Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Heterogeneous Federated Learning with Prototype Alignment and Upscaling

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ProtoNorm improves prototype-based federated learning by spreading global prototypes on a unit sphere and then upscaling them, and reports consistently higher classification accuracy under data and model heterogeneity than existing…

desk verdict A clean, honest empirical paper whose central synergy claim is undercut by its own ablation: the gains come from a tuned scalar, not from prototype alignment. read the letter →

arxiv 2507.04310 v1 pith:HZQQ7YHS submitted 2025-07-06 cs.LG cs.DC

Gyuejeong Lee , Jihwan Shin , Daeyoung Choi This is my paper
classification cs.LGcs.DC
keywords federatedlearningmodelheterogeneitydataprototype-basedprototypealignmentupscalingThomsonproblemhypersphericalenergy
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

This paper proposes Prototype Normalization (ProtoNorm), a prototype-based federated learning method that tackles the weak separation of class prototypes under heterogeneous data and model architectures. The authors claim that two server-side steps—spreading global prototypes on a unit sphere to maximize their angular separation, then magnifying them to increase Euclidean distance—produce more discriminative prototypes. They argue this makes ProtoNorm consistently outperform existing heterogeneous federated learning methods across CIFAR-10/100, Flowers-102, and Tiny ImageNet, while keeping the communication cost of prototype-based methods and moving all alignment computation to the server. If correct, the work shows that geometry alone, applied to shared prototypes rather than model parameters, can improve accuracy in federated settings without extra communication.

What carries the argument

The load-bearing object is the set of global class prototypes, normalized onto the unit sphere, optimized by a momentum-based gradient-ascent solver for the Thomson problem, the classical question of how N particles arrange themselves on a sphere to minimize repulsive potential energy. The solver minimizes the surrogate logarithmic hyperspherical energy E = Σ_{j<k} log(1/‖ĉ_j − ĉ_k‖), where each prototype moves under the repulsive force F_j = Σ_{k≠j} (ĉ_j − ĉ_k)/‖ĉ_j − ĉ_k‖², followed by renormalization. This produces maximally separated directions; the companion Prototype Upscaling step then scales the aligned unit prototypes by γ in the client-side regularization loss, which the paper identifies as necessary to avoid suboptimal local minima when prototypes remain unit vectors.

What would settle it

Run the full ProtoNorm pipeline but replace the Prototype Alignment step with randomly initialized unit vectors or with the unaligned FedProto prototypes, then re-tune γ per dataset; if accuracy on CIFAR-100, Flowers-102, and Tiny ImageNet does not fall below the reported ProtoNorm numbers, the alignment step is not the cause of the improvement. A sharper version: scale FedTGP's prototypes with its optimal γ from Table 5 and compare directly with ProtoNorm on the same three seeds; the paper reports 47.33% vs 47.41% on CIFAR-100, so a result at or above ProtoNorm's across all datasets would refute the claimed unique benefit of alignment.

Watch

Extended reading notes

Core claim

The central claim is that inter-prototype discrimination in prototype-based federated learning can be substantially improved by optimizing the global prototype configuration in two stages. In the first stage, Prototype Alignment treats the normalized global prototypes as charged particles on a unit sphere and runs a gradient-ascent solver for the Thomson problem, minimizing logarithmic hyperspherical energy so that prototypes repel each other into a well-separated angular arrangement. In the second stage, Prototype Upscaling multiplies these aligned unit vectors by a scalar γ on the client side before computing the prototype distance loss, giving the prototypes enough magnitude to guide local training. The authors report that this combination better separates prototypes and consistently outperforms six heterogeneous FL baselines, and that it also beats hyperspherical-regularization methods FedNH and FedUV in homogeneous settings. They further report that the alignment step alone, without upscaling, yields little or no accuracy gain over FedProto, so the paper's stated conclusion is that the two components work synergistically.

Load-bearing premise

The load-bearing premise is that the alignment step is what allows the upscaling step to work: if simply multiplying prototypes by a large number, without first spreading them on the sphere, gave the same accuracy, then ProtoNorm's distinctive contribution would collapse.

Editorial extensions

If this is right

  • If the central claim holds, any prototype-based federated learning pipeline can improve accuracy with a server-side post-processing step that costs no extra communication.
  • Clients no longer need to send per-class sample counts to the server, because the alignment step uses only the averaged prototype vectors, closing a privacy leak in weighted aggregation.
  • The alignment computation scales with the number of classes rather than the number of clients, which makes the method cheaper than contrastive prototype baselines as client count grows.
  • The reported results require per-dataset tuning of the scaling factor γ, so a direct corollary is that choosing γ is the main practical knob a user must set.
  • In homogeneous architecture settings, the same mechanism outperforms local hyperspherical-regularization methods, suggesting the server-side arrangement of prototypes transfers across both heterogeneous and homogeneous federations.

Reading between the lines

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

  • The paper's own numbers leave open that upscaling alone, not alignment, drives most of the gain: with tuned scaling, FedTGP reaches 47.33% versus ProtoNorm's 47.41% on CIFAR-100, so the unique role of Prototype Alignment is not yet pinned down.
  • A testable extension the authors do not explore is an adaptive or class-dependent scaling factor, derived from local prototype norms, which could remove the per-dataset grid search over γ.
  • Because the alignment acts only on a small set of vectors held by the server, the same Thomson-problem treatment could apply to classifier weight rows or logit anchors in other federated or centralized settings.
  • The experiments fix the decision-layer dimension at 512 (or 64 and 1024 in one study); whether the gains persist in very low-dimensional prototype spaces, where a sphere cannot accommodate many well-separated directions, remains open.
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

4 major / 5 minor

Summary. The paper proposes ProtoNorm, a prototype-based heterogeneous federated learning (HtFL) method. In each round, the server averages local prototypes (Eq. 7), applies a Thomson-problem-inspired gradient-ascent alignment on the unit hypersphere (Algorithm 1, Eqs. 8-14), and clients train with a prototype-distance regularizer in which the aligned global prototype is multiplied by a per-dataset scale gamma (Eq. 15). Experiments on CIFAR-10/100, Flowers-102, and Tiny ImageNet report accuracy gains over LG-FedAvg, FML, FedKD, FedDistill, FedProto, and FedTGP, plus comparisons with FedNH and FedUV in the homogeneous-model setting. The paper also includes scaling ablations, convergence plots, and communication-cost analysis.

Significance. If the claimed results hold, the contribution is practically useful: a communication-efficient, server-side geometric post-processing of prototypes that adds no client-side computation or communication overhead to prototype-based FL. The paper deserves credit for releasing code, covering several benchmark datasets and heterogeneity settings, providing a communication-cost table, and connecting the Thomson problem to prototype separation. However, the current evidence does not establish that the alignment component is necessary, nor that the full method is consistently superior to a suitably tuned scaling of existing prototypes. The central open question is whether the reported gains come from the proposed alignment or from the fitted scale parameter.

major comments (4)
  1. [Section 5.3, Tables 4-5] The claimed synergy between Prototype Alignment (PA) and Prototype Upscaling (PU) is not supported by the reported ablations. In Table 4, ProtoNorm with gamma=1 achieves 29.71% on CIFAR-100, while FedProto with gamma=1 in Table 5 achieves 29.97%, so the alignment step alone gives no measurable accuracy benefit. In Table 5, simply scaling FedTGP's existing prototypes by gamma=1.2 yields 47.33%, which is within 0.08 percentage points of ProtoNorm's 47.41% and well inside the reported standard deviation of +/-0.40. The paper never runs ProtoNorm with PU but without PA (e.g., scaling the simple average of Eq. (7) with no alignment), so the unique contribution of PA is asserted rather than demonstrated. I ask for this missing ablation with the same gamma search and multiple seeds, together with a significance test comparing ProtoNorm to FedTGP plus scaling.
  2. [Section 5.1 and Section 5.3] The scale parameter gamma is selected per dataset by grid search (values 100, 200, 200, and 400), and the results are highly sensitive to it: Table 4 shows CIFAR-100 accuracy varying from 29.71% at gamma=1 to 47.41% at gamma=200 and falling to 45.72% at gamma=1000, with Tiny ImageNet failing to converge at gamma=1000. Because the main accuracy gain is produced by this tuned scale, reporting test accuracy at the best grid-searched gamma without specifying a validation-based selection rule does not support the claim of 'consistent' superiority. Please report how gamma was chosen (e.g., a held-out validation split) and show the sensitivity of the comparison to that choice; the current Table 5 comparison shows that the gap over the strongest scaled baseline is smaller than the reported run-to-run noise.
  3. [Section 4.1, Eqs. (8)-(10)] The derivation of the surrogate energy is mathematically imprecise. The logarithmic form of the original energy is log E_s = log( sum_{j<k} h_s(||c_hat_j - c_hat_k||) ), which is not equal to the sum_{j<k} log h_s(...) minimized in Eq. (9). Jensen's inequality gives an inequality between these quantities, not an equivalence, so the claim that minimizing the original energy is 'equivalent' to minimizing the sum of logs is incorrect. Since Algorithm 1 is defined through this surrogate, the relationship between the PA objective and the stated Thomson problem should be restated as an approximation and justified, or the surrogate should be derived correctly.
  4. [Section 5.1] The evaluation protocol phrase 'average test accuracy across the best-performing local models' needs a precise operational definition. If 'best-performing' is determined after looking at the test set, the numbers are optimistically biased and comparisons across methods are unfair; if it means something else (e.g., best local validation checkpoint), that should be stated. Please specify the selection rule, confirm that it is applied identically to all baselines, and report the distribution of accuracy across the selected local models.
minor comments (5)
  1. [Section 5.1] The setup states that the FL system comprises 20 clients, but Table 2 reports experiments with 50 and 100 clients; please clarify whether 20 is the default and how the larger-client experiments are configured.
  2. [Table 4 caption] The caption says 'Classification accuracy of ProtoNorm on CIFAR-100 across varying gamma' but the table also contains Flowers-102 and Tiny ImageNet rows; the caption should be updated.
  3. [Figure 4 and surrounding text] The symbol d is used for both the decision-layer dimension and the class count; use K for the class count to avoid confusion.
  4. [Eq. (15)] The notation gamma * c_hat_j^G should be explicitly defined as scalar multiplication, and the text should consistently include gamma in the description of the local loss.
  5. [Section 5.3] The sentence 'ProtoNorm outperforms FedProto when applying the PU' is ambiguous because all ProtoNorm runs also include PA; separate the two effects when describing the ablation.

Circularity Check

2 steps flagged · score 6.0 of 10

ProtoNorm's reported gains reduce to the per-dataset tuned scaling γ; the paper's own ablation shows PA alone gives no gain and FedTGP plus scaling nearly matches ProtoNorm.

  1. fitted input called prediction [Section 5.3, 'Effect of Prototype Alignment' (Tables 4 and 5)]
    "However, when the PU is not applied, there is no noticeable performance gain, as evidenced by comparing the accuracy of ProtoNorm with γ = 1 (Table 4) against that of FedProto with γ = 1 (Table 5) for CIFAR-100. This result suggests that prototype alignment works significantly better with prototype upscaling."

    This is the paper's own evidence for the PA component. The numbers show ProtoNorm with PA and no PU (γ=1) scores 29.71 on CIFAR-100, below FedProto without PA (γ=1) at 29.97. So the alignment step, which is the first-principles novelty (Thomson-problem solver), provides zero measurable benefit; the entire gain attributed to ProtoNorm comes from the scaling factor γ. The 'synergy' claim is therefore not derived but is an assertion that is consistent only with the full, γ-tuned method. The novel component's contribution is nil by the paper's own ablation, so the central result reduces to the PU hyperparameter rather than to the alignment mechanism.

  2. fitted input called prediction [Section 5.1 'Federated Learning Setup' and Section 5.3 'Effect of Prototype Upscaling' (Table 5)]
    "Through grid search, we determined scale parameter γ values of 100, 200, 200, and 400 for CIFAR-10, CIFAR-100, Flowers-102 and Tiny ImageNet, respectively. ... Indeed, we found that their performance can be significantly improved by adjusting the scaling factor. For instance, FedTGP's accuracy on CIFAR-100 shows a remarkable increase from 36.92% to 47.33% with appropriate scaling."

    The scale γ is a fitted, dataset-specific constant chosen by grid search, and ProtoNorm's 'consistent' superiority is reported at exactly those fitted values. When the same fitted scaling is applied to FedTGP, its CIFAR-100 accuracy reaches 47.33, statistically indistinguishable from ProtoNorm's 47.41±0.40. Thus the claimed prediction 'consistently outperforms existing HtFL approaches' is forced by the fitted γ: the contrastive FedTGP baseline plus scaling alone reproduces the ProtoNorm result, so the alignment component is not an independent cause of the reported gain.

full rationale

This paper contains no load-bearing self-citation chain: the Thomson-problem energy functional is attributed to external work [22], and the FedProto/FedTGP baselines are external. The circularity is instead in the empirical derivation of the central claim. The paper's own ablation (Tables 4 and 5) shows that the novel Prototype Alignment component, evaluated alone at γ=1, does not improve over FedProto (29.71 vs 29.97 on CIFAR-100), and that the same tuned scaling applied to FedTGP reaches 47.33, within 0.08 (and within the reported ±0.40 standard deviation) of ProtoNorm's 47.41. The paper also states that 'when the PU is not applied, there is no noticeable performance gain,' conceding that the alignment step has no standalone effect. Consequently, the consistent-outperformance claim is not an independent consequence of the Thomson-problem alignment; it is a restatement of the per-dataset fitted scale parameter γ. This is partial circularity: the conceptual novelty (PA) is not supported, and the reported superiority is statistically forced by the tuned constant. Score 6.

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

The central mechanism is a post-processing step with two hyperparameters (gamma and PA settings). Gamma is fitted per dataset and the gain is largely driven by it; the alignment objective is a known approach from hyperspherical learning.

free parameters (3)
  • gamma = 100 (CIFAR-10), 200 (CIFAR-100), 200 (Flowers-102), 400 (Tiny ImageNet)
    Scale factor applied to aligned global prototypes. Selected per dataset via grid search to maximize test accuracy; the reported gains depend heavily on this choice.
  • PA learning rate eta = 0.1 with 95% decay every 10 iterations
    Chosen by hand for the gradient ascent optimizer; no sensitivity analysis is provided.
  • PA momentum mu = 0.9
    Chosen by hand for the momentum update in Algorithm 1; not varied in experiments.
assumptions (4)
  • domain assumption Minimizing hyperspherical energy of normalized global prototypes improves inter-prototype separation and downstream accuracy.
    Section 4.1 states 'We claim that prototypes arranged in a minimal energy configuration achieve better separation in Rd' without a proof or theoretical justification.
  • domain assumption Upscaling prototypes by gamma preserves consistency with the activation scale and helps local training.
    Section 4.2 asserts that small magnitudes lead to suboptimal local minima and that scaling fixes it; this is supported only by empirical tuning, not by analysis.
  • domain assumption Simple averaging in Eq. (7), without class-count weights, preserves enough information for the global prototypes to be useful.
    Section 4.1 changes the weighted average of FedProto to a simple average to reduce privacy risk, but no experiment or analysis verifies that this does not degrade prototype quality.
  • standard math The surrogate energy function in Eq. (9) is a valid proxy for the original hyperspherical energy.
    The authors appeal to Jensen's inequality to justify the log form; this is a standard mathematical step and is correctly cited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Heterogeneous Federated Learning with Prototype Alignment and Upscaling." pith.science (2026). https://pith.science/paper/HZQQ7YHS

@misc{pith2026250704310,
  author       = {Pith},
  title        = {Pith review of: Heterogeneous Federated Learning with Prototype Alignment and Upscaling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZQQ7YHS}},
  note         = {Machine review of arXiv:2507.04310}
}
read the original abstract

Heterogeneity in data distributions and model architectures remains a significant challenge in federated learning (FL). Various heterogeneous FL (HtFL) approaches have recently been proposed to address this challenge. Among them, prototype-based FL (PBFL) has emerged as a practical framework that only shares per-class mean activations from the penultimate layer. However, PBFL approaches often suffer from suboptimal prototype separation, limiting their discriminative power. We propose Prototype Normalization (ProtoNorm), a novel PBFL framework that addresses this limitation through two key components: Prototype Alignment (PA) and Prototype Upscaling (PU). The PA method draws inspiration from the Thomson problem in classical physics, optimizing global prototype configurations on a unit sphere to maximize angular separation; subsequently, the PU method increases prototype magnitudes to enhance separation in Euclidean space. Extensive evaluations on benchmark datasets show that our approach better separates prototypes and thus consistently outperforms existing HtFL approaches. Notably, since ProtoNorm inherits the communication efficiency of PBFL and the PA is performed server-side, it is particularly suitable for resource-constrained environments.

Figures

Figures reproduced from arXiv: 2507.04310 by the authors.

Figure 1
Figure 1. Spiral dataset and non-IID data distribution across clients. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of prototypes and learned representations in two-dimensional feature space on the spiral dataset. Different colors [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Comparison of normalized global prototype margins. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Pairwise global prototype distance during PA iterations. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Local prototypes and learned representations of clients. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Data distributions for the CIFAR-10 practical setting. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Data distributions for the CIFAR-100 practical setting. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 36 canonical work pages

  1. [1]

    Deep learning using rectified linear units (relu)

    Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018. 3

    arXiv 2018

  2. [2]

    Fe- drolex: Model-heterogeneous federated learning with rolling sub-model extraction

    Samiul Alam, Luyang Liu, Ming Yan, and Mi Zhang. Fe- drolex: Model-heterogeneous federated learning with rolling sub-model extraction. Advances in neural information pro- cessing systems, 35:29677–29690, 2022. 1

    work page 2022

  3. [3]

    Angular visual hardness

    Beidi Chen, Weiyang Liu, Zhiding Yu, Jan Kautz, Anshumali Shrivastava, Animesh Garg, and Animashree Anandkumar. Angular visual hardness. In International Conference on Machine Learning, pages 1637–1648. PMLR, 2020. 2

    work page 2020

  4. [4]

    Tackling data heterogeneity in federated learning with class prototypes

    Yutong Dai, Zeyuan Chen, Junnan Li, Shelby Heinecke, Lichao Sun, and Ran Xu. Tackling data heterogeneity in federated learning with class prototypes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7314– 7322, 2023. 2, 3

    work page 2023

  5. [5]

    Hyperspherical variational auto-encoders

    Tim R Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M Tomczak. Hyperspherical variational auto-encoders. arXiv preprint arXiv:1804.00891, 2018. 2

    arXiv 2018

  6. [6]

    Arcface: Additive angular margin loss for deep face recog- nition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recog- nition. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4690–4699, 2019. 2

    work page 2019

  7. [7]

    Heterofl: Compu- tation and communication efficient federated learning for het- erogeneous clients

    Enmao Diao, Jie Ding, and Vahid Tarokh. Heterofl: Compu- tation and communication efficient federated learning for het- erogeneous clients. arXiv preprint arXiv:2010.01264, 2020. 1

    arXiv 2010

  8. [8]

    Fedhp: Federated learning with hyperspherical prototypical regular- ization

    Samuele Fonio, Mirko Polato, Roberto Esposito, et al. Fedhp: Federated learning with hyperspherical prototypical regular- ization. In ESANN 2024 proceedings, pages 69–74. i6doc,

    work page 2024

Show all 46 references
  1. [9]

    Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout

    Samuel Horvath, Stefanos Laskaridis, Mario Almeida, Ilias Leontiadis, Stylianos Venieris, and Nicholas Lane. Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout. Advances in Neural Information Processing Systems, 34:12876–12889, 2021. 1

    2021

  2. [10]

    Personalized cross-silo federated learning on non-iid data

    Yutao Huang, Lingyang Chu, Zirui Zhou, Lanjun Wang, Jiangchuan Liu, Jian Pei, and Yong Zhang. Personalized cross-silo federated learning on non-iid data. In Proceed- ings of the AAAI conference on artificial intelligence, pages 7865–7873, 2021. 6

    2021

  3. [11]

    Communication- efficient on-device machine learning: Federated distillation and augmentation under non-iid private data

    Eunjeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication- efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. arXiv preprint arXiv:1811.11479, 2018. 1, 2

    2018 arXiv

  4. [12]

    Balanced open set domain adaptation via centroid alignment

    Mengmeng Jing, Jingjing Li, Lei Zhu, Zhengming Ding, Ke Lu, and Yang Yang. Balanced open set domain adaptation via centroid alignment. In Proceedings of the AAAI conference on artificial intelligence, pages 8013–8020, 2021. 2

    2021

  5. [13]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5

    2009

  6. [14]

    Tiny imagenet visual recognition challenge

    Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5

    2015

  7. [15]

    Fedmd: Heterogenous fed- erated learning via model distillation

    Daliang Li and Junpu Wang. Fedmd: Heterogenous fed- erated learning via model distillation. arXiv preprint arXiv:1910.03581, 2019. 1

    1910 arXiv

  8. [16]

    Feder- ated learning on non-iid data silos: An experimental study

    Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Feder- ated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965–978. IEEE, 2022. 1

    2022

  9. [17]

    Think locally, act globally: Fed- erated learning with local and global representations

    Paul Pu Liang, Terrance Liu, Liu Ziyin, Nicholas B Allen, Randy P Auerbach, David Brent, Ruslan Salakhutdinov, and Louis-Philippe Morency. Think locally, act globally: Fed- erated learning with local and global representations. arXiv preprint arXiv:2001.01523, 2020. 2

    2001 arXiv

  10. [18]

    Regularizing neural networks via minimizing hyperspherical energy

    Rongmei Lin, Weiyang Liu, Zhen Liu, Chen Feng, Zhiding Yu, James M Rehg, Li Xiong, and Le Song. Regularizing neural networks via minimizing hyperspherical energy. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 6917–6927, 2020. 2

    2020

  11. [19]

    Ensemble distillation for robust model fusion in federated learning

    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. 6

    2020

  12. [20]

    Large-margin softmax loss for convolutional neural networks

    Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. Large-margin softmax loss for convolutional neural networks. arXiv preprint arXiv:1612.02295, 2016. 2

    2016 arXiv

  13. [21]

    Deep hyperspherical learning

    Weiyang Liu, Yan-Ming Zhang, Xingguo Li, Zhiding Yu, Bo Dai, Tuo Zhao, and Le Song. Deep hyperspherical learning. Advances in neural information processing systems, 30, 2017

    2017

  14. [22]

    Learning towards minimum hyper- spherical energy

    Weiyang Liu, Rongmei Lin, Zhen Liu, Lixin Liu, Zhiding Yu, Bo Dai, and Le Song. Learning towards minimum hyper- spherical energy. Advances in neural information processing systems, 31, 2018. 4

    2018

  15. [23]

    Orthogonal over-parameterized training

    Weiyang Liu, Rongmei Lin, Zhen Liu, James M Rehg, Liam Paull, Li Xiong, Le Song, and Adrian Weller. Orthogonal over-parameterized training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7251–7260, 2021

    2021

  16. [24]

    Learning with hyperspherical uniformity

    Weiyang Liu, Rongmei Lin, Zhen Liu, Li Xiong, Bernhard Sch¨olkopf, and Adrian Weller. Learning with hyperspherical uniformity. In International conference on artificial intelli- gence and statistics, pages 1180–1188. PMLR, 2021. 2

    2021

  17. [25]

    Shufflenet v2: Practical guidelines for efficient cnn architec- ture design

    Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architec- ture design. In Proceedings of the European conference on computer vision (ECCV), pages 116–131, 2018. 6

    2018

  18. [26]

    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. In Artificial intelligence and statistics, pages 1273–1282. PMLR,

  19. [27]

    Hyper- spherical prototype networks

    Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyper- spherical prototype networks. Advances in neural information processing systems, 32, 2019. 2

    2019

  20. [28]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722–729. IEEE, 2008. 5

    2008

  21. [29]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510–4520, 2018. 6

    2018

  22. [30]

    Feder- ated mutual learning

    Tao Shen, Jie Zhang, Xinkang Jia, Fengda Zhang, Gang Huang, Pan Zhou, Kun Kuang, Fei Wu, and Chao Wu. Feder- ated mutual learning. arXiv preprint arXiv:2006.16765, 2020. 1, 2

    2006 arXiv

  23. [31]

    Feduv: Uniformity and variance for heterogeneous federated learning

    Ha Min Son, Moon-Hyun Kim, Tai-Myoung Chung, Chao Huang, and Xin Liu. Feduv: Uniformity and variance for heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5863–5872, 2024. 2

    2024

  24. [32]

    Personalized federated learning with moreau envelopes

    Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neu- ral Information Processing Systems, 33:21394–21405, 2020. 1, 6

    2020

  25. [33]

    Hyperspherical consistency regularization

    Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7244–7255, 2022. 2

    2022

  26. [34]

    Efficientnet: Rethinking model scal- ing for convolutional neural networks

    Mingxing Tan. Efficientnet: Rethinking model scal- ing for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 6

    1905 arXiv

  27. [35]

    Fedproto: Federated proto- type learning across heterogeneous clients

    Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated proto- type learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432– 8440, 2022. 1, 2, 3

    2022

  28. [36]

    Cosface: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 5265–5274, 2018. 2

    2018

  29. [37]

    Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation

    Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1426–1435, 2019. 2

    2019

  30. [38]

    Fedgh: Heterogeneous federated learning with general- ized global header

    Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with general- ized global header. In Proceedings of the 31st ACM Interna- tional Conference on Multimedia, pages 8686–8696, 2023. 8

    2023

  31. [39]

    Con- trolling update distance and enhancing fair trainable proto- types in federated learning under data and model heterogene- ity

    Kangning Yin, Zhen Ding, Xinhui Ji, and Zhiguo Wang. Con- trolling update distance and enhancing fair trainable proto- types in federated learning under data and model heterogene- ity. Defence Technology, 2025. 1, 2

    2025

  32. [40]

    A survey on federated learning

    Chen Zhang, Yu Xie, Hang Bai, Bin Yu, Weihong Li, and Yuan Gao. A survey on federated learning. Knowledge-Based Systems, 216:106775, 2021. 1

    2021

  33. [41]

    Fedala: Adaptive local aggregation for personalized federated learning

    Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Fedala: Adaptive local aggregation for personalized federated learning. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 11237–11244, 2023. 6

    2023

  34. [42]

    Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in fed- erated learning

    Jianqing Zhang, Yang Liu, Yang Hua, and Jian Cao. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in fed- erated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16768–1677...

    2024

  35. [43]

    Deep mutual learning

    Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, 2018. 2

    2018

  36. [44]

    Deep residual networks for hyperspectral image classi- fication

    Zilong Zhong, Jonathan Li, Lingfei Ma, Han Jiang, and He Zhao. Deep residual networks for hyperspectral image classi- fication. In 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pages 1824–1827. IEEE, 2017. 6

    2017

  37. [45]

    Data-free knowledge distillation for heterogeneous federated learn- ing

    Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learn- ing. In International conference on machine learning, pages 12878–12889. PMLR, 2021. 1, 2

    2021

  38. [46]

    Resilient and communication efficient learning for heteroge- neous federated systems

    Zhuangdi Zhu, Junyuan Hong, Steve Drew, and Jiayu Zhou. Resilient and communication efficient learning for heteroge- neous federated systems. Proceedings of machine learning research, 162:27504, 2022. 1 Heterogeneous Federated Learning with Prototype Alignment and Upscaling Su...

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.