arxiv: 2605.07888 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.CV

Recognition: no theorem link

Enhancing Federated Quadruplet Learning: Stochastic Client Selection and Embedding Stability Analysis

Ozgu Goksu , Nicolas Pugeault

Authors on Pith no claims yet

Pith reviewed 2026-05-11 03:29 UTC · model grok-4.3

classification 💻 cs.LG cs.CV

keywords federated learningquadruplet lossmetric learningdata heterogeneityrepresentation alignmentnon-IID dataimage classification

0 comments

The pith

FedQuad applies quadruplet loss in federated learning to minimize intra-class distances and maximize inter-class distances across clients, reducing misalignment during aggregation.

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

The paper tries to establish that a quadruplet loss can be optimized locally on each client to pull same-class representations closer together while pushing different-class ones farther apart, even when data distributions vary widely across devices. A sympathetic reader would care because standard federated averaging often blurs class boundaries under limited local samples and class imbalance, hurting the final global model's accuracy. The approach works by focusing on positive and negative pair distances during local training rounds, then aggregating models without ever centralizing raw data or requiring global class counts. Experiments across CIFAR-10, CIFAR-100 and Tiny-ImageNet under multiple non-IID partitions and client counts demonstrate gains over prior federated baselines, while also comparing metric learning behavior in single-machine versus distributed settings.

Core claim

The central claim is that jointly minimizing distances between positive pairs and maximizing distances between negative pairs through a quadruplet loss, together with stochastic client selection, mitigates representation misalignment introduced by model aggregation and improves generalization under data heterogeneity.

What carries the argument

The quadruplet loss that enforces smaller distances for same-class sample pairs and larger distances for different-class sample pairs, computed locally on each client's data before aggregation.

If this is right

The global model reaches higher test accuracy on non-IID image datasets than existing federated methods.
Intra-class compactness increases and inter-class separation improves after aggregation steps.
Representation collapse occurs less frequently in both centralized metric learning and federated training.
Stochastic client selection helps maintain embedding stability across training rounds.

Where Pith is reading between the lines

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

The loss structure might transfer to other federated tasks such as detection where local class imbalance is severe.
Embedding stability measurements could guide the choice of aggregation weights in broader federated optimization.
Varying the number of participating clients in new experiments would test how the gains scale with participation rate.

Load-bearing premise

Quadruplet loss terms can be stably optimized on local client data under stochastic participation and heterogeneous distributions without new instabilities or any need for unavailable global statistics.

What would settle it

If accuracy on CIFAR-10 under strong non-IID partitions shows no improvement or drops below standard federated baselines when using the quadruplet loss, the performance benefit claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.07888 by Nicolas Pugeault, Ozgu Goksu.

**Figure 2.** Figure 2: Overview of the FedQuad local training framework. Each client minimises loss com [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Inter/Intra-class ratio (↑ better) across federated learning methods on CIFAR-10 (200 clients). Error bars indicate standard deviation over runs. As an alternative, we evaluate these methods with stochastic client selection. The random client selection scenario highlights substantial variation in data distributions across training rounds. In this setting, a subset of clients is randomly selected at each c… view at source ↗

**Figure 4.** Figure 4: Inter/Intra-class ratio (↑ better) across federated learning methods on CIFAR-100 (200 clients). Error bars indicate standard deviation over runs. FedAvg SupConFL TripletFL QuadrupletFL MOON FedQuad 0 1 2 Methods IID Intra-Class Distance Inter/Intra Class Ratio 0 0.2 0.4 0.6 0.8 Intra FedAvg SupConFL TripletFL QuadrupletFL MOON FedQuad 0 1 2 Methods α = 0.5 Intra-Class Distance Inter/Intra Class Ratio 0 0.… view at source ↗

**Figure 5.** Figure 5: Inter/Intra-class ratio (↑ better) across federated learning methods on Tiny-ImageNet (200 clients). Error bars indicate standard deviation over runs. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

read the original abstract

Federated Learning (FL) enables decentralised model training across distributed clients without requiring data centralisation. However, the generalisation performance of the global model is usually degraded by data heterogeneity across clients, particularly under limited data availability and class imbalance. To address this challenge, we propose FedQuad, a novel method that explicitly enforces minimising intra-class representations while enabling inter-class splits across clients. By jointly minimising distances between positive pairs and maximising distances between negative pairs, the proposed approach mitigates representation misalignment introduced during model aggregation. We evaluate our method on CIFAR-10, CIFAR-100, and Tiny-ImageNet under diverse non-IID settings and varying numbers of clients, demonstrating consistent improvements over existing baselines. Additionally, we provide a comprehensive analysis of metric learning-based approaches in both centralised and federated environments, highlighting their effectiveness in alleviating representation collapse under heterogeneous data distributions.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

FedQuad adds quadruplet loss and stochastic client selection to federated learning to fight representation collapse, but the non-IID setup raises questions about where negative pairs come from on single-class clients.

read the letter

The main takeaway is that this paper takes quadruplet metric learning, which pulls same-class embeddings together and pushes different-class ones apart, and applies it inside federated averaging with stochastic client selection plus some stability checks on the embeddings. The goal is to reduce the misalignment that happens when you average models trained on very different local data distributions. They test on CIFAR-10, CIFAR-100, and Tiny-ImageNet under several non-IID partitions and client counts, and report better results than standard baselines. They also give a side-by-side look at how metric-learning losses behave in centralised versus federated training, which is useful for seeing where the collapse problems actually show up. That analysis and the joint formulation with stochastic selection are the clearest new pieces. The experiments appear to cover a reasonable range of heterogeneity levels, which is better than many FL papers that only show one or two splits. The stress-test point about negative samples is worth checking in the full text. Under the pathological or Dirichlet partitions they use, a fair number of clients will hold examples from only one or two classes. A valid quadruplet needs two negatives from other classes. If the method stays strictly local and avoids any global statistics or memory banks, those clients cannot compute the inter-class term, so the claimed mitigation of collapse would only work on multi-class clients. The abstract does not mention how this is handled, so the paper needs to show either that single-class clients are rare in their setups or that they have a concrete local workaround that still delivers the inter-class push. Without that, the central claim rests on an assumption that may not hold in the regimes they evaluate. This is the kind of incremental but practical extension that people building privacy-sensitive FL systems might try. It is not foundational, but the problem it targets is real and the experiments are on standard benchmarks. I would send it to a referee who knows both metric learning and federated optimisation; the sampling detail and any ablation on the quadruplet components are the obvious things to probe. Overall it is coherent enough to review rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The paper proposes FedQuad, a federated learning method that applies quadruplet loss to jointly minimize intra-class distances and maximize inter-class distances in client embeddings, aiming to counteract representation misalignment from model aggregation under data heterogeneity. It incorporates stochastic client selection and provides an embedding stability analysis, with evaluations on CIFAR-10, CIFAR-100, and Tiny-ImageNet under diverse non-IID partitions and varying client numbers, claiming consistent gains over baselines along with a broader analysis of metric learning approaches in centralized and federated settings.

Significance. If the empirical results and handling of extreme non-IID cases hold, this could provide a useful extension of metric learning techniques to federated settings for mitigating representation collapse, with the stochastic selection and stability components offering practical benefits for variable client participation. The work addresses a relevant challenge in FL generalization but would benefit from stronger quantitative grounding to establish its impact relative to existing methods.

major comments (2)

[§3] §3 (Method description of quadruplet loss): The formulation assumes clients can form valid quadruplets (anchor-positive-negative1-negative2) to enforce inter-class separation, but the evaluated non-IID regimes (pathological and Dirichlet partitions on CIFAR-10/100 and Tiny-ImageNet) commonly result in clients holding data from only one or two classes. No mechanism is specified for obtaining cross-class negatives locally without global statistics, memory banks, or cross-client sampling, which directly undermines the central claim that the loss mitigates aggregation-induced misalignment in a fully decentralized manner.
[§4] §4 (Experiments): The abstract and visible text assert 'consistent improvements' and 'comprehensive analysis' but provide no tables, numerical accuracy values, error bars, ablation results on loss components, or statistical tests. Without these, the magnitude of gains over baselines (e.g., FedAvg or other metric FL methods) and the specific contribution of the quadruplet term versus stochastic selection cannot be assessed, making the empirical support for the claims unverifiable.

minor comments (2)

[§3] The notation for the quadruplet loss and stability metric should be formalized with explicit equations early in §3 to improve readability and allow direct comparison to standard quadruplet loss definitions.
[§4] Figure captions and axis labels in the stability analysis plots could be expanded to explicitly state the non-IID partition type and client participation rate for each curve.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which has helped us identify areas for clarification and strengthening of the manuscript. We address each major comment below and outline the revisions we will make.

read point-by-point responses

Referee: [§3] §3 (Method description of quadruplet loss): The formulation assumes clients can form valid quadruplets (anchor-positive-negative1-negative2) to enforce inter-class separation, but the evaluated non-IID regimes (pathological and Dirichlet partitions on CIFAR-10/100 and Tiny-ImageNet) commonly result in clients holding data from only one or two classes. No mechanism is specified for obtaining cross-class negatives locally without global statistics, memory banks, or cross-client sampling, which directly undermines the central claim that the loss mitigates aggregation-induced misalignment in a fully decentralized manner.

Authors: We acknowledge the validity of this concern for extreme non-IID partitions where many clients may hold data from only one or two classes, limiting local quadruplet formation. Our stochastic client selection is explicitly intended to address representation misalignment by prioritizing diverse client participation across rounds, thereby allowing inter-class separation to emerge through aggregation rather than solely local computation. Locally, the quadruplet loss is applied to available same-class positives for intra-class minimization, with the stability analysis quantifying the effect of partial negatives. However, we agree the manuscript lacks an explicit description of this handling. We will revise §3 to add a clear explanation, including conditions for quadruplet construction based on local class availability and how stochastic selection ensures complementary distributions over time, while preserving the fully decentralized protocol (no extra global statistics or cross-client sampling beyond standard model aggregation). revision: yes
Referee: [§4] §4 (Experiments): The abstract and visible text assert 'consistent improvements' and 'comprehensive analysis' but provide no tables, numerical accuracy values, error bars, ablation results on loss components, or statistical tests. Without these, the magnitude of gains over baselines (e.g., FedAvg or other metric FL methods) and the specific contribution of the quadruplet term versus stochastic selection cannot be assessed, making the empirical support for the claims unverifiable.

Authors: The referee is correct that the current experimental presentation relies on qualitative claims without sufficient quantitative backing in the reviewed version. To enable proper assessment of gains and component contributions, we will expand §4 with full tables of test accuracies (including means and standard deviations over multiple random seeds), error bars on all plots, dedicated ablation studies separating the quadruplet loss from stochastic selection, and statistical tests (e.g., Wilcoxon signed-rank or paired t-tests) against baselines such as FedAvg and other metric-learning FL methods. These additions will cover all reported datasets and non-IID settings. revision: yes

Circularity Check

0 steps flagged

No circularity: method proposal supported by external benchmark evaluations

full rationale

The paper proposes FedQuad, a quadruplet-loss-based approach for federated learning under non-IID conditions, and supports its claims through empirical evaluation on standard external datasets (CIFAR-10, CIFAR-100, Tiny-ImageNet) with diverse partitions. No equations, derivations, or load-bearing steps are shown that reduce claimed improvements to quantities defined by the method itself, fitted parameters renamed as predictions, or self-citation chains. The central premise relies on the quadruplet loss formulation and stochastic client selection, which are presented as novel contributions evaluated independently rather than tautologically derived from inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are identifiable from the given description.

pith-pipeline@v0.9.0 · 5450 in / 1122 out tokens · 48604 ms · 2026-05-11T03:29:03.511402+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

[1]

Laion-5b: An open large-scale dataset for train- ing next generation image-text models,

C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al., “Laion-5b: An open large-scale dataset for train- ing next generation image-text models,” Advances in Neural Information Processing Sys- tems, vol. 35, pp. 25278–25294, 2022

work page 2022
[2]

ImageNet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2009, pp. 248–255. 18

work page 2009
[3]

Enhancing generalisation in federated learning with heterogeneous data: A comparative literature review,

A. Mora, A. Bujari, and P. Bellavista, “Enhancing generalisation in federated learning with heterogeneous data: A comparative literature review,”Future Generation Computer Systems, Elsevier, 2024

work page 2024
[4]

Federated learning on non-IID data: A survey,

H. Zhu, J. Xu, S. Liu, and Y . Jin, “Federated learning on non-IID data: A survey,”Neuro- computing, vol. 465, pp. 371–390, 2021

work page 2021
[5]

Communication- efficient learning of deep networks from decentralised data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication- efficient learning of deep networks from decentralised data,” inArtificial Intelligence and Statistics, PMLR, 2017, pp. 1273–1282

work page 2017
[6]

Federated ma- chine learning in healthcare: A systematic review on clinical applications and technical architecture

T. Zhen Ling, L. Jin, N., Siqi Li, D. Miao, X. Zhang, W. Yan Ng et al. "Federated ma- chine learning in healthcare: A systematic review on clinical applications and technical architecture." inCell Reports Medicine, 2024, vol. 2

work page 2024
[7]

Fedrsclip: Federated learning for remote sensing scene classification using vision-language models

L. Hui, Z. Chao, H. Danfeng, D. Kexin, and W. Congcong, " Fedrsclip: Federated learning for remote sensing scene classification using vision-language models", inIEEE Geoscience and Remote Sensing Magazine, 2025

work page 2025
[8]

Rethinking data heterogene- ity in federated learning: Introducing a new notion and standard benchmarks

S. Vahidian, M. Morafah, C. Chen, M. Shah, and B. Lin, "Rethinking data heterogene- ity in federated learning: Introducing a new notion and standard benchmarks." inIEEE Transactions on Artificial Intelligence, vol. 5(3), pp. 1386-1397, 2023

work page 2023
[9]

A new theoretical perspective on data hetero- geneity in federated optimization

J. Wang, S. Wang, R.R. Chen, and M. Ji, "A new theoretical perspective on data hetero- geneity in federated optimization." inarXiv preprint arXiv:2407.15567, 2024

work page arXiv 2024
[10]

Tackling data heterogeneity in federated learning through knowledge distillation with inequitable aggregation

X. Ma, S. Zhu, C. Qiu, and G. Sun, "Tackling data heterogeneity in federated learning through knowledge distillation with inequitable aggregation." inEngineering Applications of Artificial Intelligence, vol. 173, p. 114418, 2026

work page 2026
[11]

Federated minimax optimisation: Im- proved convergence analyses and algorithms,

P. Sharma, R. Panda, G. Joshi, and P. Varshney, “Federated minimax optimisation: Im- proved convergence analyses and algorithms,” inProc. Int. Conf. Mach. Learn., PMLR, 2022, pp. 19683–19730

work page 2022
[12]

and Stich, Sebastian U

S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. U. Stich, and A. T. Suresh, “Scaf- fold: Stochastic controlled averaging for on-device federated learning,”arXiv preprint arXiv:1910.06378, vol. 2, no. 6, 2019

work page arXiv 1910
[13]

Fedproc: Prototypical contrastive federated learning on non-iid data

X. Mu, Y . Shen, K. Cheng, X. Geng, J. Fu, T. Zhang, and Z. Zhang, "Fedproc: Prototypical contrastive federated learning on non-iid data." inFuture Generation Computer Systems, vol. 143, pp. 93-104, 2023

work page 2023
[14]

Model-contrastive federated learning,

Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 10713–10722. 19

work page 2021
[15]

Relaxed contrastive learning for federated learning,

S. Seo, J. Kim, G. Kim, and B. Han, “Relaxed contrastive learning for federated learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 12279–12288

work page 2024
[16]

Learning multiple layers of features from tiny images,

A. Krizhevsky, “Learning multiple layers of features from tiny images,” Master’s thesis, University of Toronto, 2009

work page 2009
[18]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” inAdvances in Neural Information Process- ing Systems, vol. 33, 2020, pp. 18661–18673

work page 2020
[19]

Deep metric learning using triplet network,

E. Hoffer and N. Ailon, “Deep metric learning using triplet network,” inSimilarity-Based Pattern Recognition: Proc. 3rd Int. Workshop, SIMBAD, Copenhagen, Denmark, Oct. 2015, pp. 84–92. Springer, 2015

work page 2015
[20]

Beyond triplet loss: A deep quadruplet network for person re-identification,

W. Chen, X. Chen, J. Zhang, and K. Huang, “Beyond triplet loss: A deep quadruplet network for person re-identification,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 403–412

work page 2017
[21]

Hard negative examples are hard, but useful,

H. Xuan, A. Stylianou, X. Liu, and R. Pless, “Hard negative examples are hard, but useful,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 126–142

work page 2020
[22]

Federated learning with metric loss,

H. Park, H. Hosseini, and S. Yun, “Federated learning with metric loss,” inProc. Workshop on Federated Learning for User Privacy and Data Confidentiality in ICML, 2021

work page 2021
[23]

Privacy- preserving and robust federated deep metric learning,

Y . Tian, X. Ke, Z. Tao, S. Ding, F. Xu, Q. Li, H. Han, S. Zhong, and X. Fu, “Privacy- preserving and robust federated deep metric learning,” inProc. IEEE/ACM 30th Int. Symp. on Quality of Service (IWQoS), 2022, pp. 1–11

work page 2022
[24]

Defending against adversarial attacks in federated learning on metric learning model,

Z. Gu, J. Shi, Y . Yang, and L. He, “Defending against adversarial attacks in federated learning on metric learning model,” inProc. IEEE 22nd Int. Conf. on Trust, Security and Privacy in Computing and Communications (TrustCom), 2023, pp. 197–206

work page 2023
[25]

Privacy preserving palmprint recognition via fed- erated metric learning,

H. Shao, C. Liu, X. Li, and D. Zhong, “Privacy preserving palmprint recognition via fed- erated metric learning,”IEEE Trans. Inf. Forensics Security, vol. 19, pp. 878–891, 2023

work page 2023
[26]

Feddc: Federated learning with non-iid data via local drift decoupling and correction,

L. Gao, H. Fu, L. Li, Y . Chen, M. Xu, and C.-Z. Xu, “Feddc: Federated learning with non-iid data via local drift decoupling and correction,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 10112–10121. 20

work page 2022
[27]

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

D. Acar, Y . Zhao, R. Navarro, M. Mattina, P. Whatmough, and V . Saligrama, “Federated learning based on dynamic regularisation,”arXiv preprint arXiv:2111.04263, 2021

work page arXiv 2021
[28]

Robust Asymmetric Heterogeneous Federated Learning with Corrupted Clients,

X. Fang, M. Ye, and B. Du, “Robust Asymmetric Heterogeneous Federated Learning with Corrupted Clients,”IEEE Trans. Pattern Anal. Mach. Intell., 2025

work page 2025
[29]

Fedproc: Prototypical contrastive federated learning on non-iid data,

X. Mu, Y . Shen, K. Cheng, X. Geng, J. Fu, T. Zhang, and Z. Zhang, “Fedproc: Prototypical contrastive federated learning on non-iid data,”Future Generation Computer Systems, vol. 143, pp. 93–104, 2023

work page 2023
[30]

FedCRL: Personalized federated learn- ing with contrastive shared representations for label heterogeneity in non-IID data,

C. Huang, X. Chen, Y . Zhang, and H. Wang, “FedCRL: Personalized federated learn- ing with contrastive shared representations for label heterogeneity in non-IID data,”arXiv preprint arXiv:2404.17916, 2024

work page arXiv 2024
[31]

Federated learning from pre-trained models: A contrastive learning approach,

Y . Tan, G. Long, J. Ma, L. Liu, T. Zhou, and J. Jiang, “Federated learning from pre-trained models: A contrastive learning approach,”Adv. Neural Inf. Process. Syst., vol. 35, pp. 19332–19344, 2022

work page 2022
[32]

FedTrip: A resource-efficient federated learning method with triplet regularization,

X. Li, M. Liu, S. Sun, Y . Wang, H. Jiang, and X. Jiang, “FedTrip: A resource-efficient federated learning method with triplet regularization,” inProc. IEEE Int. Parallel and Dis- tributed Processing Symp. (IPDPS), 2023, pp. 809–819

work page 2023
[33]

Rethinking the representation in federated unsupervised learning with non-iid data

X. Liao, W. Liu, C. Chen, P. Zhou, F. Yu, H. Zhu, B. Yao, T. Wang, X. Zheng, and Y . Tan, "Rethinking the representation in federated unsupervised learning with non-iid data." in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 22841-22850, 2024

work page 2024
[34]

Orchestra: Unsupervised federated learning via globally consistent clustering

E.S. Lubana, C.I. Tang, F. Kawsar, R.P. Dick, and A. Mathur, "Orchestra: Unsupervised federated learning via globally consistent clustering", inarXiv preprint arXiv:2205.11506, 2022

work page arXiv 2022
[35]

A mutual information perspective on feder- ated contrastive learning,

C. Louizos, M. Reisser, and D. Korzhenkov, “A mutual information perspective on feder- ated contrastive learning,”arXiv preprint arXiv:2405.02081, 2024

work page arXiv 2024
[36]

Federated contrastive learning for decentralized unlabeled medical images,

N. Dong and I. V oiculescu, “Federated contrastive learning for decentralized unlabeled medical images,” inProc. Int. Conf. on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2021, pp. 378–387

work page 2021
[37]

Local learning matters: Rethinking data heterogeneity in federated learning,

M. Mendieta, T. Yang, P. Wang, M. Lee, Z. Ding, and C. Chen, “Local learning matters: Rethinking data heterogeneity in federated learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 8397–8406. 21

work page 2022
[38]

Tackling data heterogeneity in federated learning with class prototypes,

Y . Dai, Z. Chen, J. Li, S. Heinecke, L. Sun, and R. Xu, “Tackling data heterogeneity in federated learning with class prototypes,” inProc. AAAI Conf. Artif. Intell., vol. 37, 2023, pp. 7314–7322

work page 2023
[39]

FedQuad: Federated Stochastic Quadruplet Learning to Mit- igate Data Heterogeneity

Ö. Göksu, and N. Pugeault, "FedQuad: Federated Stochastic Quadruplet Learning to Mit- igate Data Heterogeneity". In3rd International Conference on Federated Learning Tech- nologies and Applications (FLTA), 2025, pp. 262-269

work page 2025
[40]

Vision transformers in 2022: An update on tiny imagenet

E. Huynh, "Vision transformers in 2022: An update on tiny imagenet", inarXiv preprint arXiv:2205.10660, 2022. 22

work page arXiv 2022