Pith. sign in

REVIEW 4 major objections 6 minor 30 references

PFedDST: Personalized Federated Learning with Decentralized Selection Training

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read PFedDST claims that a composite peer-selection score—combining the loss of a client's model on a peer's data, cosine similarity between header layers, and a recency term—plus partial-freeze local training improves personalized accuracy…

desk verdict PFedDST combines known ideas into a new decentralized peer-selection score, but the paper's theory is internally inconsistent and its CIFAR-100 result contradicts the abstract; needs major revision. read the letter →

arxiv 2502.07750 v2 pith:FWU2DX3U submitted 2025-02-11 cs.LG cs.AI

Mengchen Fan , Keren Li , Tianyun Zhang , Qing Tian , Baocheng Geng This is my paper · ORCID
classification cs.LGcs.AI
keywords personalizedfederatedlearningdecentralizedpeerselectiondataheterogeneitypartialfreezingcommunicationefficiencyclientconvergenceacceleration
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 introduces a decentralized personalized federated learning algorithm in which each client scores every potential peer using three locally computable signals: how poorly the client's current model performs on that peer's data, how similar the peers' classification tasks are (measured by cosine distance between header-layer weights), and how long it has been since the last communication with that peer. Clients then aggregate only the feature-extraction layers with the top-scoring peers, keep their classification header strictly personalized, and train in alternating phases with one part of the model frozen. On CIFAR-10 and CIFAR-100 with pathological non-IID splits over 100 clients, the authors report that PFedDST reaches target accuracy in fewer communication rounds than the compared baselines and achieves the best CIFAR-10 accuracy at 92.25%. The intended payoff is that guided peer selection combined with partial freezing yields both faster convergence and better personalization under data heterogeneity.

What carries the argument

The composite communication score of Eq. (9), $S = s_p(\alpha s_l - s_d + c)$, where $s_l$ is the loss of client $i$'s model evaluated on peer $j$'s data (a surrogate for the selection-skew quantity $\rho$), $s_d$ is the cosine similarity between the header weight vectors of two clients (task similarity), and $s_p$ is an exponential-CDF recency term that encourages engaging peers not recently contacted. This score ranks peers; the top-scoring peers form the set $\mathcal{M}_i$ whose feature-extraction layers are averaged with the client's own, and the two-phase training (header frozen while updating $e_i$, then feature extractor frozen while updating $h_i$) carries the personalization and reduces per-round computation.

What would settle it

Run PFedDST with the composite score replaced by random peer selection while keeping the same partial-freeze training and the same number of selected peers on CIFAR-10 with 100 clients; if the random-selection version reaches 90% accuracy in the same or fewer communication rounds as the scored version, then the score itself is not the driver of the reported acceleration.

Watch

Extended reading notes

Core claim

The central claim is that a decentralized, locally computed peer-selection score, $S = s_p(\alpha s_l - s_d + c)$ from Eq. (9), identifies peers whose feature-extraction layers improve a client's local accuracy, and that alternating partial-freeze training of the shared feature extractor and the personalized header lets each client converge faster and to higher personalized accuracy than random-selection or full-parameter-exchange baselines. The score combines the loss disparity $s_l = \lVert \mathbb{E}_{(x,y)\sim D_j}[\mathcal{L}(w_i; x,y)] \rVert$, the header cosine similarity $s_d$, and the peer recency term $s_p = \phi(1 - e^{-\lambda(n_t - n_{0,j})})$. The paper frames this as extending the selection-skew convergence analysis of centralized federated learning to a decentralized personalized setting.

Load-bearing premise

The claim that the score accelerates convergence rests on the untested assumption that the loss of a client's model on a peer's data captures the selection-skew advantage proven for centralized federated learning, and that this advantage carries over to decentralized partially frozen training.

Editorial extensions

If this is right

  • If the claim holds, decentralized PFL systems can cut communication rounds to reach a target accuracy by roughly 25% compared with the strongest directed-collaboration baseline (DFedPGP) in the reported settings.
  • Partial freezing lowers the per-round update cost: the feature extractor trains for 5 epochs while the personal header trains for 1, reducing computation while preserving personalization.
  • The header-distance term can act as a filter that excludes clients with noisy labels or adversarial models, which the discussion section presents as a robustness benefit.
  • Because the score is fully local and requires no global iteration count, the method scales naturally to peer-to-peer deployments without a central server.

Reading between the lines

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

  • The paper does not ablate the three score components, so the marginal contribution of loss disparity versus task similarity versus recency is untested; a component-wise study would be needed to know which term drives the gain.
  • The convergence motivation comes from a centralized, non-convex client-selection result, and the paper does not prove that this transfers to the decentralized partially frozen setting; a dedicated convergence proof for PFedDST would close that gap.
  • The recency term's exponential CDF is one of many possible scheduling mechanisms; bandit-style exploration or graph-walk schedules could be substituted into the same composite score without changing the rest of the framework.
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 / 6 minor

Summary. PFedDST proposes a decentralized personalized federated learning framework in which each client scores candidate peers by a composite score combining loss disparity, header distance, and peer recency, then aggregates only the feature-extraction layers with the selected peers while keeping a personalized header. The paper claims that this selection strategy accelerates convergence and improves personalized test accuracy over centralized and decentralized baselines on CIFAR-10 and CIFAR-100. The contribution is presented as a practical framework with heuristic selection scores, partial-freeze training, and experimental comparisons; the paper also invokes a theoretical convergence result from Cho et al. [22] to motivate the selection rule.

Significance. If the claimed acceleration and accuracy gains were rigorously established, PFedDST would be a useful addition to the decentralized personalized federated learning literature, where peer-selection rules are often heuristic and lack convergence guarantees. The paper's partial-freeze training and composite peer scoring are reasonable engineering ideas, and the experimental comparison spans several relevant baselines. However, the manuscript's central claims are not supported by a valid theoretical argument and are partially contradicted by the reported experimental text. The significance of the work is currently limited by these correctness and reproducibility gaps.

major comments (4)
  1. [II.B, Eq. (5)] The definition of the decentralized selection skew ρ_i in Eq. (5) is internally inconsistent with the claim that random selection gives ρ=1. For random selection, the numerator is the M-average of L_j(w_i) − L_j(w*_j), while the denominator is L_i(w_i) minus the M-average of L_j(w*_j). These two quantities are equal only if the M-average of L_j(w_i) equals L_i(w_i), which is not generally true. The assertion "With purely random selection, ρ = 1" is therefore not a consequence of Eq. (5), and the subsequent motivation for score-based selection is not grounded in the stated quantity.
  2. [II.B and V] The paper claims theoretical support for accelerated convergence, but no convergence theorem is stated or proved. The conclusion says "Theoretical findings and experimental results show that our method achieved a faster convergence rate," yet Section II.B only cites Cho et al. [22] and asserts an analogy. Since the decentralized setting differs from the centralized client-selection setting of [22] (local personalized models, partial freezing, peer selection based on surrogate losses), the transfer of the selection-skew convergence result is not established. This leaves the load-bearing claim of accelerated convergence without a stated theoretical basis.
  3. [III.B] The experimental text contradicts the abstract's claim that PFedDST outperforms all state-of-the-art baselines. Section III.B states that on CIFAR-100, "DFedPGP leads with an accuracy of 79.41%, which is at least 0.7% higher than other baseline methods." This sentence, under its natural reading, says DFedPGP, not PFedDST, has the highest CIFAR-100 accuracy. The paper needs to clarify which method achieved 79.41% and, if PFedDST did not outperform on CIFAR-100, the claim in the abstract and Section I must be revised.
  4. [II.B, Eq. (6) and Eq. (9)] The surrogate loss in Eq. (6), l_j(w_{i,j}), is the loss of client i's model on peer j's data, but the theorem in Cho et al. [22] concerns the loss gap L_j(w) − L_j(w*_j) for a common global model w. No argument is given that selecting peers by this surrogate maximizes a lower bound on the correctly defined selection skew. Additionally, the final score in Eq. (9) contains unexplored free parameters α, c0, s*, and c; the paper provides no sensitivity analysis or ablation to show that the reported gains are robust to these choices rather than tuned to the experimental setup.
minor comments (6)
  1. [III.A] The experimental section does not report the number of random seeds, standard deviations, or variance across runs. Without these, the 1.0% margin on CIFAR-10 and the 0.7% gap on CIFAR-100 cannot be assessed for statistical significance.
  2. [III.A] Table I lists only some baselines with "-" entries for FedAvg and Dis-PFL, but the text does not explain why these methods fail to reach the target accuracy. A footnote or clarifying sentence would improve interpretability.
  3. [III.B] The text refers to "Table 2" for learning curves, but the manuscript includes only one table (Table I). The learning curves appear only in Figures 3 and 4. Please align the cross-references.
  4. [Algorithm 1] Line 15 of Algorithm 1 says "Update header parameters e_i" but the variable being updated is h_i. This appears to be a typo that could confuse readers.
  5. [II.B, Eq. (8)] The peer recency score in Eq. (8) applies a CDF to a CDF-like expression; the notation φ is not defined, and the composition of two CDFs is unusual. Please clarify the intended functional form.
  6. [References] Reference [28] (Freezeout) and reference [22] (Cho et al.) are cited for motivation, but the specific connection to partial freezing and selection skew is not elaborated in the text. Adding precise statements about which results are used would improve rigor.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; PFedDST's selection rule is a heuristic benchmarked externally, with a theoretical-transfer gap rather than a self-referential derivation.

full rationale

PFedDST's central claim is empirical: score-based peer selection plus partial freezing yields higher personalized accuracy and faster convergence on CIFAR-10/100 against external baselines (FedAvg, FedPer, FedBABU, DFedAvgM, Dis-PFL, DFedPGP). The selection score in Eq. (9) contains hyperparameters (alpha, c, lambda, threshold s*) that are not fitted to the reported test accuracies, and the reported curves and round-count table are independent benchmarks. The convergence argument is imported from Cho et al. [22], an external source, not from the authors' own prior work; the authors' self-citations ([2], [13]) appear only as related work and do not carry the argument. The paper does contain a derivation gap: Eq. (5)'s assertion that random selection gives rho = 1 does not follow from the displayed definition unless the average loss of client i's model over all peers equals Li(wi), and Eq. (6)'s surrogate is not shown to maximize the lower bound proved in [22] for a common global model. These are correctness and derivation concerns, not circular reductions: no equation or fitted parameter is re-labeled as a prediction, and the empirical results would stand or fall independently of the [22] analogy. Accordingly the circularity score is low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The algorithm depends on at least five unreported hand-set parameters, and the claimed convergence benefit rests on an unproved transfer of a centralized selection-skew result to a decentralized partial-freeze setting. The paper does not introduce new physical or mathematical entities.

free parameters (5)
  • alpha (loss scaling)
    Scaling parameter in Eq. (9) that balances loss disparity against header distance; no value is reported.
  • lambda (recency rate)
    Rate parameter in Eq. (8) controlling how quickly the peer recency score saturates; no value is reported.
  • c0 (recency saturation threshold)
    Threshold described in Section II.B beyond which the peer recency score reaches its maximum; no value is reported.
  • s* (selection threshold)
    Threshold in Algorithm 1 line 5 that determines which peers are selected; no value is reported.
  • c (communication cost score)
    Constant in Eq. (9); the experiments say communication cost is equal for all clients, but its value is not given.
assumptions (4)
  • ad hoc to paper The centralized selection-skew convergence benefit of Cho et al. [22] transfers to decentralized peer selection with partial freezing.
    Invoked in Section II.B to justify selecting high-loss peers; no proof is given for the decentralized partial-freeze setting.
  • ad hoc to paper The loss of client i's model on peer j's data is a valid surrogate for the selection skew rho_i.
    Section II.B says evaluating rho is NP-hard and uses l_j(w_i,j) as a surrogate; validity is asserted, not shown.
  • domain assumption Cosine similarity between header weights measures task similarity, so smaller header distance means learning from the peer is beneficial.
    Used in Eq. (7) and Section II.B; plausible in personalized federated learning but not formally justified.
  • domain assumption Aggregating feature extraction layers of selected peers while keeping personalized headers improves local accuracy.
    Standard assumption of partial personalization, inherited from FedPer and FedBABU; used throughout Section II.A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PFedDST: Personalized Federated Learning with Decentralized Selection Training." pith.science (2026). https://pith.science/paper/FWU2DX3U

@misc{pith2026250207750,
  author       = {Pith},
  title        = {Pith review of: PFedDST: Personalized Federated Learning with Decentralized Selection Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWU2DX3U}},
  note         = {Machine review of arXiv:2502.07750}
}
read the original abstract

Distributed Learning (DL) enables the training of machine learning models across multiple devices, yet it faces challenges like non-IID data distributions and device capability disparities, which can impede training efficiency. Communication bottlenecks further complicate traditional Federated Learning (FL) setups. To mitigate these issues, we introduce the Personalized Federated Learning with Decentralized Selection Training (PFedDST) framework. PFedDST enhances model training by allowing devices to strategically evaluate and select peers based on a comprehensive communication score. This score integrates loss, task similarity, and selection frequency, ensuring optimal peer connections. This selection strategy is tailored to increase local personalization and promote beneficial peer collaborations to strengthen the stability and efficiency of the training process. Our experiments demonstrate that PFedDST not only enhances model accuracy but also accelerates convergence. This approach outperforms state-of-the-art methods in handling data heterogeneity, delivering both faster and more effective training in diverse and decentralized systems.

Figures

Figures reproduced from arXiv: 2502.07750 by the authors.

Figure 1
Figure 1. Overview of PFedDST B. Strategic Peer Selection for Communication For each client, we quantify the degree of information contribution from others by assigning a score to each peer. A higher score indicates that a peer holds more valuable information for enhancing the feature extraction capabilities of the local client, thus facilitating more targeted updates. Essentially, our aim is for clients to augment their feat… view at source ↗
Figure 2
Figure 2. The validation result of each selected peer in the local [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Test accuracy on CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Test accuracy on CIFAR-100 IV. DISCUSSION Our framework is designed to prioritize the selection of the most beneficial communication peers and utilize partial personalization, ensuring optimal performance and efficiency in distributed learning scenarios. The simulation…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages

  1. [22]

    Client selec- tion in federated learning: Convergence analysis and power-of-choice selection strategies,

    Y . J. Cho, J. Wang, and G. Joshi, “Client selec- tion in federated learning: Convergence analysis and power-of-choice selection strategies,” arXiv preprint arXiv:2010.01243, 2020

    arXiv 2010

  2. [1]

    Fusion of deep neural networks for activity recognition: A regular vine copula based approach,

    S. Zhang, B. Geng, P. K. Varshney, and M. Rangaswamy, “Fusion of deep neural networks for activity recognition: A regular vine copula based approach,” in 2019 22th International Conference on Information Fusion (FU- SION). IEEE, 2019, pp. 1–7

    work page 2019

  3. [2]

    Interpretable data fusion for distributed learning: A representative approach via gradient matching,

    M. Fan, B. Geng, K. Li, X. Wang, and P. K. Varshney, “Interpretable data fusion for distributed learning: A representative approach via gradient matching,” in 2024 27th International Conference on Information Fusion (FUSION), 2024, pp. 1–8

    work page 2024

  4. [3]

    Nn- copula-cd: A copula-guided interpretable neural network for change detection in heterogeneous remote sensing images,

    W. Li, X. Wang, G. Li, B. Geng, and P. K. Varshney, “Nn- copula-cd: A copula-guided interpretable neural network for change detection in heterogeneous remote sensing images,” arXiv preprint arXiv:2303.17448 , 2023

    arXiv 2023

  5. [4]

    Machine learning predictive models to guide prevention and intervention allocation for anxiety and depressive disorders among college students,

    Y . Zhai, Y . Zhang, Z. Chu, B. Geng, M. Almaawali, R. Fulmer, Y .-W. D. Lin, Z. Xu, A. D. Daniels, Y . Liu et al. , “Machine learning predictive models to guide prevention and intervention allocation for anxiety and depressive disorders among college students,” Journal of Counseling & Development, vol. 103, no. 1, pp. 110–125, 2025

    work page 2025

  6. [5]

    Deep multimodal data fusion,

    F. Zhao, C. Zhang, and B. Geng, “Deep multimodal data fusion,” ACM Computing Surveys, vol. 56, no. 9, pp. 1– 36, 2024

    work page 2024

  7. [6]

    Impact of phased covid-19 vaccine roll- out on anxiety and depression among us adult popu- lation, january 2019–february 2023: a population-based interrupted time series analysis,

    Y . Zhai, M. Fan, B. Geng, X. Du, S. Snyder, and L. Wilkinson, “Impact of phased covid-19 vaccine roll- out on anxiety and depression among us adult popu- lation, january 2019–february 2023: a population-based interrupted time series analysis,” The Lancet Regional Health–Americas, vol. 37, 2024

    work page 2019

  8. [7]

    Distributed quantized detection of sparse signals under byzantine attacks,

    C. Quan, Y . S. Han, B. Geng, and P. K. Varshney, “Distributed quantized detection of sparse signals under byzantine attacks,” IEEE Transactions on Signal Process- ing, 2023

    work page 2023

Show all 30 references
  1. [8]

    Efficient ordered-transmission based distributed detection under data falsification attacks,

    C. Quan, N. Sriranga, H. Yang, Y . S. Han, B. Geng, and P. K. Varshney, “Efficient ordered-transmission based distributed detection under data falsification attacks,” IEEE Signal Processing Letters , vol. 30, pp. 145–149, 2023

    2023

  2. [9]

    On ordered transmission based dis- tributed gaussian shift-in-mean detection under byzantine attacks,

    C. Quan, S. Bulusu, B. Geng, Y . S. Han, N. Sriranga, and P. K. Varshney, “On ordered transmission based dis- tributed gaussian shift-in-mean detection under byzantine attacks,” IEEE Transactions on Signal Processing, 2023

    2023

  3. [10]

    Collaborative human decision making with heterogeneous agents,

    B. Geng, X. Cheng, S. Brahma, D. Kellen, and P. K. Varshney, “Collaborative human decision making with heterogeneous agents,” IEEE Transactions on Computa- tional Social Systems , vol. 9, no. 2, pp. 469–479, 2021

    2021

  4. [11]

    Enhanced audit bit based distributed bayesian detection in the presence of strategic attacks,

    C. Quan, B. Geng, Y . S. Han, and P. K. Varshney, “Enhanced audit bit based distributed bayesian detection in the presence of strategic attacks,” IEEE Transactions on Signal and Information Processing over Networks , vol. 8, pp. 49–62, 2022

    2022

  5. [12]

    Distributed learning in wireless networks: Recent progress and future chal- lenges,

    M. Chen, D. G ¨und¨uz, K. Huang, W. Saad, M. Bennis, A. V . Feljan, and H. V . Poor, “Distributed learning in wireless networks: Recent progress and future chal- lenges,” IEEE Journal on Selected Areas in Communica- tions, vol. 39, no. 12, pp. 3579–3605, 2021

    2021

  6. [13]

    Measuring heterogeneity in machine learning with distributed energy distance,

    M. Fan, B. Geng, R. Shterenberg, J. A. Casey, Z. Chen, and K. Li, “Measuring heterogeneity in machine learning with distributed energy distance,” arXiv preprint, 2025. [Online]. Available: https://arxiv.org/abs/2501.16174

    2025 arXiv

  7. [14]

    Towards personalized federated learning,

    A. Z. Tan, H. Yu, L. Cui, and Q. Yang, “Towards personalized federated learning,” IEEE transactions on neural networks and learning systems , vol. 34, no. 12, pp. 9587–9603, 2022

    2022

  8. [15]

    Federated learning with personalization layers,

    M. G. Arivazhagan, V . Aggarwal, A. K. Singh, and S. Choudhary, “Federated learning with personalization layers,” arXiv preprint arXiv:1912.00818 , 2019

    1912 arXiv

  9. [16]

    Fully decentralized federated learning,

    A. Lalitha, S. Shekhar, T. Javidi, and F. Koushanfar, “Fully decentralized federated learning,” in Third work- shop on bayesian deep learning (NeurIPS) , vol. 2, 2018

    2018

  10. [17]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” Proceedings of Machine learning and sys- tems, vol. 2, pp. 429–450, 2020

    2020

  11. [18]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,”International Journal of Computer Vision, vol. 129, no. 6, pp. 1789–1819, 2021

    2021

  12. [19]

    Federated multi-task learning under a mixture of distributions,

    O. Marfoq, G. Neglia, A. Bellet, L. Kameni, and R. Vi- dal, “Federated multi-task learning under a mixture of distributions,” Advances in Neural Information Process- ing Systems, vol. 34, pp. 15 434–15 447, 2021

    2021

  13. [20]

    Clustered fed- erated learning: Model-agnostic distributed multitask op- timization under privacy constraints,

    F. Sattler, K.-R. M ¨uller, and W. Samek, “Clustered fed- erated learning: Model-agnostic distributed multitask op- timization under privacy constraints,” IEEE transactions on neural networks and learning systems , vol. 32, no. 8, pp. 3710–3722, 2020

    2020

  14. [21]

    Fedbabu: Towards en- hanced representation for federated image classification,

    J. Oh, S. Kim, and S.-Y . Yun, “Fedbabu: Towards en- hanced representation for federated image classification,” arXiv preprint arXiv:2106.06042 , 2021

    2021 arXiv

  15. [23]

    Decentralized federated averaging,

    T. Sun, D. Li, and B. Wang, “Decentralized federated averaging,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 4, pp. 4289–4301, 2022

    2022

  16. [24]

    Dispfl: Towards communication-efficient personalized federated learning via decentralized sparse training,

    R. Dai, L. Shen, F. He, X. Tian, and D. Tao, “Dispfl: Towards communication-efficient personalized federated learning via decentralized sparse training,” in Interna- tional conference on machine learning . PMLR, 2022, pp. 4587–4604

    2022

  17. [25]

    Stochastic gradient push for distributed deep learn- ing,

    M. Assran, N. Loizou, N. Ballas, and M. Rabbat, “Stochastic gradient push for distributed deep learn- ing,” in International Conference on Machine Learning . PMLR, 2019, pp. 344–353

    2019

  18. [26]

    Decentralized directed collaboration for personalized federated learning,

    Y . Liu, Y . Shi, Q. Li, B. Wu, X. Wang, and L. Shen, “Decentralized directed collaboration for personalized federated learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 168–23 178

    2024

  19. [27]

    Enhancing decentralized federated learning for non-iid data on het- erogeneous devices,

    M. Chen, Y . Xu, H. Xu, and L. Huang, “Enhancing decentralized federated learning for non-iid data on het- erogeneous devices,” in 2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 2023, pp. 2289–2302

    2023

  20. [28]

    Freeze- out: Accelerate training by progressively freezing layers,

    A. Brock, T. Lim, J. M. Ritchie, and N. Weston, “Freeze- out: Accelerate training by progressively freezing layers,” arXiv preprint arXiv:1706.04983 , 2017

    2017 arXiv

  21. [29]

    Personalized fed- erated learning: A unified framework and universal opti- mization techniques,

    F. Hanzely, B. Zhao, and M. Kolar, “Personalized fed- erated learning: A unified framework and universal opti- mization techniques,” arXiv preprint arXiv:2102.09743 , 2021

    2021 arXiv

  22. [30]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics. PMLR, 2017, pp. 1273–1282

    2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.