Pith. sign in

REVIEW 5 major objections 6 minor 60 references

FEDTAIL: Federated Long-Tailed Domain Generalization with Sharpness-Guided Gradient Matching

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read FedTAIL claims that aligning gradients, applying per-class sharpness minimization, and weighting by curvature yields state-of-the-art domain generalization under long-tailed, decentralized data.

desk verdict Competent recombination of known sharpness tricks, but the paper's headline claims about federated and long-tailed DG are never actually tested. read the letter →

arxiv 2506.08518 v1 pith:M4T6LD6H submitted 2025-06-10 cs.AI cs.CVcs.LG

classification cs.AIcs.CVcs.LG
keywords FedTAILdomaingeneralizationfederatedlearninglong-tailedclassificationsharpness-awareminimizationgradientcoherenceclassimbalanceadversarialalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that domain generalization can be made robust to class imbalance and to conflicts between optimization objectives at the same time. It proposes FedTAIL, one training objective with four interacting additions: a gradient-coherence penalty that keeps classification and adversarial gradients aligned, a separate sharpness-aware perturbation for each class, a curvature-based weight that amplifies tail classes sitting in sharp regions of the loss, and a sharpness-aware entropy term that smooths predictions across domains. On the standard leave-one-domain-out protocol, FedTAIL reports average accuracies of 90.2% on PACS, 73.1% on OfficeHome, 89.2% on Digits-DG, and 68.8% on mini-DomainNet, each ahead of the prior results listed. A sympathetic reading is that a single harmonized objective can deliver these gains without centralized data, since all components are computed locally and only model updates are exchanged.

What carries the argument

The load-bearing mechanism is the combined FedTAIL objective, in which each term targets a separate failure mode. $L_{\text{coh}} = -\alpha\langle\nabla_\theta L_{\text{cls}}, \nabla_\theta L_{\text{adv}}\rangle$ penalizes opposite gradient directions so adversarial alignment does not fight classification; class-wise perturbations $\epsilon_c = \rho\,\nabla_\theta L_c/\|\nabla_\theta L_c\|_2$ localize sharpness for each class; curvature weights $\gamma_c = 1/(1+\sigma_{\max}(\nabla^2 L_c))$ up-weight classes whose loss surface is still sharp; and the sharpness-aware entropy term $L_{\text{sharp-er}}$ aligns perturbed conditional predictions with a class-frequency target $Q_T$, preventing confident head classes from dominating. The gradient coherence term is the part that makes the adversarial objective stable enough for the other components to help.

What would settle it

Compute the per-class frequency curve of each benchmark and run the federated protocol exactly as Section 5.3 describes, with one client per domain, one local epoch, and FedAvg aggregation, reporting accuracy and communication cost against a federated ERM baseline; the long-tailed and federated claims collapse if the datasets turn out near-balanced or the federated variant is never run.

Watch

Extended reading notes

Core claim

FedTAIL's central claim is that three failure modes can be corrected in one objective: conflicting gradients between the classification and adversarial losses, class-agnostic sharpness minimization that overlooks tail classes, and entropy regularization that lets confident head classes dominate. The proposed solution is the combined loss $$L_{\text{FedTAIL}} = L_{\text{cls}} + L_{\text{adv}} + L_{\text{sharp-er}} + \sum_{c}\gamma_c L_c + L_{\text{coh}},$$ with a gradient-coherence term $L_{\text{coh}} = -\alpha\langle\nabla_\theta L_{\text{cls}}, \nabla_\theta L_{\text{adv}}\rangle$, per-class SAM perturbations, curvature weights $\gamma_c = 1/(1+\sigma_{\max}(\nabla^2 L_c))$, and a sharpness-aware entropy alignment to a target distribution $Q_T$. The paper reports that this objective raises average leave-one-domain-out accuracy to 90.2% on PACS, 73.1% on OfficeHome, 89.2% on Digits-DG, and 68.8% on mini-DomainNet, all above the listed prior methods.

Load-bearing premise

The load-bearing premise is that PACS, OfficeHome, Digits-DG, and mini-DomainNet actually have skewed class counts and that a leave-one-domain-out run with each domain as one client is enough to validate a federated method; if either is untrue, the long-tailed and federated claims are unverified.

Editorial extensions

If this is right

  • On PACS, FedTAIL's reported 90.2% average exceeds the strongest listed baseline (89.8%), with the largest margins on Cartoon and Sketch.
  • On Digits-DG, the reported 89.2% average is roughly 7.4 points above the best prior baseline listed (81.8%).
  • The PACS ablation traces the gain from 83.8% to 90.2% across four added components, each contributing positively: adversarial alignment (+2.2), sharpness-aware entropy (+2.2), class-wise curvature balancing (+1.4), and gradient coherence (+0.6).
  • Because clients compute class perturbations and local gradients while sharing only model updates, the objective is compatible with federated averaging and does not require pooling raw data.
  • The method also adapts to balanced benchmarks: the entropy target $Q_T$ becomes uniform when classes are balanced, so the alignment term does not distort training on non-long-tailed data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension the authors leave implicit: if the curvature-aware class weighting is the active ingredient, FedTAIL should also improve standard long-tailed classification and semi-supervised learning, where the same high-confidence gradient bias appears; a test on those settings would isolate the mechanism.
  • Extension: a full federated run with one client per domain under FedAvg, reporting accuracy and communication cost against federated baselines, would settle the federated claim, since the tables report centralized leave-one-domain-out runs.
  • Extension: the frequency-based $Q_T$ becomes uniform on balanced benchmarks like Digits-DG, so a learned or momentum-updated $Q_T$ could keep the alignment term informative even when label counts are balanced.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes FedTAIL, a domain generalization method that combines (i) a gradient coherence regularizer intended to align classification and domain-adversarial gradients, (ii) class-wise sharpness-aware minimization with a curvature-based weighting gamma_c, and (iii) a sharpness-aware entropy regularization term that matches predictions to a target-like distribution Q_T. The claimed contribution is a federated, long-tailed domain generalization framework. Experiments are reported on PACS, OfficeHome, Digits-DG, and mini-DomainNet under a leave-one-domain-out evaluation, together with an ablation study and t-SNE visualizations in the appendix. The central empirical claim is state-of-the-art accuracy, particularly under domain shift and label imbalance.

Significance. The combination of existing ideas is sensible, the code is released, and the ablation in Table 5 shows monotonic PACS gains from the added components. However, the significance as a federated long-tailed DG method is not established: the experiments are centralized and balanced, the closest prior method SAMALTDG is absent from the comparisons, and the Q_T construction is internally inconsistent. If genuine federated and long-tailed evaluations were supplied, the method could be a useful contribution; with the current evidence, the claims run well ahead of the experiments.

major comments (5)
  1. [§5.2–5.3 and Tables 1–4] The evaluation never instantiates the claimed federated setting. Section 5.2 describes the standard centralized leave-one-domain-out protocol, and although Section 5.3 states that 'in federated settings, each domain corresponds to a separate client' and that FedAvg is used, no table reports federated results, and there is no client partitioning, no local-epoch parameter, no communication-cost analysis, and no federated baseline. The abstract's claim that results 'validate effectiveness in both centralized and federated settings' is therefore unsupported.
  2. [§5.1, Tables 1–4, Appendix Tables 6–7] The benchmarks are not long-tailed. PACS, OfficeHome, and mini-DomainNet are used with their original splits, and Appendix Table 7 explicitly reports uniform Q_T=0.1 for Digits-DG. No class-imbalance ratio, resampled long-tailed split, or class-balanced metric is reported. Consequently the stated advantage 'particularly in the presence of label imbalance' is not tested by any experiment in the paper.
  3. [§4 Eq. (12) and Appendix] There is an inconsistency in the definition of Q_T. Eq. (12) defines Q_T as a target-like predictive distribution conditioned on F(X+epsilon) and computed from an ensemble or momentum-updated model, but the appendix states that Q_T is calculated as the relative frequency of each class (freq_class/freq_total), and Tables 6–7 report unconditional class proportions. These are different objects; the KL term in Eq. (12) conditions on features, whereas the reported implementation uses training-set class priors. This discrepancy affects the interpretation and reproducibility of the L_sharp-er component, which is central to the method.
  4. [§2 and Tables 1–4] SAMALTDG, the method this paper explicitly extends ('SAMALTDG addresses class imbalance ... but remains centralized'), is not included in any comparison table, nor is any other long-tailed DG baseline. Without such comparisons, the claim of improved long-tailed domain generalization over the closest prior work is unverified even in the centralized setting.
  5. [§5.2–5.3] The reported numbers are averages over three seeds, but no standard deviations, confidence intervals, or per-class accuracies are shown despite the text stating that per-class accuracy is reported. Hyperparameters rho, alpha, and gamma are set to single defaults without a sensitivity study, so the margins over prior methods (e.g., 89.2 vs. 81.8 on Digits-DG in Table 3) cannot be assessed for statistical reliability or hyperparameter dependence.
minor comments (6)
  1. [§5.2] Section 5.2 states that 'We report both overall accuracy and average per-class accuracy to account for class imbalance,' but Tables 1–4 contain only overall accuracy; please either add per-class columns or delete the sentence.
  2. [§4 Eq. (12) and §5.3] Equation (12) is written as a KL divergence between conditional distributions, but Section 5.3 says the KL is 'computed over batch-wise class distributions'; please specify the estimator and how the conditioning on F(X+epsilon) is realized in the implementation.
  3. [Abstract and §6] The abstract and conclusion claim validation 'in both centralized and federated settings'; given the absence of federated experiments, these claims should be limited to centralized leave-one-domain-out results.
  4. [References] Reference [60] appears to duplicate Reference [49] (same title and authors, one listing 'Zijian Lin' and the other 'Zijian Ling'); please check and consolidate.
  5. [Figure 2] The caption refers to 'Standard DG' without identifying which method is plotted; please name the baseline.
  6. [Table 5] In the ablation, the row labeled '+ Class Bal.' is said to include class-wise sharpness minimization with curvature-aware weighting, but the text does not specify whether the per-class perturbations of Eqs. (9)–(10) are already active in the baseline or are introduced only in that row; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the proposed FedTAIL objective is an ab initio combination of cited components, and its benchmark accuracies are external measurements. The unsubstantiated federated and long-tailed evaluation is a validity concern, not a circularity.

full rationale

The paper does not derive any quantity from a fitted parameter and then relabel it as a prediction. Equations (8)-(13) define new regularization terms (gradient coherence, class-wise SAM perturbations, curvature-aware weights, and sharpness-aware entropy) directly from the model's own gradients and losses; none of these terms is defined in terms of the benchmark accuracies they are claimed to explain. The empirical claims rest on leave-one-domain-out accuracy against external baselines (Tables 1-4), and the ablation (Table 5) is an incremental loss-term study, not a fitted-parameter prediction. There are no self-citations by the present authors; the method builds on cited prior work (SAM, SAGM, SAMALTDG) rather than on an author-specific uniqueness theorem. The appendix does contain a self-undermining admission: Table 7 shows Digits-DG has a uniform QT = 0.1, i.e., it is class-balanced rather than long-tailed, and Section 5.3 describes a federated FedAvg loop whose results never appear in Tables 1-4. These are evidentiary gaps for the paper's 'federated, long-tailed' claims, but they are not circular reductions: no equation equals its input by construction, and no fitted value is re-reported as a prediction. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The method depends on two domain assumptions: that flat minima help OOD generalization, and that adversarial alignment helps. It also introduces Q_T as a target-like distribution, but Q_T is derived from training labels and computed as class frequency, making the alignment term partially self-confirming. Hyperparameters rho and alpha are tuned without reported searches.

free parameters (3)
  • rho (SAM perturbation radius) = 0.05
    Default value set in Section 5.3; controls all sharpness perturbations, including class-wise epsilon_c in Eq 9 and entropy perturbation in Eq 12.
  • alpha (gradient coherence coefficient) = not reported
    Weight of Lcoh in Eq 8; controls trade-off between classification and adversarial gradients, but no search protocol or final value is given.
  • gamma (maximum square loss coefficient) = 1
    Mentioned in Section 5.3, though the maximum square loss of Eq 7 does not appear in the final FedTAIL objective Eq 13.
assumptions (5)
  • domain assumption Flat minima generalize better out-of-distribution
    The entire method inherits this from SAM [8] and related work; Section 3.3 and 4 rely on it without further justification.
  • domain assumption Adversarial domain alignment encourages domain-invariant representations
    Eq 2 and the coherence term Eq 8 assume fooling a domain discriminator improves target-domain accuracy.
  • domain assumption PACS, OfficeHome, Digits-DG, and mini-DomainNet contain meaningful long-tailed class imbalance
    The paper's abstract and introduction claim long-tailed settings, but Section 5.1 describes standard datasets without constructing imbalanced splits; this premise appears questionable.
  • ad hoc to paper Q_T estimated from training class frequencies is a valid proxy for the target conditional distribution
    Appendix Tables 6-7 compute Q_T as class frequency, while Eq 12 calls it a target-like predictive distribution from a momentum model; no justification reconciles these.
  • ad hoc to paper The largest Hessian eigenvalue sigma_max can be estimated reliably per class
    Eq 11 assumes a practical per-class Hessian computation, but the paper does not describe how sigma_max is estimated in a federated, mini-batch setting.
invented entities (1)
  • Target-like conditional distribution Q_T
    purpose: Used as the reference distribution in the sharpness-aware entropy KL term Eq 12 to align conditional predictions across domains.
    Q_T is described as momentum-updated predictive distribution in Section 4, but computed as raw per-domain class frequency in the appendix. It has no external validation and is estimated from the training data itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FEDTAIL: Federated Long-Tailed Domain Generalization with Sharpness-Guided Gradient Matching." pith.science (2026). https://pith.science/paper/M4T6LD6H

@misc{pith2026250608518,
  author       = {Pith},
  title        = {Pith review of: FEDTAIL: Federated Long-Tailed Domain Generalization with Sharpness-Guided Gradient Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4T6LD6H}},
  note         = {Machine review of arXiv:2506.08518}
}
read the original abstract

Domain Generalization (DG) seeks to train models that perform reliably on unseen target domains without access to target data during training. While recent progress in smoothing the loss landscape has improved generalization, existing methods often falter under long-tailed class distributions and conflicting optimization objectives. We introduce FedTAIL, a federated domain generalization framework that explicitly addresses these challenges through sharpness-guided, gradient-aligned optimization. Our method incorporates a gradient coherence regularizer to mitigate conflicts between classification and adversarial objectives, leading to more stable convergence. To combat class imbalance, we perform class-wise sharpness minimization and propose a curvature-aware dynamic weighting scheme that adaptively emphasizes underrepresented tail classes. Furthermore, we enhance conditional distribution alignment by integrating sharpness-aware perturbations into entropy regularization, improving robustness under domain shift. FedTAIL unifies optimization harmonization, class-aware regularization, and conditional alignment into a scalable, federated-compatible framework. Extensive evaluations across standard domain generalization benchmarks demonstrate that FedTAIL achieves state-of-the-art performance, particularly in the presence of domain shifts and label imbalance, validating its effectiveness in both centralized and federated settings. Code: https://github.com/sunnyinAI/FedTail

Figures

Figures reproduced from arXiv: 2506.08518 by the authors.

Figure 1
Figure 1. Accuracy vs. Epoch comparison between FedTAIL and DGviaER across different domains (Art, Cartoon, Photo, Sketch) on the PACS dataset using ResNet-50 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualizations of feature embeddings across class and domain. Top row: [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages

  1. [1]

    Towards principled disentanglement for domain generalization

    Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. Towards principled disentanglement for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8014–8024, 2022

  2. [2]

    Learning to learn single domain generalization

    Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12556–12565, 2020

  3. [3]

    Balaji, S

    Y . Balaji, S. Sankaranarayanan, and R. Chellappa. Metareg: Towards domain generalization using meta- regularization. In NeurIPS, 2018

  4. [4]

    Domain generalization via invariant feature representation

    Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning, pages 10–18. PMLR, 2013

  5. [5]

    Learning to generalize: Meta-learning for domain generalization

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

  6. [6]

    Semantic data augmentation based distance metric learning for domain generalization

    Mengzhu Wang, Jianlong Yuan, Qi Qian, Zhibin Wang, and Hao Li. Semantic data augmentation based distance metric learning for domain generalization. InProceedings of the 30th ACM international conference on multimedia, pages 3214–3223, 2022

  7. [7]

    Domain generalization via entropy regularization

    Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao. Domain generalization via entropy regularization. Advances in neural information processing systems, 33:16096–16107, 2020

  8. [8]

    Sharpness-aware minimization for efficiently improving generalization

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020

Show all 60 references
  1. [9]

    Escaping saddle points for effective generalization on class-imbalanced data

    Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, et al. Escaping saddle points for effective generalization on class-imbalanced data. Advances in Neural Information Processing Systems, 35:22791–22805, 2022

  2. [10]

    Domain adaptation for semantic segmentation with maximum squares loss

    Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2090–2099, 2019

  3. [11]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004

  4. [12]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016

  5. [13]

    Domain generalization with adversarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, pages 5400–5409, 2018

  6. [14]

    Learning de-biased represen- tations with biased representations

    Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased represen- tations with biased representations. In International conference on machine learning, pages 528–539. PMLR, 2020

  7. [15]

    Adaptive risk minimization: Learning to adapt to domain shift

    Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems , 34:23664–23678, 2021. 11 FedTail

  8. [16]

    Domain generalization via model-agnostic learning of semantic features

    Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, volume 32, 2019

  9. [17]

    Domain adaptive ensemble learning.IEEE Transactions on Image Processing, 30:8008–8018, 2021

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain adaptive ensemble learning.IEEE Transactions on Image Processing, 30:8008–8018, 2021

  10. [18]

    Generalizing across domains via cross-gradient training

    Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations (ICLR), 2018

  11. [19]

    Domain generaliza- tion by solving jigsaw puzzles

    Fabio M Carlucci, Antonio D’Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generaliza- tion by solving jigsaw puzzles. In CVPR, pages 2229–2238, 2019

  12. [20]

    Domain agnostic learning with disentangled representations

    Xingchao Peng, Zijun Huang, Ximeng Sun, and Kate Saenko. Domain agnostic learning with disentangled representations. In ICML, pages 5102–5112, 2019

  13. [21]

    Undoing the damage of dataset bias

    Aditya Khosla, Tinghui Zhou, Tomasz Malisiewicz, Alexei A Efros, and Antonio Torralba. Undoing the damage of dataset bias. In Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 158–171. Spr...

  14. [22]

    Cross-domain face presentation attack detection via multi-domain disentangled representation learning

    Guoqing Wang, Hu Han, Shiguang Shan, and Xilin Chen. Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6678–6687, 2020

  15. [23]

    Out-of-distribution generalization via risk extrapolation (rex)

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International conference on machine learning, pages 5815–5826. PMLR, 2021

  16. [24]

    Invariant risk minimization

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

  17. [25]

    Domain generalization via gradient surgery

    Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6630–6638, 2021

  18. [26]

    Towards efficient and scalable sharpness- aware minimization

    Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. Towards efficient and scalable sharpness- aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12360–12370, 2022

  19. [27]

    Efficient sharpness-aware minimization for improved training of neural networks.arXiv preprint arXiv:2110.03141, 2021

    Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, and Vincent YF Tan. Efficient sharpness-aware minimization for improved training of neural networks.arXiv preprint arXiv:2110.03141, 2021

  20. [28]

    Simplifying neural nets by discovering flat minima

    Sepp Hochreiter and Jürgen Schmidhuber. Simplifying neural nets by discovering flat minima. Advances in neural information processing systems, 7, 1994

  21. [29]

    On large-batch training for deep learning: Generalization gap and sharp minima

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016

  22. [30]

    Sharp minima can generalize for deep nets

    Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can generalize for deep nets. In International Conference on Machine Learning, pages 1019–1028. PMLR, 2017

  23. [31]

    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. Advances in Neural Information Processing Systems, 34:22405–22418, 2021

  24. [32]

    Learning from extrinsic and intrinsic supervisions for domain generalization

    Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng. Learning from extrinsic and intrinsic supervisions for domain generalization. In European Conference on Computer Vision, pages 159–176. Springer, 2020

  25. [33]

    Sharpness-aware model-agnostic long-tailed domain generalization

    Houcheng Su, Weihao Luo, Daixian Liu, Mengzhu Wang, Jing Tang, Junyang Chen, Cong Wang, and Zhenghan Chen. Sharpness-aware model-agnostic long-tailed domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15091–15099, 2024

  26. [34]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180–1189, 2015

  27. [35]

    Sharpness-aware gradient matching for domain generalization

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

  28. [36]

    D. Li, Y . Yang, Y .-Z. Song, and T. M. Hospedales. Deeper, broader and artier domain generalization. InICCV, 2017. 12 FedTail

  29. [37]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018–5027, 2017

  30. [38]

    Learning to generate novel domains for domain generalization

    Kaiyang Zhou, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In European Conference on Computer Vision (ECCV), pages 561–578, 2020

  31. [39]

    X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019

  32. [40]

    D’Innocente and B

    A. D’Innocente and B. Caputo. Domain generalization with domain-specific aggregation modules. In GCPR, 2018

  33. [41]

    Statistical learning theory

    Vladimir Naumovich Vapnik. Statistical learning theory. Wiley-Interscience, 1998

  34. [42]

    Epi-fcr: Episodic fine-grained cross-domain few-shot learning via feature calibration and relation alignment

    Yuhang Li, Yaqing Wang, Yifan Li, Yuxin Li, Yujie Zhang, and Liang Wang. Epi-fcr: Episodic fine-grained cross-domain few-shot learning via feature calibration and relation alignment. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1234–1243. ACM, 2022

  35. [43]

    Yufei Wang, Haoliang Li, and Alex C. Kot. Heterogeneous domain generalization via domain mixup. arXiv preprint arXiv:2009.05448, 2020

  36. [44]

    Hospedales

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 5542–5550, Oct 2017

  37. [45]

    Embracing the dark knowledge: Domain generalization using regularized knowledge distillation

    Yufei Wang, Haoliang Li, Lap-pui Chau, and Alex C Kot. Embracing the dark knowledge: Domain generalization using regularized knowledge distillation. In ACM International Conference on Multimedia, pages 2595–2604, 2021

  38. [46]

    Symmetric self-paced learning for domain generalization

    Di Zhao, Yun Sing Koh, Gillian Dobbie, Hongsheng Hu, and Philippe Fournier-Viger. Symmetric self-paced learning for domain generalization. In AAAI Conference on Artificial Intelligence, volume 38, pages 16961–16969, 2024

  39. [47]

    Domain generalization via entropy regularization

    Shiqi Zhao, Mingkui Gong, Ting Liu, Yunchao Fu, and Dacheng Tao. Domain generalization via entropy regularization. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 16096–16107, 2020

  40. [48]

    Deep domain-adversarial image generation for domain generalisation

    Kaiyang Zhou, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Deep domain-adversarial image generation for domain generalisation. In AAAI Conference on Artificial Intelligence (AAAI), pages 13025–13032, 2020

  41. [49]

    A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance

    Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In International Conference on Computer Vision, pages 11685–11695, 2023

  42. [50]

    Self-challenging improves cross-domain generalization

    Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124–140. Springer, 2020

  43. [51]

    A fourier-based framework for domain generalization

    Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383– 14392, 2021

  44. [52]

    Domain generalization using causal matching

    Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In ICML, pages 7313–7324, 2021

  45. [53]

    Feature stylization and domain-aware contrastive learning for domain generalization

    Seogkyu Jeon, Kibeom Hong, Pilhyeon Lee, Jewook Lee, and Hyeran Byun. Feature stylization and domain-aware contrastive learning for domain generalization. In ACM Multimedia, pages 22–31, 2021

  46. [54]

    Domain generalization via frequency-domain-based feature disentanglement and interaction

    Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and interaction. In Proceedings of the 30th ACM International Conference on Multimedia, MM ’22, page 4821–4829, New York, NY , USA, 2...

  47. [55]

    Pcl: Proxy- based contrastive learning for domain generalization

    Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy- based contrastive learning for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7087–7097, 2022

  48. [56]

    Style neophile: Constantly seeking novel styles for domain generalization

    Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7130–7140, June 2022

  49. [57]

    Improving generalization with domain convex game

    Fangrui Lv, Jian Liang, Shuang Li, Jinming Zhang, and Di Liu. Improving generalization with domain convex game. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24315– 24324, 2023. 13 FedTail

  50. [58]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, 2015

  51. [59]

    J. Deng, W. Dong, R. Socher, L.J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009

  52. [60]

    A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance

    Zeyi Huang, Andy Zhou, Zijian Lin, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. arXiv preprint arXiv:2309.12530, 2023. 14

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.