Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Air-FedGA: A Grouping Asynchronous Federated Learning Mechanism Exploiting Over-the-air Computation

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

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

Pith's one-line read Air-FedGA claims that grouping workers for over-the-air aggregation and letting groups update asynchronously converges at a rate set by staleness and data skew, cutting federated training time by 29.9%-71.6%.

desk verdict A sensible combination of AirComp and grouped asynchrony with a directly measured speedup, but the proof has a noise-term slip and the optimization proxy is unvalidated. read the letter →

arxiv 2507.05704 v1 pith:XYEY7UJ7 submitted 2025-07-08 cs.DC

Qianpiao Ma , Junlong Zhou , Xiangpeng Hou , Jianchun Liu , Hongli Xu , Jianeng Miao , Qingmin Jia This is my paper
classification cs.DC
keywords federatedlearningover-the-aircomputationasynchronousworkergroupingedgeheterogeneitynon-IIDdataconvergenceanalysispowercontrol
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

Federated learning over wireless edge networks faces two opposing constraints: over-the-air computation (AirComp) can aggregate model updates in one transmission, but it demands that workers transmit together, which forces a server to wait for the slowest worker. This paper proposes Air-FedGA, which splits workers into groups, performs AirComp aggregation inside each group, and lets groups update the global model asynchronously. It proves a linear convergence bound for the resulting procedure, and derives a worst-case training-time expression that it optimizes by jointly choosing transmit-power scaling factors, denoising factors, and the worker grouping. If the bound is tight, the mechanism converts the straggler wait into a tunable staleness that can be balanced against group size, yielding reported speedups of 29.9%-71.6% over state-of-the-art AirComp and asynchronous baselines.

What carries the argument

The central object is the grouped asynchronous AirComp update rule (Eq. 10): each round $t$, one group $V_{j_t}$ transmits concurrently, the parameter server receives $y_t = \sum_{i \in V_{j_t}} d_i \sigma_t w_i^t + z_t$, and forms $w_t = (1 - \sum_{i \in V_{j_t}} \alpha_i) w_{t-1} + y_t/(D \sqrt{\eta_t})$. The two design knobs are the power scaling factor $\sigma_t$ at each worker and the denoising factor $\eta_t$ at the server, which together set the per-round aggregation error $C_t$. The convergence argument rests on a delayed-recursion lemma (Lemma 1, from [33]) that turns a recurrence $Q(t) \leq x Q(t-1) + y Q(l_t) + z$ into $Q(t) \leq \rho^t Q(0) + \delta$ with $\rho = (x+y)^{1/(1+\tau_{\max})}$, which is what converts staleness $\tau_{\max}$ into a slower contraction rate. The grouping algorithm sorts workers by data size and greedily assigns each to the group that minimizes the worst-case training-time proxy while keeping per-group training times within a factor $\xi$ of each other.

What would settle it

Run Air-FedGA on a simulated heterogeneous edge with controlled staleness, and compare the measured number of rounds to reach target accuracy against the bound's prediction $(1+\tau_{\max}) \log_B A$ for multiple groupings. If a grouping that the algorithm rejects (for example, one with larger $\tau_{\max}$ but much smaller per-round time) reaches the target accuracy faster than the algorithm's chosen grouping, then the bound-based proxy is not optimizing true training time.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that asynchronous, grouped over-the-air aggregation converges at a rate governed by staleness and inter-group data-distribution mismatch, and that this structure can be optimized to cut wall-clock training time. Theorem 1 gives $E[F(w_T)] - F(w^*) \leq \rho^T (F(w_0)-F(w^*)) + \delta$, with $\rho = [1 - (2\mu\gamma - \mu/L)\sum_j \psi_j \beta_j]^{1/(1+\tau_{\max})}$ and $\delta$ a residual that grows with the earth-mover distance $\Lambda_j$ between each group's label distribution and the global one, and with the AirComp error term $C_t = (\sigma_t/\sqrt{\eta_t} - 1)^2 W_t^2 + \sigma_0^2/(D_{j_t}^2 \eta_t)$. The paper then replaces the true time-to-accuracy objective with the bound $T \geq (1+\tau_{\max}) \log_B A$, where $A = (\epsilon - \delta)/(F(w_0)-F(w^*))$ and $B = 1 - (2\mu\gamma - \mu/L)\sum_j \psi_j \beta_j$, and proposes a greedy grouping algorithm plus an alternating power and denoising control algorithm to minimize that expression. The load-bearing prediction is that grouping workers with similar local training times, while rebalancing label distributions across groups to shrink $\Lambda_j$, reduces both the per-round wait and the number of rounds needed.

Load-bearing premise

The optimization's optimality rests on treating the worst-case round bound $T \geq (1+\tau_{\max}) \log_B A$ as a faithful proxy for real training time, with $\tau_{\max}$ estimated by a heuristic (Eq. 39); if that bound is loose, the grouping and power choices the algorithm returns need not minimize actual time-to-accuracy.

Editorial extensions

If this is right

  • If Air-FedGA is correct, it preserves AirComp's communication efficiency while avoiding full synchrony: the server waits only for the slowest worker in a group, not the slowest in the whole system.
  • The convergence bound predicts that reducing inter-group data skew (EMD $\Lambda_j$) shrinks the residual error $\delta$, so grouping to make distributions more IID should improve final accuracy as well as speed.
  • The greedy grouping algorithm, by aligning training times within groups and balancing distributions across groups, should scale to large $N$: total training time decreases with the number of workers in the simulations.
  • The power control algorithm, by minimizing $C_t$ under per-worker energy budgets, should keep Air-FedGA's extra energy consumption modest compared with synchronous AirComp.
  • The convergence rate depends on the maximum staleness $\tau_{\max}$, so the mechanism's benefit is largest when group sizes are chosen so that no group falls too far behind.

Reading between the lines

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

  • The bound-based optimization is a worst-case proxy; a direct comparison of predicted versus actual round counts on real workloads would reveal whether the proxy's looseness changes which grouping is optimal.
  • Lemma 1 treats staleness only through its maximum $\tau_{\max}$; if the staleness distribution is heavy-tailed, a bound using the full staleness profile might yield different grouping decisions, an extension the paper does not explore.
  • The convergence proof assumes L-smoothness and strong convexity, while experiments use CNNs and VGG; a testable extension is to measure whether the same $\rho$ and $\delta$ formulas approximately predict convergence on non-convex models in practice.
  • Air-FedGA's grouping could be combined with staleness-weighted aggregation (assigning smaller weights to staler groups) as a further speedup; the paper does not consider this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper proposes Air-FedGA, a federated learning mechanism that combines over-the-air computation with group-wise asynchronous updates. Workers are partitioned into groups; within each group, local models are aggregated over a noisy MAC, while different groups update the global model asynchronously. The authors provide a convergence analysis (Theorem 1), a training-time minimization formulation (P1/P2) with power-control and worker-grouping algorithms, and experiments on MNIST, CIFAR-10, and ImageNet-100 against several baselines. The headline empirical claim is a 29.9%-71.6% training-time speedup over state-of-the-art solutions, supported by direct time-to-accuracy measurements.

Significance. If the identified issues are fixed, the paper makes a useful contribution: it proposes a practical middle ground between fully synchronous AirComp and fully asynchronous FL, gives a quantitative convergence bound that exposes the roles of staleness, data heterogeneity (EMD), and AirComp noise, and backs the mechanism with multi-dataset experiments. The direct time-to-accuracy measurements are a strength, since the speedup claim is supported by measured curves rather than by the theoretical bound alone. The convergence derivation is a competent rerun of the standard template, but with an inconsistency in Eq. (17) and some missing assumptions; the optimization component is not fully validated as a true training-time minimizer.

major comments (4)
  1. [IV-B, Eq. (17)] In the definition of the aggregation error ε_j^t, the noise term is written as ∑_{vi∈V_jt} z_t/(D_jt√η_t). Since z_t is a single AWGN vector added in Eq. (9), the noise contribution should be z_t/(D_jt√η_t). As written, the noise variance is inflated by |V_jt|^2, which is inconsistent with Ct in Eq. (30) that uses σ_0^2/(D_jt^2 η_t). This inconsistency directly affects the residual term δ in Theorem 1 and must be corrected.
  2. [IV-A, Assumption 2; IV-B, Theorem 1] Theorem 1 states that ρ ∈ (0,1), but this requires 2μγ - μ/L > 0, i.e., μ > 0 and γ > 1/(2L). Assumption 2 only assumes μ ≥ 0; with μ = 0, the contraction factor is 1 and δ has a zero denominator. The assumption should be changed to μ > 0, and the remark in Section VI-A that the mechanism also performs well for non-convex CNN models should be separated from the theoretical guarantee, which requires strong convexity.
  3. [V-A, Eqs. (37)-(40)] The transformation of P1 into P2 is valid only when ε > δ and A = (ε - δ)/(F(w0) - F(w*)) lies in (0,1); otherwise log_B A is undefined or the lower bound is vacuous. The paper never checks these conditions for the experimental settings. In addition, τmax in Eq. (39) is a heuristic (L_max * Σ_j 1/L_j) with no derivation or comparison to measured staleness, and the paper does not verify that minimizing the P2 proxy minimizes the actual training time. The direct time-to-accuracy measurements in Fig. 3 do support the empirical speedup, but the optimality claim for the grouping and power-control algorithms is not established.
  4. [IV-B, Eq. (29)] The proof takes an expectation over group selection using fixed probabilities ψ_j, but the asynchronous protocol does not specify a stochastic process for group participation; the sequence of participating groups is determined by completion times and is not necessarily i.i.d. with probabilities ψ_j. The theorem should either model the group selection process explicitly or state that ψ_j is the empirical relative frequency and interpret the result as an average-over-time bound. As written, the expectation step in the proof of Theorem 1 is not fully justified.
minor comments (5)
  1. [V-A, Eq. (35)] The symbol L is overloaded: it denotes the smoothness constant in Assumption 1 and the average per-round time in Eq. (35). Please use a different symbol for one of them to avoid confusion.
  2. [V-B, Algorithm 2, line 4] The expression min{√η_t} ⋃ { h_i^t √E_i/(d_i W_t) | ∀v_i∈V } is ambiguous; use min( {√η_t} ∪ { h_i^t √E_i/(d_i W_t) | ∀v_i∈V } ) for clarity.
  3. [VI-B-1] The abstract claims a 29.9%-71.6% speedup, but the only numerical comparison reported in the text is for LR on MNIST at 80% accuracy; please report the corresponding speedups for the other model/dataset combinations or qualify the claim.
  4. [IV-C, Corollary 2] The statement that τmax depends partly on the number M of groups is vague; for M > 1, the maximum staleness also depends on completion times and the scheduling policy. Clarify the relationship.
  5. [Fig. 2] The displayed equation for w_1 lacks parentheses around the numerator; the term z_1/(D√η_1) appears outside the sum, making the expression ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence bound is derived from stated assumptions and an external lemma, and the claimed speedups are measured empirically rather than predicted from fitted constants.

full rationale

The paper's central derivation chain is self-contained. Theorem 1 (Section IV-B, Eq. 13) is obtained from Assumptions 1-4 plus Lemma 1, which is an external delayed-iteration convergence lemma (Feyzmahdavian et al., ref [33]); the proof of Lemma 1 is not supplied in the paper, but it is an independent mathematical result, not a self-citation. The constants ρ and δ are expressed in terms of stated quantities (γ, μ, L, ψ_j, β_j, Λ_j, G, C_t), and no fitted parameters are introduced to calibrate the bound to the experiments. The 29.9%-71.6% speedup claim in the abstract is supported by direct time-to-accuracy measurements (Section VI-B, Fig. 3: 1077s vs 1536s vs 3794s), not by evaluating the convergence bound. The optimization in Section V-A does replace training time with the worst-case bound L(1+τ̂max) log_B A (Eq. 40a), which is a proxy whose tightness is not measured; this is a legitimate correctness or optimality caveat, but it is not circularity because the bound is not fitted to the data and the grouping algorithm's optimality is stated relative to P2, not claimed to be derived from empirical time-to-accuracy. Self-citations [25] and [37] appear only in related work and as support for a bias statement, and they are not load-bearing. Hence no circular step is identifiable.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central claim rests on the standard convex-FL assumption bundle (L-smoothness, mu-strong convexity, bounded gradients, bounded model norm), the delayed-recursion lemma cited from [33] rather than proven, a perfect-CSI channel-inversion transmit scheme, and a pair of timing models used both in the problem formulation and in the simulator. On the experimental side, the reported speedups depend on free parameters the paper never fixes or reports: the heterogeneity factors kappa_i (uniform in [1,10]), the model-norm bound W_t, the learning rate gamma, the participation frequencies psi_j, and the tuning of xi to 0.3. No invented physical entities are introduced.

free parameters (5)
  • xi (grouping similarity threshold) = 0.3
    Controls the allowed intra-group spread of local training times (constraint 36d). Tuned by sweeping [0,1] on CNN/MNIST (Fig. 8), then used for the headline benchmark numbers.
  • W_t (model norm bound) = not reported
    Assumption 4's bound enters both the convergence term C_t (Eq. 30) and the energy-feasibility limit on sigma_t (Eq. 46). The paper never states how W_t is obtained in the simulations.
  • gamma (learning rate) = not reported
    Theorem 1 requires gamma in (1/(2L), 1/L). The experimental section never states the learning rate, so the assumption cannot be checked.
  • kappa_i (heterogeneity scaling factors) = uniform random in [1,10]
    The simulated local training time is li = kappa_i * lhat_i, and every reported training-time number is computed on this synthetic timeline; the 29.9%-71.6% speedup range is conditional on this choice.
  • psi_j (group participation frequency) = not defined operationally
    The convergence bound in Theorem 1 weights groups by psi_j, but the paper never states how psi_j is measured or estimated in the algorithm or the experiments.
assumptions (7)
  • domain assumption Every local loss fi is L-smooth and mu-strongly convex (Assumptions 1-2, Section IV-A).
    Standard in the FL convergence literature; the paper acknowledges it only holds for convex smooth models. No reported experiment (MLP labeled LR, CNN, VGG-16) satisfies it.
  • domain assumption Gradients are uniformly bounded: ||G^k(w)||^2 <= G^2 (Assumption 3).
    Standard bounded-gradient assumption used to bound the EMD term in Eq. (24).
  • ad hoc to paper Local model norms satisfy ||w_i^t||^2 <= W_t^2 for all workers and rounds (Assumption 4).
    Asserted to hold for the whole trajectory, justified only by 'the model size is relatively stable'; W_t is used in the energy constraint Eq. (46) and in C_t Eq. (30), but the paper never measures or sets W_t in experiments.
  • standard math Lemma 1: if Q(t) <= xQ(t-1) + yQ(l_t) + z with x+y<1 then Q(t) <= rho^t Q(0) + delta (Section IV-B).
    The contraction lemma is not proved in the paper; it is cited to [33] with 'can be proved by mathematical induction'. The entire theorem's rate depends on it.
  • ad hoc to paper Staleness estimate: tau_max = L_max * sum_j 1/L_j (Eq. 39).
    Heuristic replacing the true max staleness by slowest group completion time times total update frequency; stated as 'obvious' with no derivation or validation.
  • domain assumption Perfect channel state information and channel-inversion transmit power pi^t = di * sigma_t / hi^t (Eq. 6).
    Requires hi^t known at each worker and constant within a round; no CSI error model; the experiments never simulate fading at all.
  • domain assumption Simplified timing models: aggregation time L_u = q/(R Ls) (Eq. 33, from [10]) and group completion L_j = max_i l_i + L_u (Eq. 34).
    These timing models define both the optimization objective P1-P4 and the experimental timeline, so the reported wall-clock speedups inherit their validity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Air-FedGA: A Grouping Asynchronous Federated Learning Mechanism Exploiting Over-the-air Computation." pith.science (2026). https://pith.science/paper/XYEY7UJ7

@misc{pith2026250705704,
  author       = {Pith},
  title        = {Pith review of: Air-FedGA: A Grouping Asynchronous Federated Learning Mechanism Exploiting Over-the-air Computation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYEY7UJ7}},
  note         = {Machine review of arXiv:2507.05704}
}
read the original abstract

Federated learning (FL) is a new paradigm to train AI models over distributed edge devices (i.e., workers) using their local data, while confronting various challenges including communication resource constraints, edge heterogeneity and data Non-IID. Over-the-air computation (AirComp) is a promising technique to achieve efficient utilization of communication resource for model aggregation by leveraging the superposition property of a wireless multiple access channel (MAC). However, AirComp requires strict synchronization among edge devices, which is hard to achieve in heterogeneous scenarios. In this paper, we propose an AirComp-based grouping asynchronous federated learning mechanism (Air-FedGA), which combines the advantages of AirComp and asynchronous FL to address the communication and heterogeneity challenges. Specifically, Air-FedGA organizes workers into groups and performs over-the-air aggregation within each group, while groups asynchronously communicate with the parameter server to update the global model. In this way, Air-FedGA accelerates the FL model training by over-the-air aggregation, while relaxing the synchronization requirement of this aggregation technology. We theoretically prove the convergence of Air-FedGA. We formulate a training time minimization problem for Air-FedGA and propose the power control and worker grouping algorithm to solve it, which jointly optimizes the power scaling factors at edge devices, the denoising factors at the parameter server, as well as the worker grouping strategy. We conduct experiments on classical models and datasets, and the results demonstrate that our proposed mechanism and algorithm can speed up FL model training by 29.9%-71.6% compared with the state-of-the-art solutions.

Figures

Figures reproduced from arXiv: 2507.05704 by the authors.

Figure 1
Figure 1. The Architecture of Air-FedGA. access channels (MAC) to sum these vectors over-the-air [14]. The implementation of AirComp has a bifacial impact on FL. On one hand, due to the efficient spectrum utilization of AirComp-based aggregation, it is expected to significantly reduce the transmission delay compared to the OMA-based aggregation which decouples communication and computa￾tion. On the other hand, strict synchron… view at source ↗
Figure 2
Figure 2. The workflow of Air-FedGA. v1 and v2 performs over-the-air aggregation simultaneously at round 1, i.e., w1 = (1−α1−α2)w0+ α1σ1w1 1+α2σ1w2 √ 1 η1 + z1 D √η1 . Since workers v1 and v2 receive the global model w0 at round 1, the staleness τ1 = 0. For another instance, workers v5 and v6 performs over-the-air aggregation simultaneously at round 4, i.e., w4 = (1 − α5 − α6)w3 + α5σ4w5 4+α6σ4w6 √ 4 η4 + z4 D √η4 . Since the… view at source ↗
Figure 3
Figure 3. Loss/Accuracy vs. Time (LR on MNIST). Left: Loss; Right: Accuracy. parameter server. This adjusted time li is then used to calculate its training completion time and recorded in a dynamically maintained list, L. By monitoring the training completion times of all workers recorded in L, we determine when each group performs over-the-air aggregations. Additionally, we set the bandwidth B = 1MHz, the noisy variance σ 2 … view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Loss/Accuracy vs. Time (VGG-16 on ImageNet-100). [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Grouping of workers with different local training ti [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Training time under different values of ξ. TABLE III: The impact of the grouping methods on EMD Methods Original TiFL Air-FedGA EMD 1.8 0.69 0.21 4) Energy Consumption for Over-the-air Aggregation [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: Training Time vs. Number of Workers. Left: Single Round; Right: Total. VII. CONCLUSION In this paper, we have proposed an AirComp-based grouping asynchronous federated learning mechanism (Air-FedGA) to address the challenges of communication resource constraint, heter…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

    Towar ds an intelligent edge: Wireless communication meets machine learning,

    G. Zhu, D. Liu, Y . Du, C. Y ou, J. Zhang, and K. Huang, “Towar ds an intelligent edge: Wireless communication meets machine learning,” arXiv preprint arXiv:1809.00343 , 2018

    arXiv 2018

  2. [2]

    In- edge ai: Intelligentizing mobile edge computing, caching and co mmunication by federated learning,

    X. Wang, Y . Han, C. Wang, Q. Zhao, X. Chen, and M. Chen, “In- edge ai: Intelligentizing mobile edge computing, caching and co mmunication by federated learning,” IEEE Network, vol. 33, no. 5, pp. 156–165, 2019

    work page 2019

  3. [3]

    Federated optimization: Distributed machine learning for on-device intelligence,

    J. Kone ˇcn`y, H. B. McMahan, D. Ramage, and P . Richtárik, “Federated optimization: Distributed machine learning for on-device intelligence,” arXiv preprint arXiv:1610.02527 , 2016

    arXiv 2016

  4. [4]

    Scaling distributed ma chine learning with the parameter server,

    M. Li, D. G. Andersen, J. W. Park, A. J. Smola, A. Ahmed, V . J osifovski, J. Long, E. J. Shekita, and B.-Y . Su, “Scaling distributed ma chine learning with the parameter server,” in 11th {USENIX} Symposium on Operating Systems Design and Implementation ( {OSDI} 14), 2014, pp. 583–598

    work page 2014

  5. [5]

    Federated learning over wireless networks: Optimization model design and analysis,

    N. H. Tran, W. Bao, A. Zomaya, M. N. H. Nguyen, and C. S. Hong , “Federated learning over wireless networks: Optimization model design and analysis,” in IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, 2019, pp. 1387–1395

    work page 2019

  6. [6]

    Energy-efficient federated edge learnin g with joint communication and computation design,

    X. Mo and J. Xu, “Energy-efficient federated edge learnin g with joint communication and computation design,” Journal of Communications and Information Networks , vol. 6, no. 2, pp. 110–124, 2021

    work page 2021

  7. [7]

    Cost-e ffective federated learning in mobile edge networks,

    B. Luo, X. Li, S. Wang, J. Huang, and L. Tassiulas, “Cost-e ffective federated learning in mobile edge networks,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 12, pp. 3606–3621, 2021

    work page 2021

  8. [8]

    Convergence time optimiza- tion for federated learning over wireless networks,

    M. Chen, H. V . Poor, W. Saad, and S. Cui, “Convergence time optimiza- tion for federated learning over wireless networks,” IEEE Transactions on Wireless Communications , vol. 20, no. 4, pp. 2457–2471, 2021

    work page 2021

Show all 40 references
  1. [9]

    Energy efficient federated learning over wireless communication n etworks,

    Z. Y ang, M. Chen, W. Saad, C. S. Hong, and M. Shikh-Bahaei, “Energy efficient federated learning over wireless communication n etworks,” IEEE Transactions on Wireless Communications , vol. 20, no. 3, pp. 1935–1949, 2021

    1935

  2. [10]

    Broadband analog aggrega tion for low-latency federated edge learning,

    G. Zhu, Y . Wang, and K. Huang, “Broadband analog aggrega tion for low-latency federated edge learning,” IEEE Transactions on Wireless Communications, vol. 19, no. 1, pp. 491–506, 2020

    2020

  3. [11]

    Communication-efficient learning of deep networks from de centralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and A. y. B. A rcas, “Communication-efficient learning of deep networks from de centralized data,” AISTATS, pp. 1273–1282, 2017

    2017

  4. [12]

    Joint optimization of com- munications and federated learning over the air,

    X. Fan, Y . Wang, Y . Huo, and Z. Tian, “Joint optimization of com- munications and federated learning over the air,” IEEE Transactions on Wireless Communications, vol. 21, no. 6, pp. 4434–4449, 2022

    2022

  5. [13]

    Computation over multiple-ac cess channels,

    B. Nazer and M. Gastpar, “Computation over multiple-ac cess channels,” IEEE Transactions on Information Theory , vol. 53, no. 10, pp. 3498– 3516, 2007

    2007

  6. [14]

    Optimized powe r control design for over-the-air federated edge learning,

    X. Cao, G. Zhu, J. Xu, Z. Wang, and S. Cui, “Optimized powe r control design for over-the-air federated edge learning,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 1, pp. 342–358, 2021

    2021

  7. [15]

    One-bit over-the -air aggre- gation for communication-efficient federated edge learnin g: Design and convergence analysis,

    G. Zhu, Y . Du, D. Gündüz, and K. Huang, “One-bit over-the -air aggre- gation for communication-efficient federated edge learnin g: Design and convergence analysis,” IEEE Transactions on Wireless Communications , vol. 20, no. 3, pp. 2120–2135, 2020

    2020

  8. [16]

    Federated learni ng via over- the-air computation,

    K. Y ang, T. Jiang, Y . Shi, and Z. Ding, “Federated learni ng via over- the-air computation,” IEEE Transactions on Wireless Communications , vol. 19, no. 3, pp. 2022–2035, 2020

    2022

  9. [17]

    Machine learning at the wirel ess edge: Distributed stochastic gradient descent over-the-air,

    M. M. Amiri and D. Gündüz, “Machine learning at the wirel ess edge: Distributed stochastic gradient descent over-the-air,” IEEE Transactions on Signal Processing , vol. 68, pp. 2155–2169, 2020

    2020

  10. [18]

    Transmission power con trol for over- the-air federated averaging at network edge,

    X. Cao, G. Zhu, J. Xu, and S. Cui, “Transmission power con trol for over- the-air federated averaging at network edge,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 5, pp. 1571–1586, 2022

    2022

  11. [19]

    Federated learning over wire less fading channels,

    M. M. Amiri and D. Gündüz, “Federated learning over wire less fading channels,” IEEE Transactions on Wireless Communications , vol. 19, no. 5, pp. 3546–3557, 2020

    2020

  12. [20]

    SA FA: a semi-asynchronous protocol for fast federated learning w ith low overhead,

    W. Wu, L. He, W. Lin, R. Mao, C. Maple, and S. A. Jarvis, “SA FA: a semi-asynchronous protocol for fast federated learning w ith low overhead,” IEEE Transactions on Computers , 2020

    2020

  13. [21]

    Asynchronous federate d optimization,

    C. Xie, S. Koyejo, and I. Gupta, “Asynchronous federate d optimization,” arXiv preprint arXiv:1903.03934 , 2019

    1903 arXiv

  14. [22]

    Asynchro nous online federated learning for edge devices with non-iid data,

    Y . Chen, Y . Ning, M. Slawski, and H. Rangwala, “Asynchro nous online federated learning for edge devices with non-iid data,” in 2020 IEEE International Conference on Big Data (Big Data) . IEEE, 2020, pp. 15–24

    2020

  15. [23]

    Asynchronous stochastic gradient descent with delay comp ensation,

    S. Zheng, Q. Meng, T. Wang, W. Chen, N. Y u, Z.-M. Ma, and T. -Y . Liu, “Asynchronous stochastic gradient descent with delay comp ensation,” in International Conference on Machine Learning . PMLR, 2017, pp. 4120–4129

    2017

  16. [24]

    Client selection with staleness compensation in asynchronous federated learning,

    H. Zhu, J. Kuang, M. Y ang, and H. Qian, “Client selection with staleness compensation in asynchronous federated learning,” IEEE Transactions on V ehicular Technology, vol. 72, no. 3, pp. 4124–4129, 2022

    2022

  17. [25]

    FedS A: A semi-asynchronous federated learning mechanism in heter ogeneous edge computing,

    Q. Ma, Y . Xu, H. Xu, Z. Jiang, L. Huang, and H. Huang, “FedS A: A semi-asynchronous federated learning mechanism in heter ogeneous edge computing,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 12, pp. 3654–3672, 2021

    2021

  18. [26]

    Tifl: A tier-based federated learning system,

    Z. Chai, A. Ali, S. Zawad, S. Truex, A. Anwar, N. Baracald o, Y . Zhou, H. Ludwig, F. Y an, and Y . Cheng, “Tifl: A tier-based federated learning system,” in Proceedings of the 29th international symposium on high- performance parallel and distributed computing , 2020, pp. 125–136

    2020

  19. [27]

    Adaptive asynchronous federated learning in resource-co nstrained edge computing,

    J. Liu, H. Xu, L. Wang, Y . Xu, C. Qian, J. Huang, and H. Huan g, “Adaptive asynchronous federated learning in resource-co nstrained edge computing,” IEEE Transactions on Mobile Computing , 2021

    2021

  20. [28]

    Decen tral- ized federated learning with adaptive configuration for het erogeneous participants,

    Y . Liao, Y . Xu, H. Xu, L. Wang, C. Qian, and C. Qiao, “Decen tral- ized federated learning with adaptive configuration for het erogeneous participants,” IEEE Transactions on Mobile Computing , 2023

    2023

  21. [29]

    Gradient statistics aware power co ntrol for over- the-air federated learning,

    N. Zhang and M. Tao, “Gradient statistics aware power co ntrol for over- the-air federated learning,” IEEE Transactions on Wireless Communica- tions, vol. 20, no. 8, pp. 5115–5128, 2021

    2021

  22. [30]

    On the pairing of the soft max activation and cross-entropy penalty functions and the der ivation of the softmax activation function,

    R. A. Dunne and N. A. Campbell, “On the pairing of the soft max activation and cross-entropy penalty functions and the der ivation of the softmax activation function,” in Proc. 8th Aust. Conf. on the Neural Networks, Melbourne , vol. 181. Citeseer, 1997, p. 185

    1997

  23. [31]

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

    Y . Sun, S. Zhou, Z. Niu, and D. Gündüz, “Dynamic scheduli ng for over- the-air federated edge learning with energy constraints,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 1, pp. 227–242, 2022

    2022

  24. [32]

    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

  25. [33]

    A de layed proximal gradient method with linear convergence rate,

    H. R. Feyzmahdavian, A. Aytekin, and M. Johansson, “A de layed proximal gradient method with linear convergence rate,” in 2014 IEEE International W orkshop on Machine Learning for Signal Proc essing (MLSP), 2014, pp. 1–6

    2014

  26. [34]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, P . Haffner et al. , “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  27. [35]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton et al. , Learning multiple layers of features from tiny images . Citeseer, 2009

    2009

  28. [36]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S . Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” International journal of computer vision, vol. 115, pp. 211–252, 2015

    2015

  29. [37]

    FedUC: A Unified C luster- ing Approach for Hierarchical Federated Learning,

    Q. Ma, Y . Xu, H. Xu, J. Liu, and L. Huang, “FedUC: A Unified C luster- ing Approach for Hierarchical Federated Learning,” IEEE Transactions on Mobile Computing , no. 01, pp. 1–18, 2024

    2024

  30. [38]

    D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant, Applied logistic regression. John Wiley & Sons, 2013, vol. 398

    2013

  31. [39]

    Shalev-Shwartz and S

    S. Shalev-Shwartz and S. Ben-David, Understanding machine learning: From theory to algorithms . Cambridge university press, 2014

    2014

  32. [40]

    V ery deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “V ery deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    2014 arXiv

Pith tools

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