Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

FedCGD: Collective Gradient Divergence Optimized Scheduling for Wireless Federated Learning

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 Wireless federated learning converges at a rate set by the scheduled group's collective gradient divergence plus sampling variance, so FedCGD picks complementary devices that cancel each other's data bias.

desk verdict A good collective-divergence idea, but the scheduling algorithms ignore the bandwidth constraint they are supposed to solve. read the letter →

arxiv 2506.07581 v1 pith:BA344EOE submitted 2025-06-09 cs.LG cs.AIcs.DC

Tan Chen , Jintao Yan , Yuxuan Sun , Sheng Zhou , Zhisheng Niu This is my paper · ORCID
classification cs.LGcs.AIcs.DC
keywords federatedlearningwirelessnetworksdeviceschedulingdataheterogeneitycollectivegradientdivergenceweightedearthmovingdistancesamplingvarianceconvergenceanalysis
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 claims that in wireless federated learning, what slows convergence is not the data bias of individual devices but the collective gradient divergence of the scheduled device group — the gap between the group's average gradient and the global gradient — added to the sampling variance of the scheduled data. The authors prove a bound (Theorem 1) placing exactly these two terms, plus a local-iteration bias of order $\tau^2$, on the federal-central difference, and they show that the device-level term is not minimized by picking individually 'good' devices: two complementary biased devices can form a group closer to the global distribution than two individually unbiased ones. To make the bound actionable for classification tasks, the device-level term is converted into the weighted earth moving distance (WEMD) between the scheduled group's class distribution and the global distribution, and the FedCGD algorithm schedules devices by minimizing WEMD plus sampling variance, with a greedy solver of $O(V^2)$ complexity and a fix-sum coordinate descent solver. If this is right, device selection in wireless FL should be treated as group composition rather than individual ranking, and the reported simulations show up to 4.2% higher accuracy on CIFAR-10 while using 41.8% fewer devices than the baselines.

What carries the argument

The load-bearing object is the collective gradient divergence (CGD) itself: the device-level term $\Delta^{(j)}$, the norm of the difference between the data-weighted average gradient of the scheduled group and the global gradient, together with the sample-level term bounded by $\sigma/\sqrt{|\Pi^{(j)}| b}$. The argument's hinge is Theorem 1, which shows these two terms plus a local-iteration bias of order $\tau(\tau-1)$ upper-bound the federal-central difference $U_j$, and the follow-up transformation Eq. (8), which bounds $\Delta^{(j)}$ for classification losses by the weighted earth moving distance (WEMD) between the scheduled group's class distribution and the global class distribution, weighted by class-gradient norms $G_c$. This converts an unobservable gradient quantity into a measurable scheduling objective over label distributions, formulated as problem P1 (minimize WEMD plus sampling variance under the total bandwidth budget). The NP-hardness proof (reduction from partition, Lemma 4) and the two approximate solvers — greedy scheduling at $O(|V|^2)$ and fix-sum coordinate descent (FSCD) at pseudo-polynomial cost — plus the online estimators for $\sigma$ and $G_c$ complete the machinery, since the schedules are chosen without ever observing the true gradients.

What would settle it

Run one FL task with two schedules matched in every way except collective divergence — for example, group A holding class distributions [0.8,0.2] and [0.2,0.8] and group B holding two identical [0.8,0.2] distributions, so the per-device divergences are the same while the group divergences differ. If the measured federal-central difference $\mathbb{E}[U_j]$ (or final accuracy) does not favor the complementary group A, the collective-divergence mechanism — the paper's central claim — is not what controls convergence speed.

Watch

Extended reading notes

Core claim

The central discovery is that data heterogeneity in FL device scheduling is a collective property. Theorem 1 bounds the federal-central difference by $\mathbb{E}[U_j] \le \tfrac{1}{2} \tau(\tau-1)\eta\beta g + \eta\tau\bigl(\sigma/\sqrt{|\Pi^{(j)}| b} + \Delta^{(j)}\bigr)$, where $\Delta^{(j)} = \bigl\|\sum_{v\in\Pi^{(j)}} \alpha_v^{(j)}\nabla f_v(w^{(j-1)}) - \nabla F(w^{(j-1)})\bigr\|$ is the norm of the difference between the scheduled group's average gradient and the global gradient, and $\sigma/\sqrt{|\Pi^{(j)}| b}$ is the sampling variance of the scheduled data. Because the device-level term is collective, individually divergent devices can cancel each other: the paper's example shows distributions [0.8,0.2] and [0.2,0.8] forming a group with near-zero divergence against a global [0.5,0.5] distribution, beating two devices each at [0.51,0.49]. For classification problems, $\Delta^{(j)}$ is bounded by the weighted earth moving distance (WEMD) between the scheduled group's class distribution and the global class distribution, which makes the scheduling problem a tractable optimization over label distributions: minimize WEMD plus sampling variance. That problem is NP-hard, and the paper proposes a polynomial greedy algorithm and a pseudo-polynomial fix-sum coordinate descent algorithm (average relative error 0.19%), with simulations reporting up to 4.2% higher accuracy on CIFAR-10 while scheduling 41.8% fewer devices.

Load-bearing premise

The load-bearing premise is that the total bandwidth budget does not need to be checked while scheduling: both proposed algorithms pick devices using only the data-divergence and sampling-variance objective, so on a tight channel the set they return can request more bandwidth than the system actually has, making the schedule infeasible in the wireless setting the paper claims to address.

Editorial extensions

If this is right

  • Device selection becomes a group-composition problem: the scheduler should seek a complementary set whose combined class distribution approaches the global one, rather than a ranking of individually 'good' devices; the paper's example shows two far-from-global devices forming a better group than two near-global ones.
  • The optimal schedule balances WEMD against sampling variance, so the number of devices to schedule is itself an output of the optimization — fewer devices when complementarity is easy (small $\alpha$), more devices when it is hard (large $\alpha$) — matching the switching behavior shown in the Dirichlet experiments.
  • Because Theorem 2 extends the CGD decomposition to non-convex losses, the same WEMD-plus-variance objective applies to deep models such as the CNN and ResNet-18 used in the simulations, not only to convex objectives.
  • The estimated ratio $G/\sigma$ acts as an indicator of when device-level divergence matters: when it is large, heterogeneity-aware scheduling such as FedCGD is needed; when small, a channel-only baseline suffices, suggesting the schedule can be adapted mid-training.
  • If the claim holds, wireless FL deployments can improve accuracy by up to 4.2% on CIFAR-10 while using 41.8% fewer devices than the best-effort baselines, because the scheduled group better represents the global distribution per unit of bandwidth.

Reading between the lines

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

  • The bandwidth budget is the place to probe: neither Algorithm 1 nor Algorithm 2 verifies constraint (7c) — Algorithm 2 even acknowledges that a 0-to-1 transit 'may violate' the bandwidth constraint — so on a tight 20 MHz channel the returned schedule can be infeasible; adding a post-hoc budget check that drops devices until $\sum_{v} x_v B_v^* \le B$ would show whether the reported gains survive r
  • The WEMD surrogate assumes a classification loss with equal-size device datasets; a direct estimator of $\Delta^{(j)}$ from the uploaded model deltas would extend the same collective-divergence principle to regression tasks or unequal dataset sizes without the label-distribution simplification.
  • The $G/\sigma$ ratio found on CIFAR-100 is effectively a heterogeneity signal-to-noise ratio, and it suggests a threshold-based scheduler that runs the cheap channel-only baseline when $G/\sigma$ is small and switches to FedCGD when heterogeneity dominates — a testable regime-switching extension the paper does not develop.
  • The NP-hardness reduction implies no polynomial exact algorithm for P1, so the 0.19% relative error of FSCD is only empirical; a provable approximation guarantee for the greedy algorithm would be the natural next step, and the paper's bound structure gives a benchmark for it.
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 studies joint device scheduling and bandwidth allocation for federated learning (FL) over wireless networks, arguing that data heterogeneity should be measured collectively for the scheduled group rather than per device. The authors prove, under convexity and gradient bounds, an upper bound on the federated-central difference, E[U_j] ≤ 1/2 τ(τ−1)ηβg + ητ(σ/√(|Π|b)+Δ), where Δ is the divergence between the weighted average gradient of scheduled devices and the global gradient. For classification with equal device dataset sizes, they upper-bound Δ by a weighted earth mover distance (WEMD) between the group class distribution and the global distribution. They then formulate P1, minimizing WEMD plus sampling variance subject to a total minimum-bandwidth constraint, claim P1 is NP-hard, and propose two heuristics (greedy GS and fix-sum coordinate descent FSCD) integrated into FedCGD. Simulations on CIFAR-10/100 report improved accuracy with fewer scheduled devices.

Significance. If correct, the collective-divergence viewpoint is a useful conceptual contribution: it explains why scheduling devices with small individual divergence may be suboptimal and why complementary devices can form a better group. The convergence decomposition is standard but cleanly isolates device-level and sample-level terms, and the WEMD-based proxy is a plausible tractable surrogate. The authors provide detailed appendix proofs for the main bounds and give a clear system model. However, as written, the paper does not establish that the proposed algorithms solve P1 or that the experimental gains are achieved under the stated bandwidth constraint. Several load-bearing assumptions (equal dataset sizes, the NP-hardness reduction) are either unstated or invalid. With careful fixes to the algorithms, the assumptions, and the reduction, the paper could become a valuable contribution.

major comments (3)
  1. [Section V-B, Algorithms 1 and 2] Neither Algorithm 1 nor Algorithm 2 enforces the total-minimum-bandwidth constraint (7c), sum_{v in Π} x_v B*_v ≤ B. Algorithm 1's stopping rule (line 7) compares only the WEMD-plus-variance objective, and bandwidth never enters the greedy selection. Algorithm 2 initializes with the S lowest-bandwidth devices but its coordinate-descent swaps can replace a low-bandwidth device with a high-bandwidth one, and the final selection over S (line 14) is made by objective value alone; the early-exit condition (line 10) also ignores bandwidth. The returned schedule can therefore violate (7c), so the algorithms do not solve P1 as claimed. Because the reported accuracy gains are generated by these schedules, the simulation results do not support the paper's wireless-FL setting with limited bandwidth.
  2. [Lemma 2 and Appendix C] Lemma 2 states that sample-level gradient divergence is bounded by σ/√(|Π|b), but the proof uses α_v = 1/|Π| and rewrites the divergence as 1/(|Π|b) times a sum over all samples. This is valid only when all devices have equal dataset sizes, |D_v| = |D| for all v. That assumption is not stated in Lemma 2 or elsewhere in Section IV-B; it appears only later in Section V-A. Since Theorem 1 and the objective in P1 rely on Lemma 2, the convergence bound and the optimization objective are not established for arbitrary device dataset sizes. The lemma should either state the equal-size assumption or carry the general aggregation weights α_v = |D_v|/Σ|D_v| through the proof.
  3. [Lemma 4 and Appendix F] The claimed polynomial-time reduction from Partition sets C=1 and p_{v,0}=r_v, p_0=c_sum/(2s). For a classification problem with a single class, the only valid class-distribution vector is p_{v,0}=p_0=1, not arbitrary integers. The constructed instance is therefore not a valid instance of P1, and the Partition problem is not shown to reduce to P1 by the given construction. This affects the paper's claim that P1 is NP-hard and weakens the motivation for the heuristics. The reduction needs to be repaired, for example by using at least two classes with normalized proportions.
minor comments (6)
  1. [Algorithm 1, line 7] The greedy algorithm initializes Π = ∅, but the stopping condition uses 1/sqrt(|Π|), which is undefined for |Π| = 0 at the first iteration; the base case needs to be specified explicitly.
  2. [Section III and Section V-A] The equal-|D_v| assumption is introduced only for the WEMD transformation, but it is also needed by Lemma 2 and by the definition of the global loss F(w)=E_v[f_v(w)] in Section III; the assumption should be stated in the system model.
  3. [Appendix E, proof of Theorem 2] The line 'the last inequality holds because of Lemma 3 and the fact EX≤√E[X^2]' should refer to Lemma 2, since Lemma 3 bounds local iteration bias, not sample-level divergence.
  4. [Section V-B, Algorithm 2] The early-exit condition (line 10) uses W(Π) where Π is the current local optimum for a fixed schedule size S; the notation would be clearer if the dependence on S were explicit, since W was previously defined for arbitrary sets.
  5. [Section VI, Table II discussion] The text writes 'only our method FCGD schedules much fewer devices' where the acronym should be FedCGD; please unify the notation.
  6. [Section VI-B, Figure 3] The text reports relative errors of 5.16% and 0.19% for GS and FSCD but does not state how the optimal objective is computed or what the 'CD algorithm' baseline is; please clarify the evaluation protocol for Fig. 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the WEMD-plus-variance objective in P1 is a direct upper bound from Theorem 1 and Eq. (8), and the reported gains are empirical rather than fitted.

full rationale

The derivation chain starts with a standard FL convergence decomposition (Lemma 1, Theorem 1 in Section IV) that bounds the federated-central difference by local-iteration bias, sampling variance, and the collective device-level gradient divergence Δ. Section V-A then upper-bounds Δ by the weighted L1 divergence (WEMD) in Eq. (8), and P1 minimizes the sum of this bound and the Lemma 2 variance term. This is a conventional bound-minimization design: the algorithm is an application of the bound, not an input to it. The parameters σ and G_c are plug-in estimates (Eqs. (10)-(12)) computed from online local updates, not constants fitted to the reported accuracy, so no fitted quantity is renamed as a prediction. The self-citations ([9] for the Lambert-W minimum-bandwidth formula, [41] for standard assumptions, [47] for coordinate descent) are peripheral and not load-bearing; none is invoked as a uniqueness theorem or as the justification for the central convergence result, which is proven in the appendices. The main weakness is non-circular: Algorithms 1 and 2 never check constraint (7c), so the returned schedule can violate the total bandwidth budget; this is an infeasibility/correctness risk in the wireless setting, not a circularity of the derivation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard FL smoothness and variance assumptions, plus a classification-specific loss structure and an equal-dataset-size assumption that is introduced later than the theorems that need it. The algorithm also depends on online estimates of sigma and G. No physical or formal entities are invented; the 'collective gradient divergence' is a defined composite quantity.

free parameters (2)
  • sigma_hat (per-round estimated gradient variance bound) = estimated online via Eq. (10)-(11)
    Plug-in estimate of sigma used in the sampling variance term sigma/sqrt(|Pi|b); the scheduling decision depends on its accuracy.
  • G_hat (estimated class gradient norm bound) = estimated online via Eq. (12)
    Weights the WEMD term in P1; estimation is heuristic and directly changes the balance between WEMD and sampling variance.
assumptions (4)
  • domain assumption Assumption 1: each local loss is convex, rho-Lipschitz, beta-smooth, gradients uniformly bounded by g, and stochastic gradients are unbiased with variance bounded by sigma^2
    Invoked in Section IV (Assumption 1) and used in the proofs of Proposition 1, Lemmas 2-3, and Theorems 1-2.
  • ad hoc to paper Equal per-device dataset sizes: |D_v| equal for all v, giving alpha_v = 1/|Pi|
    Stated only in Section V.A; Lemma 2 and Theorem 1 are proved and presented without this assumption, so the general claim is not supported when dataset sizes differ.
  • domain assumption Classification loss structure F(w)=sum_c p_c E_{x|y=c}[log g_i(w)] with class-gradient bound |grad_w E_{x|y=c}[log g_i(w)]| <= G_c
    Required for the WEMD transformation in Eq. (8); confines the theoretical justification to classification, not general FL tasks.
  • domain assumption Each device samples a batch of size b uniformly from its local dataset, and sampling is independent across devices
    Used in Lemma 2 and the P1 objective; batch size b is fixed in simulations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedCGD: Collective Gradient Divergence Optimized Scheduling for Wireless Federated Learning." pith.science (2026). https://pith.science/paper/BA344EOE

@misc{pith2026250607581,
  author       = {Pith},
  title        = {Pith review of: FedCGD: Collective Gradient Divergence Optimized Scheduling for Wireless Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BA344EOE}},
  note         = {Machine review of arXiv:2506.07581}
}
read the original abstract

Federated learning (FL) is a promising paradigm for multiple devices to cooperatively train a model. When applied in wireless networks, two issues consistently affect the performance of FL, i.e., data heterogeneity of devices and limited bandwidth. Many papers have investigated device scheduling strategies considering the two issues. However, most of them recognize data heterogeneity as a property of individual devices. In this paper, we prove that the convergence speed of FL is affected by the sum of device-level and sample-level collective gradient divergence (CGD). The device-level CGD refers to the gradient divergence of the scheduled device group, instead of the sum of the individual device divergence. The sample-level CGD is statistically upper bounded by sampling variance, which is inversely proportional to the total number of samples scheduled for local update. To derive a tractable form of the device-level CGD, we further consider a classification problem and transform it into the weighted earth moving distance (WEMD) between the group distribution and the global distribution. Then we propose FedCGD algorithm to minimize the sum of multi-level CGDs by balancing WEMD and sampling variance, within polynomial time. Simulation shows that the proposed strategy increases classification accuracy on the CIFAR-10 dataset by up to 4.2\% while scheduling 41.8\% fewer devices, and flexibly switches between reducing WEMD and reducing sampling variance.

Figures

Figures reproduced from arXiv: 2506.07581 by the authors.

Figure 1
Figure 1. Illustration of wireless FL with heterogeneous data. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of multi-level collective gradient divergences. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance of proposed algorithms. FL simulations are conducted on the CIFAR-10 dataset and CIFAR-100 dataset1 for the image classification task. Both of them have 50000 training samples and 10000 testing samples in total. CIFAR-10 has 10 classes, with 5000 training samples and 1000 test samples for each class; CIFAR-100 has 100 classes, with 500 training samples and 100 test samples for each class. Two types of da… view at source ↗
Figures from the paper (5 more)
Figure 6
Figure 6. Figure 6: Dirichlet data distribution with different [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Epochs to reach the target test accuracy of baselines with different [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 5
Figure 5. Figure 5: Max test accuracy of baselines with different imbalance ratios. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 8
Figure 8. Figure 8: Average scheduled number of baselines with different [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Performance of baselines on the CIFAR-100. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FedTeddi: Temporal Drift and Divergence Aware Scheduling for Timely Federated Edge Learning

    cs.LG 2025-09 conditional novelty 5.0 of 10

    A temporal-drift and collective-divergence aware greedy client scheduler plus bandwidth allocator accelerates convergence in federated edge learning with streaming, non-i.i.d. data.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [39]

    Coalitional FL: Coalition formation and selection in federated learning with heterogeneous data,

    N. Zhang, Q. Ma, W. Mao, and X. Chen, “Coalitional FL: Coalition formation and selection in federated learning with heterogeneous data,” IEEE Trans. Mobile Comput., vol. 23, no. 11, pp. 10494–10508, Nov. 2024

    work page 2024

  2. [40]

    Fed-cbs: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction,

    J. Zhang, A. Li, M. Tang, J. Sun, X. Chen, F. Zhang, C. Chen, Y . Chen, and H. Li, “Fed-cbs: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction,” inInt. Conf. on Machine Learning, Honolulu, HI, USA, Jul. 2023, pp. 41354–41381

    work page 2023

  3. [1]

    Communication-efficient learning of deep networks from decentral- ized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentral- ized data,” inProc. Artificial Intelligence and Statistics (AIStats), Ft. Lauderdale, FL, USA, Apr. 2017, pp. 1273–1282

    work page 2017

  4. [2]

    Adaptive federated optimization,

    S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Kone ˇcn`y, S. Kumar, and H. B. McMahan, “Adaptive federated optimization,” in Int. Conf. on Learning Representations (ICLR), May 2021

    work page 2021

  5. [3]

    Federated learning in mobile edge networks: A comprehensive survey,

    W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y .-C. Liang, Q. Yang, D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A comprehensive survey,”IEEE Commun. Surveys & Tutorials, vol. 22, no. 3, pp. 2031-2063, Feb. 2020

    work page 2020

  6. [4]

    Federated learning and wireless communications,

    Z. Qin, G. Y . Li, and H. Ye, “Federated learning and wireless communications,”IEEE Wireless Commun., vol. 28, no. 5, pp. 134– 140, Oct. 2021

    work page 2021

  7. [5]

    Addressing class imbalance in federated learning,

    L. Wang, S. Xu, X. Wang, and Q. Zhu, “Addressing class imbalance in federated learning,” inProc. AAAI Conf. Artif. Intell., Feb. 2021, vol. 35, pp. 10165–10173

    work page 2021

  8. [6]

    Edge learning with timeliness constraints: Challenges and solutions,

    Y . Sun, W. Shi, X. Huang, S. Zhou and Z. Niu, “Edge learning with timeliness constraints: Challenges and solutions,”IEEE Commun. Mag., vol. 58, no. 12, pp. 27–33, Dec. 2020

    work page 2020

Show all 52 references
  1. [7]

    MOB-FL: Mobility-aware federated learning for intelligent connected vehicles,

    B. Xie, Y . Sun, S. Zhou, Z. Niu, Y . Xu, J. Chen, and D. Gunduz, “MOB-FL: Mobility-aware federated learning for intelligent connected vehicles,” inIEEE Int. Conf. Commun. (ICC), Rome, Italy, May 2023, pp. 3951-3957

    2023

  2. [8]

    Dynamic scheduling for vehicle-to-vehicle communications enhanced federated learning,

    J. Yan, T. Chen, Y . Sun, Z. Nan, S. Zhou and Z. Niu, “Dynamic scheduling for vehicle-to-vehicle communications enhanced federated learning,”IEEE Trans. Wireless Commun., early access, Jun. 2025

    2025

  3. [9]

    Joint device scheduling and resource allocation for latency constrained wireless federated learning,

    W. Shi, S. Zhou, Z. Niu, M. Jiang, and L. Geng, “Joint device scheduling and resource allocation for latency constrained wireless federated learning,”IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 453–467, Jan. 2021

    2021

  4. [10]

    Federated learning with non-iid data,

    Y . Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V . Chandra, “Federated learning with non-iid data,”arXiv preprint arXiv:1806.00582, 2018

    2018 arXiv

  5. [11]

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

    E. Jeong, S. Oh, H. Kim, J. Park, M. Bennis, and S.-L. Kim, “Communication-efficient on-device machine learning: Federated dis- tillation and augmentation under non-iid private data,” inNeural Inform. Process. Syst. Workshop, Montr ´eal, Canada, Dec. 2018

    2018

  6. [12]

    Scaffold: Stochastic controlled averaging for federated learn- ing,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “Scaffold: Stochastic controlled averaging for federated learn- ing,” inInt. Conf. on Machine Learning, Jul. 2020, pp. 5132–5143

    2020

  7. [13]

    Federated learning with label distribution skew via logits calibration,

    J. Zhang, Z. Li, B. Li, J. Xu, S. Wu, S. Ding, and C. Wu, “Federated learning with label distribution skew via logits calibration,” inInt. Conf. on Machine Learning, Baltimore, MD, USA, Jul. 2022, pp. 26311– 26329

    2022

  8. [14]

    Federated learning with personalization layers,

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

    1912 arXiv

  9. [15]

    Adaptive federated learning in resource constrained edge computing systems,

    S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and K. Chan, “Adaptive federated learning in resource constrained edge computing systems,”IEEE J. Sel. Areas Commun., vol. 37, no. 6, pp. 1205–1221, Jun. 2019

    2019

  10. [16]

    Convergence time optimization for federated learning over wireless networks,

    M. Chen, H. V . Poor, W. Saad, and S. Cui, “Convergence time optimization for federated learning over wireless networks,”IEEE Trans. Wireless Commun., vol. 20, no. 4, pp. 2457–2471, Apr. 2021

    2021

  11. [17]

    Joint model pruning and device selection for communication-efficient federated edge learning ,

    S. Liu, G. Yu, R. Yin, J. Yuan, L. Shen, and C. Liu, “Joint model pruning and device selection for communication-efficient federated edge learning ,”IEEE Trans. Commun., vol. 70, no. 1, pp. 231–244, Jan. 2022

    2022

  12. [18]

    Communication-efficient federated edge learning via opti- mal probabilistic device scheduling,

    M. Zhang, G. Zhu, S. Wang, J. Jiang, Q. Liao, C. Zhong, and S. Cui, “Communication-efficient federated edge learning via opti- mal probabilistic device scheduling,”IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8536–8551, Oct. 2022

    2022

  13. [19]

    Beamforming vector design and device selection in over-the-air federated learning,

    M. Kim, A. L. Swindlehurst, and D. Park, “Beamforming vector design and device selection in over-the-air federated learning,”IEEE Trans. Wireless Commun., vol. 22, no. 11, pp. 7464–7477, Nov. 2023

    2023

  14. [20]

    Asynchronous wireless federated learning with probabilistic client selection,

    J. Yang, Y . Liu, F. Chen, W. Chen, and C. Li, “Asynchronous wireless federated learning with probabilistic client selection,”IEEE Trans. Wireless Commun., vol. 23, no. 7, pp. 7144–7158, Jul. 2024

    2024

  15. [21]

    Energy harvesting aware client selection for over-the-air federated learning,

    C. Chen, Y .-H. Chiang, H. Lin, J. C. Lui, and Y . Ji, “Energy harvesting aware client selection for over-the-air federated learning,” inIEEE Global Commun. Conf., Rio de Janeiro, Brazil, Dec. 2022, pp. 5069- 5074

    2022

  16. [22]

    Scheduling policies for federated learning in wireless networks,

    H. H. Yang, Z. Liu, T. Q. S. Quek, and H. V . Poor, “Scheduling policies for federated learning in wireless networks,”IEEE Trans. Commun., vol. 68, no. 1, pp. 317–333, Jan. 2020

    2020

  17. [23]

    Context-aware online client selection for hierarchical federated learning,

    Z. Qu, R. Duan, L. Chen, J. Xu, Z. Lu, and Y . Liu, “Context-aware online client selection for hierarchical federated learning,”IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 12, pp. 4353–4367, Dec. 2022

    2022

  18. [24]

    Green federated learning via energy-aware client selection,

    R. Albelaihi, L. Yu, W. D. Craft, X. Sun, C. Wang, and R. Gazda, “Green federated learning via energy-aware client selection,” inIEEE Global Commun. Conf., Rio de Janeiro, Brazil, Dec. 2022, pp. 13-18

    2022

  19. [25]

    Joint device scheduling and bandwidth allocation for federated learning over wireless networks,

    T. Zhang, K.-Y . Lam, J. Zhao, and J. Feng, “Joint device scheduling and bandwidth allocation for federated learning over wireless networks,” IEEE Trans. Wireless Commun., vol. 24, no. 1, pp. 3–18, Jan. 2025

    2025

  20. [26]

    Joint client selection and resource allocation for federated learning in mobile edge networks,

    L. Luo, Q. Cai, Z. Li, and H. Yu, “Joint client selection and resource allocation for federated learning in mobile edge networks,” inIEEE Wireless Commun. and Netw. Conf. (WCNC), Austin, TX, USA, Apr. 2022, pp. 1218-1223

    2022

  21. [27]

    Contextual client selection for efficient federated learning over edge devices,

    Q. Pan, H. Cao, Y . Zhu, J. Liu, and B. Li, “Contextual client selection for efficient federated learning over edge devices,”IEEE Trans. Mobile Comput., vol. 23, no. 6, pp. 6538–6548, June 2024

    2024

  22. [28]

    Scheduling for cellular federated edge learning with importance and channel awareness,

    J. Ren, Y . He, D. Wen, G. Yu, K. Huang, and D. Guo, “Scheduling for cellular federated edge learning with importance and channel awareness,”IEEE Trans. Wireless Commun., vol. 19, no. 11, pp. 7690– 7703, Nov. 2020

    2020

  23. [29]

    Convergence of update aware device scheduling for federated learning at the wireless edge,

    M. M. Amiri, D. G ¨und¨uz, S. R. Kulkarni, and H. V . Poor, “Convergence of update aware device scheduling for federated learning at the wireless edge,”IEEE Trans. Wireless Commun., vol. 20, pp. 3643–3658, no. 6, Jun. 2021

    2021

  24. [30]

    Adaptive heterogeneous client sampling for federated learning over wireless networks,

    B. Luo, W. Xiao, S. Wang, J. Huang, and L. Tassiulas, “Adaptive heterogeneous client sampling for federated learning over wireless networks,”IEEE Trans. Mobile Comput., vol. 23, no. 10, pp. 9663- 9677, Oct. 2024

    2024

  25. [31]

    Data-importance aware user scheduling for communication-efficient edge machine learning,

    D. Liu, G. Zhu, J. Zhang, and K. Huang, “Data-importance aware user scheduling for communication-efficient edge machine learning,”IEEE Trans. Cognitive Commun. Netw., vol. 7, no. 1, pp. 265–278, Mar. 2021

    2021

  26. [32]

    Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective,

    J. Xu and H. Wang, “Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective,”IEEE Trans. Wireless Commun., vol. 20, no. 2, pp. 1188–1200, Feb. 2021

    2021

  27. [33]

    FedCor: Correlation-based active client selection strategy for hetero- geneous federated learning,

    M. Tang, X. Ning, Y . Wang, J. Sun, Y . Wang, H. Li, and Y . Chen, “FedCor: Correlation-based active client selection strategy for hetero- geneous federated learning,” inProceed. IEEE Conf. on Comput. Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, Jun. 2022, pp. ...

    2022

  28. [34]

    AUCTION: Automated and quality-aware client selection framework for efficient federated learning,

    Y . Deng, F. Lyu, J. Ren, H. Wu, Y . Zhou, Y . Zhang, and X. Shen, “AUCTION: Automated and quality-aware client selection framework for efficient federated learning,”IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 8, pp. 1996–2009, Aug. 2022

    1996

  29. [35]

    Data-centric client selection for federated learning over distributed edge networks,

    R. Saha, S. Misra, A. Chakraborty, C. Chatterjee, and P. K. Deb, “Data-centric client selection for federated learning over distributed edge networks,”IEEE Trans. Parallel Distrib. Syst., vol. 34, no. 2, pp. 675–686, Feb. 2023

    2023

  30. [36]

    Data distribution-aware online client selection algorithm for federated learning in heterogeneous networks,

    J. Lee, H. Ko, S. Seo, and S. Pack, “Data distribution-aware online client selection algorithm for federated learning in heterogeneous networks,”IEEE Trans. Veh. Technol., vol. 72, no. 1, pp. 1127–1136, Jan. 2023

    2023

  31. [37]

    Stabilizing and accelerating federated learning on heterogeneous data with partial client participation,

    H. Zhang, C. Li, W. Dai, Z. Zheng, J. Zou, and H. Xiong, “Stabilizing and accelerating federated learning on heterogeneous data with partial client participation,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 1, pp. 67–83, Jan. 2025

    2025

  32. [38]

    Channel and gradient- importance aware device scheduling for over-the-air federated learn- ing,

    Y . Sun, Z. Lin, Y . Mao, S. Jin, and J. Zhang, “Channel and gradient- importance aware device scheduling for over-the-air federated learn- ing,”IEEE Trans. Wireless Commun., vol. 23, no. 7, pp. 6905–6920, Jul. 2024

    2024

  33. [41]

    Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,

    T. Chen, J. Yan, Y . Sun, S. Zhou, D. G ¨und¨uz, and Z. Niu, “Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,”IEEE Trans. Veh. Technol., vol. 74, no. 1, pp. 1657–1673, Jan. 2025

    2025

  34. [42]

    On the conver- gence of fedavg on non-iid data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the conver- gence of fedavg on non-iid data,”Proc. of the Int. Conf. on Learning Representations, New Orleans, LA, USA, May 2019

    2019

  35. [43]

    Mobility-aware cluster federated learning in hierarchical wireless 19 networks,

    C. Feng, H. H. Yang, D. Hu, Z. Zhao, T. Q. Quek, and G. Min, “Mobility-aware cluster federated learning in hierarchical wireless 19 networks,”IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8441– 8458, Oct. 2022

    2022

  36. [44]

    Dynamic scheduling for over-the-air federated edge learning with energy constraints,

    Y . Sun, S. Zhou, Z. Niu, and D. G ¨und¨uz, “Dynamic scheduling for over-the-air federated edge learning with energy constraints,” inIEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 227-242, Jan. 2022

    2022

  37. [45]

    Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization,

    A. Reisizadeh, A. Jadbabaie, A. Mokhtari, H. Hassani, and R. Pedarsani, “Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization,” inProc. Artificial Intelligence and Statistics (AIStats), Aug. 2020, pp. 2021–2031

    2020

  38. [46]

    Federated learning with flexible control,

    S. Wang, J. Perazzone, M. Ji, and K. S. Chan, “Federated learning with flexible control,” inIEEE Conf. on Comput. Commun., New York City, NY , USA, 2023, pp. 1-10

    2023

  39. [47]

    Robust task offloading and resource allocation under imperfect computing capacity informa- tion in edge intelligence systems,

    Z. Nan, Y . Han, J. Yan, S. Zhou, and Z. Niu, “Robust task offloading and resource allocation under imperfect computing capacity informa- tion in edge intelligence systems,”IEEE Trans. Mobile Comput., early access, Feb. 2025

    2025

  40. [48]

    Measuring the effects of non- identical data distribution for federated visual classification,

    T.-M. H. Hsu, H. Qi, and M. Brown, “Measuring the effects of non- identical data distribution for federated visual classification,”arXiv preprint arXiv:1909.06335, 2019

    1909 arXiv

  41. [49]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceed. IEEE Conf. on Comput. Vision and Pattern Recognition (CVPR), Las Vegas, NV , USA, Jun. 2016, pp. 770–778

    2016

  42. [50]

    Why batch normalization damage federated learning on non-iid data?,

    Y . Wang, Q. Shi, and T.-H. Chang, “Why batch normalization damage federated learning on non-iid data?,”IEEE Trans. Neural Netw. Learn. Syst, vol. 36, no. 1, pp. 1692-1706, Jan. 2025

    2025

  43. [51]

    Towards understanding biased client selection in federated learning,

    Y . J. Cho, J. Wang, and G. Joshi, “Towards understanding biased client selection in federated learning,” in Proc. Artificial Intelligence and Statistics (AIStats), May 2022, pp. 10351–10375

    2022

  44. [52]

    Study on channel model for frequencies from 0.5 to 100 GHz,

    3rd Generation Partnership Project (3GPP), “Study on channel model for frequencies from 0.5 to 100 GHz,” 3GPP, Sophia Antipolis, France, Tech. Rep. 38.901, v18.0.0, Mar. 2024

    2024

Pith tools

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