Pith. sign in

REVIEW 3 major objections 6 minor 30 references

UniVarFL: Uniformity and Variance Regularized Federated Learning for Heterogeneous Data

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Two local regularizers let federated clients train as if their data were IID, the paper claims, matching or beating global-regularization methods without global-model comparisons.

desk verdict A well-motivated local-only regularizer pair that is undermined by the paper's own STL-10 α=1.0 result, where UniVarFL is worst by ~10 points. read the letter →

arxiv 2506.08167 v1 pith:QLIFEICJ submitted 2025-06-09 cs.LG cs.AIcs.CVcs.DC

Sunny Gupta , Nikita Jangid , Amit Sethi This is my paper
classification cs.LGcs.AIcs.CVcs.DC
keywords federatedlearningnon-IIDdataclientdriftclassifiervarianceregularizationhypersphericaluniformitylabelshiftfeaturelocaltraining
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

UniVarFL is a proposed federated-learning framework whose central claim is that non-IID performance loss can be fixed entirely inside each client's local training, without comparing local models to the global model or freezing layers. It adds two regularization terms to the usual cross-entropy loss: one that keeps class-wise predicted-probability variance above a floor derived from an idealized IID one-hot setup, and one that spreads feature embeddings uniformly on the hypersphere. The paper argues these terms counter local classifier bias and feature collapse, and reports higher accuracy than FedAvg, FedProx, MOON, and a freezing baseline on label-shift and feature-shift benchmarks. A convergence theorem is also stated, bounding the expected gradient norm of the global objective at $O(1/\sqrt{TE})$ under smoothness and bounded-heterogeneity assumptions. If these claims hold, the practical payoff is a scalable FL method that avoids expensive global-model comparisons.

What carries the argument

The machinery is the combined local loss $\mathcal{L} = \mathcal{L}_{CE} + \mu \mathcal{L}_{HE} + \lambda \mathcal{L}_V$. The classifier variance term $\mathcal{L}_V = \frac{1}{D}\sum_{j=1}^D \max(0, c - \mathrm{Var}_{x_i\sim X}(f_\theta(x_i)_j))$ enforces a per-class variance floor, where $D$ is the number of classes and the floor $c$ is defined as the average per-column variance of a $D\times D$ identity matrix, i.e. the variance expected from perfect one-hot predictions. The hyperspherical energy term $\mathcal{L}_{HE} = \frac{1}{n^2}\sum_{i,j}(1 - z_i^\top z_j + \epsilon)^{-1}$ penalizes small angular separation between $\ell_2$-normalized feature vectors $z_i = g_\theta(x_i)$. The threshold $c$ is the object that carries the variance argument: it translates the intuition of IID-like predictions into a fixed numeric target that each client enforces locally, while the energy term plays the analogous role for feature geometry.

What would settle it

Run UniVarFL on a low-class-count dataset with mild label skew (e.g., STL-10 at $\alpha=1.0$), sweeping the variance floor $c$ from near zero to above the identity-derived value, and compare test accuracy against FedAvg: if accuracy improves at a $c$ different from the default and drops below FedAvg at the default, the fixed-floor assumption is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that client drift under non-IID data can be counteracted by a pair of local-only regularizers: classifier variance regularization $\mathcal{L}_V$, which penalizes class-wise prediction variance falling below a threshold $c$ computed from the variance of one-hot identity vectors, and hyperspherical energy regularization $\mathcal{L}_{HE}$, which penalizes pairs of $\ell_2$-normalized feature embeddings that are too close in cosine similarity. The authors argue that these two terms jointly emulate IID-like training dynamics at the client level, eliminating the need for global model dependency while addressing both label shift and feature shift. They support this with experiments showing UniVarFL outperforming FedAvg, FedProx, MOON, and Freeze in most reported settings, including partial client participation, and with a nonconvex convergence result stating that the global objective reaches an approximate stationary point at rate $O(1/\sqrt{TE})$.

Load-bearing premise

The load-bearing premise is that the class-wise variance floor $c$, computed once from one-hot identity vectors, is a universally good proxy for IID prediction statistics; if it is too high or too low for a given dataset and skew level, the variance regularizer can reduce accuracy instead of preventing bias.

Editorial extensions

If this is right

  • If UniVarFL is correct, client drift can be reduced without layer-wise weight comparisons, extra forward passes, or frozen classifier layers, lowering per-round compute.
  • The same two regularizers apply to both label-shift and feature-shift heterogeneity, extending beyond Dirichlet-only evaluations.
  • Under partial client participation, UniVarFL is reported to hold an accuracy advantage, suggesting the method tolerates realistic participation ratios.
  • The stated $O(1/\sqrt{TE})$ convergence rate, if valid, puts UniVarFL on par with standard nonconvex local-SGD analyses.

Reading between the lines

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

  • A testable extension is replacing the fixed identity-derived floor $c$ with a per-dataset or per-round estimate from an actual IID batch; the STL-10 $\alpha=1.0$ result reported in the paper, where UniVarFL (59.9%) trails FedAvg (69.1%) and Freeze (72.0%), suggests the fixed floor can be miscalibrated under mild skew.
  • Because both regularizers are local and require no global comparisons, they could in principle be composed with aggregation-side methods or personalized FL schemes, though the paper does not test such combinations.
  • An ablation that varies $\mu$ and $\lambda$ independently on a feature-shift dataset would test whether classifier-variance and hyperspherical-uniformity regularizers address distinct failure modes or merely duplicate the same effect.
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

3 major / 6 minor

Summary. The paper proposes UniVarFL, a federated learning method that adds two client-side regularizers to local training: a classifier variance regularizer (LV) that penalizes class-wise prediction-variance collapse relative to a fixed threshold c, and a hyperspherical energy regularizer (LHE) that encourages angular uniformity of normalized feature embeddings. The method is evaluated on STL-10, CIFAR-100, PACS, and HAM10000 under label-shift and feature-shift settings, and a nonconvex convergence theorem is stated in Section 6. The central claim is that UniVarFL outperforms existing FL methods in accuracy while avoiding global-model dependence and extra computational overhead.

Significance. If the empirical claim were fully supported, UniVarFL would be a useful addition: a local-only, cheap regularizer pair that reduces client drift and avoids the extra forward passes or global-model comparisons of FedProx, MOON, and FedAlign. The paper also provides a public code repository, experiments across multiple datasets, and a client-participation analysis. However, the paper's own Table 1 contains a large counterexample to the unqualified superiority claim: on STL-10 with alpha=1.0, UniVarFL (59.9) is about 10 points below FedAvg (69.1) and Freeze (72.0). In addition, the convergence result in Section 6 is a generic local-SGD bound that does not use the specific regularizer structure and is not proved. These issues materially weaken the central claims as currently stated, though they are addressable in revision.

major comments (3)
  1. [Abstract, §5 Table 1, §7] The abstract and conclusion claim that UniVarFL 'outperforms existing methods in accuracy' and 'achieves state-of-the-art performance', but Table 1 directly contradicts this for STL-10 with alpha=1.0: UniVarFL reaches 59.9±0.6, while FedAvg reaches 69.1±0.6 and Freeze reaches 72.0±0.1. The Results section itself concedes that 'Freeze excels on STL-10', so the unqualified superiority claim is not supported by the paper's own evidence. Please revise the claims to state precisely the settings in which UniVarFL is competitive or superior, or re-tune hyperparameters and re-run the experiments to remove this counterexample.
  2. [§6, Theorem 1] Theorem 1 is stated without proof, and the assumptions (A1)-(A3) together with the displayed bound are exactly a standard nonconvex local-SGD analysis. The theorem does not use the specific forms of L_V or L_HE except through generic smoothness/heterogeneity constants; in particular, L_V contains the nonsmooth operation max(0, c - Var(...)), and L_HE contains the reciprocal term (1 - cos(theta) + epsilon)^{-1}, neither of which is analyzed. As written, the theorem is a restatement of existing convergence results rather than a guarantee for the UniVarFL objective. Please provide a proof that explicitly handles the regularizers, or clearly state that the theoretical result is a generic bound that does not distinguish UniVarFL from other local-SGD methods.
  3. [§4, Hyperparameters; §3.1 Eq. (2)] The hyperparameters lambda=D/4 and mu=0.5 are tuned only on CIFAR-100 and then applied to all datasets, including STL-10, PACS, and HAM10000. The variance floor c in Eq. (2) is computed from a D x D identity matrix and therefore depends on the number of classes, but its scaling relative to the actual softmax variance may not transfer across datasets. The large STL-10 alpha=1.0 degradation could plausibly be caused by this fixed configuration. Please report a hyperparameter sensitivity analysis or per-dataset tuning results, and at minimum show that the chosen fixed values are not responsible for the STL-10 result.
minor comments (6)
  1. [§3.2, Eq. (3)] The sum in Eq. (3) runs over all i,j, including i=j, which contributes a constant term 1/epsilon, and each unordered pair (i,j) with i≠j is counted twice. Please restrict the sum to i<j or otherwise define the double-counting convention explicitly.
  2. [§3.2] The phrase 'Theoretical and Empirical Justification: This approach is inspired by insights from hyperspherical representation learning...' appears twice verbatim in the same subsection; please delete the duplicate.
  3. [§5, Table 2 and surrounding text] The text defines the participation ratio as rho, but Table 2 uses p (e.g., p = 0.1). Please unify the notation.
  4. [§4, Models] The Models paragraph says 'For STL-10 and PACS, we use a lightweight CNN...' and then 'For CIFAR-100 and PACS, we employ a ResNet-18...'; the second mention of PACS is likely a typo for HAM10000. Please correct.
  5. [§6] The local objective in Section 6 uses the acronyms LHUR and LCVR, while the rest of the paper uses LHE and LV. Please use consistent notation throughout.
  6. [§6.1, Theorem 1] The theorem statement writes the rate as O(1/sqrt(T E)), but the exact displayed bound uses min over t of the expected gradient norm; please clarify whether T is the number of communication rounds and E is the number of local steps, and whether the bound holds for the final iterate or the minimum over iterates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the variance floor is a closed-form constant, no self-citations are load-bearing, and the empirical claims are not forced by construction.

full rationale

The derivation chain is self-contained. The classifier variance threshold c in Eq. (2) is a closed-form function of the class count D (c=(D-1)/D^2 for the identity matrix), computed from one-hot vectors rather than fitted to any benchmark or derived from the method's own outputs. The regularizers in Eqs. (1)-(3) define a new client-side objective but do not encode the experimental results they are claimed to explain. Hyperparameters lambda=D/4 and mu=0.5 are tuned on CIFAR-100 and transferred to other datasets, which is ordinary cross-dataset hyperparameter transfer, not fitting a predictor to the target datasets. The reference list contains no self-citations by Gupta/Jangid/Sethi, so no argument rests on an author-imported uniqueness theorem. Theorem 1 is admittedly a generic local-SGD-style bound, stated as 'consistent with classical nonconvex SGD analysis and existing work on local/parallel SGD,' and it does not substantively use the specific regularizer terms; this is a novelty/support weakness rather than a circular reduction, because the bound is not equivalent to the regularizer definitions and is not used to fit any empirical result. The STL-10 alpha=1.0 result (UniVarFL 59.9 vs FedAvg 69.1 and Freeze 72.0) conflicts with the abstract's blanket 'outperforms' claim, but that is an empirical-support/correctness issue, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on two tuned hyperparameters and an unproven set of standard FL assumptions for the convergence theorem. The variance threshold c is an ad hoc reference derived from one-hot vectors. No new entities are introduced.

free parameters (3)
  • mu (LHE weight) = 0.5
    Tuned on CIFAR-100; controls the strength of the hyperspherical uniformity loss.
  • lambda (LV weight) = D/4 (D = number of classes)
    Set as a heuristic depending on the number of classes; tuned on CIFAR-100.
  • epsilon in LHE = not specified
    Numerical stability constant in Eq. 3; value not given in the paper.
assumptions (4)
  • domain assumption The local loss functions are L-smooth (A1).
    Assumption A1 in Section 6, needed for the convergence theorem.
  • domain assumption Bounded local gradient variance (A2).
    Assumption A2 in Section 6, standard in FL analysis.
  • domain assumption Bounded gradient heterogeneity (A3).
    Assumption A3 in Section 6, standard in FL analysis.
  • ad hoc to paper The class-wise variance floor c, computed from one-hot identity vectors, is an appropriate reference for IID prediction variance.
    Defined in Eq. 2; this assumption is not validated and may be the cause of the STL-10 α=1.0 degradation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniVarFL: Uniformity and Variance Regularized Federated Learning for Heterogeneous Data." pith.science (2026). https://pith.science/paper/QLIFEICJ

@misc{pith2026250608167,
  author       = {Pith},
  title        = {Pith review of: UniVarFL: Uniformity and Variance Regularized Federated Learning for Heterogeneous Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLIFEICJ}},
  note         = {Machine review of arXiv:2506.08167}
}
read the original abstract

Federated Learning (FL) often suffers from severe performance degradation when faced with non-IID data, largely due to local classifier bias. Traditional remedies such as global model regularization or layer freezing either incur high computational costs or struggle to adapt to feature shifts. In this work, we propose UniVarFL, a novel FL framework that emulates IID-like training dynamics directly at the client level, eliminating the need for global model dependency. UniVarFL leverages two complementary regularization strategies during local training: Classifier Variance Regularization, which aligns class-wise probability distributions with those expected under IID conditions, effectively mitigating local classifier bias; and Hyperspherical Uniformity Regularization, which encourages a uniform distribution of feature representations across the hypersphere, thereby enhancing the model's ability to generalize under diverse data distributions. Extensive experiments on multiple benchmark datasets demonstrate that UniVarFL outperforms existing methods in accuracy, highlighting its potential as a highly scalable and efficient solution for real-world FL deployments, especially in resource-constrained settings. Code: https://github.com/sunnyinAI/UniVarFL

Figures

Figures reproduced from arXiv: 2506.08167 by the authors.

Figure 1
Figure 1. Presents the singular values of the classifier’s final layer weights, obtained from a model trained on the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages

  1. [1]

    Communication- efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication- efficient learning of deep networks from decentralized data. InArtificial intelligence and statistics, pages 1273–1282. PMLR, 2017

    2017

  2. [2]

    A survey on distributed machine learning.Acm computing surveys (csur), 53(2):1–33, 2020

    Joost Verbraeken, Matthijs Wolting, Jonathan Katzy, Jeroen Kloppenburg, Tim Verbelen, and Jan S Rellermeyer. A survey on distributed machine learning.Acm computing surveys (csur), 53(2):1–33, 2020

    work page 2020

  3. [3]

    Federated learning based on dynamic regularization.arXiv preprint arXiv:2111.04263, 2021

    Durmus Alp Emre Acar, Yue Zhao, Ramon Matas Navarro, Matthew Mattina, Paul N Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization.arXiv preprint arXiv:2111.04263, 2021

    arXiv 2021

  4. [4]

    Model-contrastive federated learning

    Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021

    work page 2021

  5. [5]

    Federated optimization in heterogeneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020

    2020

  6. [6]

    No fear of heterogeneity: Classifier calibration for federated learning with non-iid data.Advances in Neural Information Processing Systems, 34:5972– 5984, 2021

    Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, Jian Liang, and Jiashi Feng. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data.Advances in Neural Information Processing Systems, 34:5972– 5984, 2021

    work page 2021

  7. [7]

    Gradaug: A new regularization method for deep neural networks

    Taojiannan Yang, Sijie Zhu, and Chen Chen. Gradaug: A new regularization method for deep neural networks. Advances in neural information processing systems, 33:14207–14218, 2020

    work page 2020

  8. [8]

    Fedbabu: Towards enhanced representation for federated image classification.arXiv preprint arXiv:2106.06042, 2021

    Jaehoon Oh, Sangmook Kim, and Se-Young Yun. Fedbabu: Towards enhanced representation for federated image classification.arXiv preprint arXiv:2106.06042, 2021

    arXiv 2021

Show all 30 references
  1. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  2. [10]

    Learning multiple layers of features from tiny images.https://www

    Krizhevsky Alex. Learning multiple layers of features from tiny images.https://www. cs. toronto. edu/kriz/learning- features-2009-TR. pdf, 2009

    2009

  3. [11]

    Scaffold: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. InInternational conference on machine learning, pages 5132–5143. PMLR, 2020. 8 PRIME AI paper

    2020

  4. [12]

    Local learning matters: Rethinking data heterogeneity in federated learning

    Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. Local learning matters: Rethinking data heterogeneity in federated learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8397–8406, 2022

    2022

  5. [13]

    Is your data relevant?: Dynamic selection of relevant data for federated learning

    Lokesh Nagalapatti, Ruhi Sharma Mittal, and Ramasuri Narayanam. Is your data relevant?: Dynamic selection of relevant data for federated learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 7859–7867, 2022

    2022

  6. [14]

    Fedcor: Correlation-based active client selection strategy for heterogeneous federated learning

    Minxue Tang, Xuefei Ning, Yitu Wang, Jingwei Sun, Yu Wang, Hai Li, and Yiran Chen. Fedcor: Correlation-based active client selection strategy for heterogeneous federated learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10102–...

    2022

  7. [15]

    Fedmix: Approximation of mixup under mean augmented federated learning.arXiv preprint arXiv:2107.00233, 2021

    Tehrim Yoon, Sumin Shin, Sung Ju Hwang, and Eunho Yang. Fedmix: Approximation of mixup under mean augmented federated learning.arXiv preprint arXiv:2107.00233, 2021

    2021 arXiv

  8. [16]

    Differentially private federated learning with local regularization and sparsification

    Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10122–10131, 2022

    2022

  9. [17]

    Byzantine-robust learning on heterogeneous datasets via bucketing.arXiv preprint arXiv:2006.09365, 2020

    Sai Praneeth Karimireddy, Lie He, and Martin Jaggi. Byzantine-robust learning on heterogeneous datasets via bucketing.arXiv preprint arXiv:2006.09365, 2020

    2006 arXiv

  10. [18]

    Heterofl: Computation and communication efficient federated learning for heterogeneous clients.arXiv preprint arXiv:2010.01264, 2020

    Enmao Diao, Jie Ding, and Vahid Tarokh. Heterofl: Computation and communication efficient federated learning for heterogeneous clients.arXiv preprint arXiv:2010.01264, 2020

    2010 arXiv

  11. [19]

    Smartidx: Reducing communication cost in federated learning by exploiting the cnns structures

    Donglei Wu, Xiangyu Zou, Shuyu Zhang, Haoyu Jin, Wen Xia, and Binxing Fang. Smartidx: Reducing communication cost in federated learning by exploiting the cnns structures. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 4254–4262, 2022

    2022

  12. [20]

    Achieving linear speedup with partial worker participation in non-iid federated learning.arXiv preprint arXiv:2101.11203, 2021

    Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-iid federated learning.arXiv preprint arXiv:2101.11203, 2021

    2021 arXiv

  13. [21]

    Closing the generalization gap of cross-silo federated medical image segmentation

    An Xu, Wenqi Li, Pengfei Guo, Dong Yang, Holger R Roth, Ali Hatamizadeh, Can Zhao, Daguang Xu, Heng Huang, and Ziyue Xu. Closing the generalization gap of cross-silo federated medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2022

  14. [22]

    Cd2-pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning

    Yiqing Shen, Yuyin Zhou, and Lequan Yu. Cd2-pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041–10050, 2022

    2022

  15. [23]

    Fine-tuning global model via data-free knowledge distillation for non-iid federated learning

    Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10174–10183, 2022

    2022

  16. [24]

    Measuring the effects of non-identical data distribution for federated visual classification.arXiv preprint arXiv:1909.06335, 2019

    Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification.arXiv preprint arXiv:1909.06335, 2019

    1909 arXiv

  17. [25]

    Layer-wised model aggregation for personalized federated learning

    Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Layer-wised model aggregation for personalized federated learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10092–10101, 2022

    2022

  18. [26]

    Personalized federated learning using hypernetworks

    Aviv Shamsian, Aviv Navon, Ethan Fetaya, and Gal Chechik. Personalized federated learning using hypernetworks. InInternational Conference on Machine Learning, pages 9489–9502. PMLR, 2021

    2021

  19. [27]

    Bayesian nonparametric federated learning of neural networks

    Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. InInternational conference on machine learning, pages 7252–7261. PMLR, 2019

    2019

  20. [28]

    Federated learning with matched averaging.arXiv preprint arXiv:2002.06440, 2020

    Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging.arXiv preprint arXiv:2002.06440, 2020

    2002 arXiv

  21. [29]

    Federated learning with position-aware neurons

    Xin-Chun Li, Yi-Chu Xu, Shaoming Song, Bingshuai Li, Yinchuan Li, Yunfeng Shao, and De-Chuan Zhan. Federated learning with position-aware neurons. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10082–10091, 2022

    2022

  22. [30]

    Ensemble distillation for robust model fusion in federated learning.Advances in neural information processing systems, 33:2351–2363, 2020

    Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning.Advances in neural information processing systems, 33:2351–2363, 2020. 9

    2020

Pith tools

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