Pith. sign in

REVIEW 4 major objections 3 minor 28 references

Kuramoto-FedAvg: Using Synchronization Dynamics to Improve Federated Learning Optimization under Statistical Heterogeneity

T0 review · 4 major / 3 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 Reweighting each client's update by the sine of its phase angle relative to the global mean provably shrinks the drift term in the FedAvg bound, giving Kuramoto-FedAvg a strict round-count advantage on non-IID data.

desk verdict A simple direction-based reweighting with an unproven theory and shaky experiments; the Kuramoto framing doesn't rescue the core claims. read the letter →

arxiv 2505.19605 v1 pith:F623JQ2R submitted 2025-05-26 cs.LG

Aggrey Muhebwa , Khotso Selialia , Fatima Anwar , Khalid K. Osman This is my paper · ORCID
classification cs.LG
keywords federatedlearningclientdriftstatisticalheterogeneitynon-IIDdataKuramotomodelsynchronizationgradientalignmentconvergenceanalysis
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 tries to establish that client drift in federated learning—the divergence of local models from the global trajectory when data are non-IID—can be attacked at the aggregation step rather than by constraining local training. It proposes Kuramoto-FedAvg, which treats each client's model update as an oscillator and assigns it a phase equal to the angle between that update and the average update. Aggregation then weights each client by the sine of its phase difference from the mean phase, amplifying aligned updates and suppressing misaligned ones. The paper claims this reweighting strictly shrinks the gradient-drift term $\Gamma(t)$ in the FedAvg convergence bound, and therefore proves a communication-round advantage: $T_{\mathrm{Kuramoto\text{-}FedAvg}} < T_{\mathrm{FedAvg}}$ for any target error under heterogeneous data. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 report faster convergence and accuracy gains, with the largest margins under high heterogeneity.

What carries the argument

The load-bearing object is the phase-alignment weight $\rho_k(t) = \sin(\bar{\theta}(t)-\theta_k(t)) / \sum_j \sin(\bar{\theta}(t)-\theta_j(t))$, where $\theta_k(t)$ is the angle in parameter space between client $k$'s local update $\Delta w_k^t$ and the mean update $\bar{\Delta}^t$. This weight replaces the data-proportional weight $p_k$ in the FedAvg aggregation step $w^{t+1} = w^t + \kappa_t \sum_k \rho_k(t)\Delta w_k^t$. The argument runs through the drift term $\Gamma(t) := \sum_k p_k\|\nabla F_k(w^t)\|^2 - \|\nabla F(w^t)\|^2$ in the FedAvg convergence bound; the paper redefines it with $\rho_k(t)^2$ in place of $p_k$ and asserts that the sine weighting, which suppresses large phase differences, makes $\Gamma_{\mathrm{Kuramoto}}(t) < \Gamma(t)$. The mechanism is an analogy to the Kuramoto model of coupled oscillators, where phases synchronize when coupling strength exceeds a threshold; here the coupling acts once per round at the server through reweighting rather than through continuous phase dynamics.

What would settle it

One decisive check is to compute, on a non-IID benchmark with known local gradients, the ratio $\Gamma_{\mathrm{Kuramoto}}(t)/\Gamma(t)$ using the paper's weight formula $\rho_k(t)$ across rounds: if the ratio is ever $\ge 1$, or if $\sum_j \sin(\bar{\theta}(t)-\theta_j(t))$ vanishes or produces negative weights that violate the descent inequality, the claimed strict bound $T_{\mathrm{Kuramoto\text{-}FedAvg}} < T_{\mathrm{FedAvg}}$ fails in that setting. A simpler observational test: on a public non-IID federated benchmark, run FedAvg and Kuramoto-FedAvg with identical hyperparameters and record the round count to reach a fixed test accuracy; the paper's theorem predicts a strict round-count advantage at every target accuracy, which is contradicted if FedAvg ever matches or beats it.

Watch

Extended reading notes

Core claim

The paper's central claim is that the aggregation rule itself, rather than local regularization or control variates, can neutralize client drift. At each round the server computes each client's phase $\theta_k(t) = \arccos\big(\langle \Delta w_k^t, \bar{\Delta}^t\rangle / (\|\Delta w_k^t\|\,\|\bar{\Delta}^t\|)\big)$, the angle between that client's local update and the weighted average update, and reweights updates by $\rho_k(t) = \sin(\bar{\theta}(t)-\theta_k(t)) / \sum_j \sin(\bar{\theta}(t)-\theta_j(t))$. Under $L$-smoothness, unbiased stochastic gradients, and $\mu$-strong convexity, the FedAvg bound (Theorem 1) reads $\mathbb{E}[F(w^{t+1})] \le F(w^t) - \eta_t\|\nabla F(w^t)\|^2 + (L\eta_t^2/2)\Gamma(t) + (L\eta_t^2/2)\sigma^2$. Theorem 2 asserts that replacing the fixed client weights $p_k$ by $\rho_k(t)^2$ inside the drift term gives $\Gamma_{\mathrm{Kuramoto}}(t) < \Gamma(t)$ at every round, hence a strictly tighter per-round bound and strictly fewer communication rounds to reach any target error $\varepsilon$: $T_{\mathrm{Kuramoto\text{-}FedAvg}}(\varepsilon,\Gamma) < T_{\mathrm{FedAvg}}(\varepsilon,\Gamma)$. The claim is that phase alignment alone guarantees this advantage, with no change to local training.

Load-bearing premise

The load-bearing premise is that the angle between one round's noisy local update and the global average update reliably reveals which clients' true gradients point toward the global direction, and that reweighting by the sine of that phase difference shrinks the drift term without inflating stochastic noise or producing degenerate weights.

Editorial extensions

If this is right

  • If Theorem 2 holds, federated training on non-IID data reaches a target accuracy in strictly fewer communication rounds than FedAvg, and communication rounds are the dominant cost in federated systems.
  • The advantage is bought entirely at the server: clients keep their local training unchanged, and the only added computation is one cosine similarity per client per round, with no extra communication.
  • Because the drift term $\Gamma(t)$ grows with statistical heterogeneity, the bound predicts that the round-count advantage widens exactly when data are most skewed, matching the paper's largest empirical gains under strong label sharding.
  • The bound is stated per communication round with round-dependent step size $\eta_t$, so the claimed advantage applies to each round's descent and accumulates over the whole training trajectory.

Reading between the lines

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

  • A deployment detail the paper leaves implicit is a rule for handling negative or vanishing sine weights: the normalizing sum $\sum_j \sin(\bar{\theta}(t)-\theta_j(t))$ can be zero or negative when updates straddle the mean phase, so a practical implementation needs a clamping or sign convention.
  • Because the phase is computed from the same noisy updates being reweighted, a natural testable variant computes the phase from a second independent mini-batch; if the round advantage survives, the mechanism is driven by genuine gradient alignment rather than by noise filtering.
  • The aggregation-only design means the synchronization weighting could be stacked on top of local-correction methods such as proximal terms or control variates; the paper does not run this combination, but its bound treats local training as unchanged.
  • The drift-reduction claim rests on an analogy to phase synchronization; supplying an explicit algebraic comparison of $\Gamma_{\mathrm{Kuramoto}}(t)$ with $\Gamma(t)$ under the weight formula would turn that analogy into a directly checkable inequality.
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 / 3 minor

Summary. The paper proposes Kuramoto-FedAvg, a federated aggregation scheme that reweights client updates by their cosine-similarity phase relative to the mean update, motivated by the Kuramoto model of coupled oscillators. The authors claim a theoretical result (Theorem 1) for FedAvg, a stronger result (Theorem 2) that Kuramoto-FedAvg has strictly fewer communication rounds than FedAvg under data heterogeneity, and empirical gains on MNIST, Fashion-MNIST, and CIFAR-10. The central contribution is the synchronization-based aggregation rule and the claimed proof that it reduces client drift.

Significance. If the theoretical claims were valid, a lightweight, aggregation-only method that provably reduces client drift would be a useful alternative to local-correction methods such as FedProx and SCAFFOLD. The paper does not, however, establish these claims: the proof of Theorem 1 is algebraically incorrect, the proof of Theorem 2 rests on an asserted inequality that does not follow from the stated assumptions, and the algorithm as written contains mechanical flaws (negative weights, division by zero, an undefined coupling parameter). The empirical section provides no code, seeds, or error bars, so the reported improvements cannot be checked. The paper therefore does not currently support its main claims.

major comments (4)
  1. [Section 4, Appendix A.1, Theorem 1] The derivation of the FedAvg bound is algebraically incorrect. After taking expectations, the proof obtains E||Σ p_k g_k||² = ||∇F(w)||² + Σ p_k² σ_k², but then it invokes the identity Σ p_k||∇F_k||² = ||∇F||² + Γ(t) to replace the ||∇F||² term by Γ(t). That identity concerns a different quantity, and the correct one-step bound would contain Lη²/2 (||∇F||² + σ²), not Lη²/2 Γ(t). The FedAvg benchmark bound used throughout the paper is therefore not established.
  2. [Section 4, Appendix A.2, Theorem 2] The claimed inequality Γ_Kuramoto(t) < Γ(t) is asserted, not derived. The weights ρ_k(t) from Eq. (3) are not nonnegative: sin(θ̄−θ_k) can be negative, and the denominator is zero when all phase differences are zero. Even when the weights are well defined, Σ ρ_k² ||∇F_k||² need not be smaller than Σ p_k ||∇F_k||², and the proof gives no argument for this. The proof also treats ρ_k as deterministic constants even though they are computed from the same stochastic gradients used in the update, and it reuses the FedAvg variance σ² in Eq. (5) although the reweighted update has variance Σ ρ_k² σ_k². The conclusion T_Kuramoto-FedAvg(ε,Γ) < T_FedAvg(ε,Γ) does not follow from the stated assumptions.
  3. [Section 3.2 and Algorithm 1] The algorithm is not the object of the theory. Algorithm 1 line 23 applies an additional factor κ_t to the phase-weighted sum, while Eq. (4) and Theorem 2 omit κ entirely, and κ_t is never defined; Section 5.1 introduces a different hyperparameter κ0. The phase in line 17 is undefined when any Δw_k or the mean update is zero, and the synchronization weights can be negative, so the server update need not be a convex combination of client updates. These mechanical flaws mean the implemented method differs from the analyzed update rule.
  4. [Section 5] The empirical evidence is not verifiable. The figures report no error bars or seeds, the number of independent runs is not stated, and no code is provided. The text uses inconsistent notation for heterogeneity (s in the figure captions versus λ in the main text), and Table 1 reports results at round 200 although Section 5.1 states training is conducted over 100 rounds; the 'no-sync' baseline in Table 1 is also not defined among the earlier baselines. These issues prevent checking the claimed 10–15% accuracy gains on CIFAR-10.
minor comments (3)
  1. [Throughout] There are numerous typographical errors, including 'Theoratical Analysis' in the Section 4 heading, 'tje' in Section 5.4, and 'Kuramoto-FedA vg' in the title line; these should be corrected.
  2. [Section 2] The citation for FedNova is given as [12], but reference [12] is FedBN; FedNova is reference [24]. Please correct the citation.
  3. [Section 5.1] The text refers to 'Kuramoto-FL' rather than 'Kuramoto-FedAvg' when describing the coupling mechanism, and the relationship between the coupling strength κ in Section 5.1 and the κ_t in Algorithm 1 is unclear.

Circularity Check

2 steps flagged · score 7.0 of 10

Theorem 2's faster-rounds claim is assumed, not derived: Γ_Kuramoto is defined using the sync weights, the strict decrease Γ_Kuramoto<Γ is asserted, and the round inequality is introduced with 'we can assume'.

  1. self definitional [Section 4, Theorem 2 (paragraph defining Γ_Kuramoto); proof deferred to Appendix A.2]
    "The client drift is thus represented as: Γ(t) := Σ_{k=1}^N ρ_k(t)^2 ||∇F_k(w^t)||^2 − ||∇F(w^t)||^2. Based on the Kuramoto model, we know that clients with large phase differences(highly mis-aligned) receive lower weights. Then, we have: Γ_Kuramoto(t)<Γ(t)."

    The proof defines the Kuramoto drift Γ_Kuramoto directly using the algorithm's own sync weights ρ_k, then asserts Γ_Kuramoto(t)<Γ(t) without any computation of E||Σρ_k g_k||² for the actual reweighted update. Faster convergence is equivalent to the reweighted drift being smaller, so the theorem's conclusion is loaded into the definition; the 'then' step is an unproved assertion, not a derived inequality. Moreover, ρ_k can be negative or undefined when the denominator vanishes, and ρ_k depends on the same stochastic gradients as the update, so a correct bound would contain extra first-order and variance terms that are never analyzed.

  2. other [Section 4, Theorem 2 (last paragraph, after Eq. (5))]
    "Therefore, we can assume that for any target error ϵ > 0, the number of iterations (or communication rounds with the central server) required by Kuramoto-FedAvg, T_Kuramoto-FedAvg(ϵ,Γ), is strictly smaller than that of standard FedAvg. T_Kuramoto-FedAvg(ϵ,Γ)< T_FedAvg(ϵ,Γ)."

    The theorem's conclusion is introduced by 'we can assume' rather than proved. The only bridge from the descent bound to this statement is the asserted Γ_Kuramoto(t)<Γ(t), which itself is not established. The proof therefore assumes exactly the strict round-count inequality it is supposed to establish; no convergence-rate argument or round-count inversion is supplied.

full rationale

No load-bearing self-citations appear, so the circularity here is not of the self-citation kind. The central theoretical claim, however, is structurally circular in that its conclusion is assumed. In Theorem 2, the proof defines Γ_Kuramoto using the synchronization weights ρ_k and then asserts Γ_Kuramoto<Γ; that inequality is precisely what would be needed to show faster convergence, but it is never derived from the dynamics of the reweighted update. The statement is then completed by 'we can assume' the strict round-count inequality, which is the theorem's own target. The comparison also inherits an invalid FedAvg bound: in Appendix A.1, after correctly evaluating E||Σp_k g_k||² as ||∇F||²+σ², the proof converts the quadratic term into Lη²/2 Γ(t)+Lη²/2 σ² using the identity Σp_k||∇F_k||²=||∇F||²+Γ(t), a substitution that does not follow algebraically. Thus the claimed tighter convergence bound for Kuramoto-FedAvg is not derived from the assumptions; it is placed into the definition of Γ_Kuramoto and then assumed. The empirical evaluations are external benchmarks and are not circular, but they do not rescue the theoretical comparison. Overall, the central theorem reduces to an assumption of the result, giving a circularity score of 7.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The theory rests on unproven assumptions about the reliability of stochastic phases and the invariance of the variance term, plus a tuned synchronization strength; these carry the main burden for the claimed advantage.

free parameters (1)
  • synchronization strength κ0 (κ_t) = 0.005 (best on all datasets); also tested 0.1, 0.3, 1.0
    Tuned per dataset in Table 1; performance is sensitive to this value, with large values degrading accuracy.
assumptions (3)
  • standard math Local objective F_k is L-smooth and µ-strongly convex; stochastic gradients are unbiased with bounded variance.
    Assumptions 2 in Section 4; these are standard in FL convergence analysis but not verified for the CNN models used in experiments.
  • domain assumption The phase θ_k(t) from a single round of stochastic updates is a reliable proxy for the population gradient direction, and reweighting by it reduces the population drift Γ(t).
    Theorem 2 depends on this; no theoretical or empirical justification is given.
  • ad hoc to paper The reweighted stochastic variance Σ_k ρ_k(t)^2 σ_k^2 is bounded by the FedAvg variance σ^2, so Eq. (5) can reuse the same σ^2 term.
    Eq. (5) in Section 4 reuses σ^2 without proof; since ρ_k can be negative or large, this is not guaranteed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kuramoto-FedAvg: Using Synchronization Dynamics to Improve Federated Learning Optimization under Statistical Heterogeneity." pith.science (2026). https://pith.science/paper/F623JQ2R

@misc{pith2026250519605,
  author       = {Pith},
  title        = {Pith review of: Kuramoto-FedAvg: Using Synchronization Dynamics to Improve Federated Learning Optimization under Statistical Heterogeneity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F623JQ2R}},
  note         = {Machine review of arXiv:2505.19605}
}
read the original abstract

Federated learning on heterogeneous (non-IID) client data experiences slow convergence due to client drift. To address this challenge, we propose Kuramoto-FedAvg, a federated optimization algorithm that reframes the weight aggregation step as a synchronization problem inspired by the Kuramoto model of coupled oscillators. The server dynamically weighs each client's update based on its phase alignment with the global update, amplifying contributions that align with the global gradient direction while minimizing the impact of updates that are out of phase. We theoretically prove that this synchronization mechanism reduces client drift, providing a tighter convergence bound compared to the standard FedAvg under heterogeneous data distributions. Empirical validation supports our theoretical findings, showing that Kuramoto-FedAvg significantly accelerates convergence and improves accuracy across multiple benchmark datasets. Our work highlights the potential of coordination and synchronization-based strategies for managing gradient diversity and accelerating federated optimization in realistic non-IID settings.

Figures

Figures reproduced from arXiv: 2505.19605 by the authors.

Figure 1
Figure 1. Comparison of Kuramoto-FedAvg with standard FedAvg and SCAFFOLD under varying [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Mean test accuracy across varying degrees of statistical heterogeneity. While all methods [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of test accuracy for Kuramoto-FedAvg, FedAvg, and SCAFFOLD over several [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 18 canonical work pages

  1. [1]

    The kuramoto model: A simple paradigm for synchronization phenomena.Reviews of modern physics, 77(1):137–185, 2005

    Juan A Acebrón, Luis L Bonilla, Conrad J Pérez Vicente, Félix Ritort, and Renato Spigler. The kuramoto model: A simple paradigm for synchronization phenomena.Reviews of modern physics, 77(1):137–185, 2005

    work page 2005

  2. [2]

    Syn- chronization in complex networks.Physics reports, 469(3):93–153, 2008

    Alex Arenas, Albert Díaz-Guilera, Jurgen Kurths, Yamir Moreno, and Changsong Zhou. Syn- chronization in complex networks.Physics reports, 469(3):93–153, 2008

    work page 2008

  3. [3]

    Emnist: Extending mnist to handwritten letters

    Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emnist: Extending mnist to handwritten letters. In2017 international joint conference on neural networks (IJCNN), pages 2921–2926. IEEE, 2017

    2017

  4. [4]

    Kuramoto model with frequency- degree correlations on complex networks.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 87(3):032106, 2013

    BC Coutinho, A V Goltsev, SN Dorogovtsev, and JFF Mendes. Kuramoto model with frequency- degree correlations on complex networks.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 87(3):032106, 2013

    work page 2013

  5. [5]

    Amplitude expansions for instabilities in populations of globally-coupled oscillators.Journal of statistical physics, 74:1047–1084, 1994

    John David Crawford. Amplitude expansions for instabilities in populations of globally-coupled oscillators.Journal of statistical physics, 74:1047–1084, 1994

    work page 1994

  6. [6]

    Synchronization and transient stability in power networks and nonuniform kuramoto oscillators.SIAM Journal on Control and Optimization, 50(3):1616– 1642, 2012

    Florian Dorfler and Francesco Bullo. Synchronization and transient stability in power networks and nonuniform kuramoto oscillators.SIAM Journal on Control and Optimization, 50(3):1616– 1642, 2012

    work page 2012

  7. [7]

    Novel insights into lossless ac and dc power flow

    Florian Dörfler and Francesco Bullo. Novel insights into lossless ac and dc power flow. In2013 IEEE Power & Energy Society General Meeting, pages 1–5. IEEE, 2013

    work page 2013

  8. [8]

    An adaptive model for synchrony in the firefly pteroptyx malaccae.Journal of Mathematical Biology, 29(6):571–585, 1991

    Bard Ermentrout. An adaptive model for synchrony in the firefly pteroptyx malaccae.Journal of Mathematical Biology, 29(6):571–585, 1991

    work page 1991

Show all 28 references
  1. [9]

    Scaffold: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132–5143. PMLR, 2020

    2020

  2. [10]

    Self-entrainment of a population of coupled non-linear oscillators

    Yoshiki Kuramoto. Self-entrainment of a population of coupled non-linear oscillators. In International symposium on mathematical problems in theoretical physics: January 23–29, 1975, kyoto university, kyoto/Japan, pages 420–422. Springer, 1975

    1975

  3. [11]

    Federated optimization in heterogeneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020

    2020

  4. [12]

    Fedbn: Federated learning on non-iid features via local batch normalization.arXiv preprint arXiv:2102.07623, 2021

    Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization.arXiv preprint arXiv:2102.07623, 2021

    2021 arXiv

  5. [13]

    Synchronization in the random- field kuramoto model on complex networks.Physical Review E, 94(1):012308, 2016

    MA Lopes, EM Lopes, S Yoon, JFF Mendes, and A V Goltsev. Synchronization in the random- field kuramoto model on complex networks.Physical Review E, 94(1):012308, 2016

    2016

  6. [14]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. InArtificial intelligence and statistics, pages 1273–1282. PMLR, 2017

    2017

  7. [15]

    An introduction to convolutional neural networks.arXiv preprint arXiv:1511.08458, 2015

    Keiron O’shea and Ryan Nash. An introduction to convolutional neural networks.arXiv preprint arXiv:1511.08458, 2015

    2015 arXiv

  8. [16]

    Network dynamics of coupled oscillators and phase reduction techniques.Physics Reports, 819:1–105, 2019

    Bastian Pietras and Andreas Daffertshofer. Network dynamics of coupled oscillators and phase reduction techniques.Physics Reports, 819:1–105, 2019. 10

    2019

  9. [17]

    Do cifar-10 classifiers generalize to cifar-10?arXiv preprint arXiv:1806.00451, 2018

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do cifar-10 classifiers generalize to cifar-10?arXiv preprint arXiv:1806.00451, 2018

    2018 arXiv

  10. [18]

    The kuramoto model in complex networks.Physics Reports, 610:1–98, 2016

    Francisco A Rodrigues, Thomas K DM Peron, Peng Ji, and Jürgen Kurths. The kuramoto model in complex networks.Physics Reports, 610:1–98, 2016

    2016

  11. [19]

    Understanding federated learning from iid to non-iid dataset: An experimental study.arXiv preprint arXiv:2502.00182, 2025

    Jungwon Seo, Ferhat Ozgur Catak, and Chunming Rong. Understanding federated learning from iid to non-iid dataset: An experimental study.arXiv preprint arXiv:2502.00182, 2025

    2025 arXiv

  12. [20]

    Higher order interactions in complex networks of phase oscillators promote abrupt synchronization switching.Communications Physics, 3(1):218, 2020

    Per Sebastian Skardal and Alex Arenas. Higher order interactions in complex networks of phase oscillators promote abrupt synchronization switching.Communications Physics, 3(1):218, 2020

    2020

  13. [21]

    Sync: The emerging science of spontaneous order

    Steven Strogatz. Sync: The emerging science of spontaneous order. 2004

    2004

  14. [22]

    From kuramoto to crawford: exploring the onset of synchronization in populations of coupled oscillators.Physica D: Nonlinear Phenomena, 143(1-4):1–20, 2000

    Steven H Strogatz. From kuramoto to crawford: exploring the onset of synchronization in populations of coupled oscillators.Physica D: Nonlinear Phenomena, 143(1-4):1–20, 2000

    2000

  15. [23]

    Srikanth Tammina. Transfer learning using vgg-16 with deep convolutional neural network for classifying images.International Journal of Scientific and Research Publications (IJSRP), 9(10):143–150, 2019

    2019

  16. [24]

    Tackling the objective inconsistency problem in heterogeneous federated optimization.Advances in neural information processing systems, 33:7611–7623, 2020

    Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization.Advances in neural information processing systems, 33:7611–7623, 2020

    2020

  17. [25]

    Synchronization transitions in a disordered josephson series array.Physical review letters, 76(3):404, 1996

    Kurt Wiesenfeld, Pere Colet, and Steven H Strogatz. Synchronization transitions in a disordered josephson series array.Physical review letters, 76(3):404, 1996

    1996

  18. [26]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

    2017 arXiv

  19. [27]

    Federated machine learning: Concept and applications.ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1–19, 2019

    Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications.ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1–19, 2019

    2019

  20. [28]

    Federated learning with non-iid data.arXiv preprint arXiv:1806.00582, 2018

    Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data.arXiv preprint arXiv:1806.00582, 2018. A Proofs A.1 Proof of Theorem 1 Proof.By theL-smoothness ofF(w), we have: F(w t+1)≤F(w t) +∇F(w t)⊤(wt+1 −w t) + L 2 ||wt+...

    2018 arXiv

Pith tools

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