang. Bi- level meta-learning for few-shot domain generalization. InIEEE/CVF conference on computer vision and pattern recognition (CVPR), 2023. 3

2023
[32]

Addressing model vulnerability to distributional shifts over image transforma- tion sets

Riccardo V olpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transforma- tion sets. InInternational Conference on Computer Vision (ICCV), 2019. 7

2019
[33]

Generalizing to unseen domains via adversarial data augmentation

Riccardo V olpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. InAd- vances in Neural Information Processing Systems (NeurIPS),
[34]

Learning to diversify for single do- main generalization

Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single do- main generalization. InInternational Conference on Com- puter Vision (ICCV), 2021. 2, 6, 7

2021
[35]

Simde: A simple domain ex- pansion approach for single-source domain generalization

Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain ex- pansion approach for single-source domain generalization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

2023
[36]

Cutmix: Regu- larization strategy to train strong classifiers with localizable features

Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. InInternational Conference on Computer Vision (ICCV), 2019. 7

2019
[37]

mixup: Beyond empirical risk minimiza- tion

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. InInternational Conference on Learning Representa- tions (ICLR), 2018. 2, 7

2018
[38]

Flatness-aware minimization for domain generalization

Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. InInternational Conference on Computer Vision (ICCV), 2023. 1

2023
[39]

Adversarial style augmentation for domain gener- alization

Yabin Zhang, Bin Deng, Ruihuang Li, Kui Jia, and Lei Zhang. Adversarial style augmentation for domain gener- alization. InarXiv preprint arXiv:2301.12643, 2023. 2

work page arXiv 2023
[40]

Maximum-entropy adversarial data augmentation for im- proved generalization and robustness

Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for im- proved generalization and robustness. InAdvances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 7

2020
[41]

AdvST: revisiting data augmentations for single domain generaliza- tion

Guangtao Zheng, Mengdi Huai, and Aidong Zhang. AdvST: revisiting data augmentations for single domain generaliza- tion. InAAAI conference on artificial intelligence (AAAI),