Pith. sign in

REVIEW 4 major objections 6 minor 20 references

GLow -- A Novel, Flower-Based Simulated Gossip Learning Strategy

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read GLow builds a Flower-based gossip learning simulator that matches federated accuracy in tests.

desk verdict Useful Flower-based decentralized learning simulator, but the 'performs similarly' claim overreaches and the synchronous round-robin design means it is not yet an asynchronous gossip simulator. read the letter →

arxiv 2501.10463 v1 pith:EWCPWRKE submitted 2025-01-15 cs.LG cs.AIcs.DC

Aitor Belenguer , Jose A. Pascual , Javier Navaridas This is my paper · ORCID
classification cs.LGcs.AIcs.DC
keywords gossiplearningfederateddecentralizedFlowerframeworknetworktopologysimulationMNISTCIFAR10
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 introduces GLow, a gossip learning strategy built on the Flower framework, which simulates fully decentralized training without a central server. The authors claim GLow lets researchers test convergence, scalability, and topology choices for IoT-style agent networks before physical deployment. Across MNIST and CIFAR10, with 8 and 16 connected agents plus special control agents, GLow reaches accuracies over 0.98 and 0.75, and performs similarly to centralized and federated learning in most experiments. The point is that fully decentralized gossip learning can be studied and validated in simulation before dealing with the complexity and faults of real devices.

What carries the argument

The central mechanism is a round-robin head-aggregation loop: each iteration selects one agent as head, that agent trains locally for a fixed number of epochs, pulls current weights from its neighbors, and replaces its model with a weighted average of those weights. This replaces the server-driven parameter spread of a centralized federated strategy with a neighbor-request step, reducing the number of messages. A topology generator creates graphs from disconnected to fully connected, and special agents (empty, disconnected, empty-and-disconnected) act as controls to verify that information propagates through the network.

What would settle it

Run GLow's double-ring 8+2 agent CIFAR10 configuration on physical devices or in an emulator with random message drops, delayed arrivals, or a malicious agent sending corrupted weights; if the connected agents' average accuracy drops substantially below the disconnected self-learning baseline, the simulator's convergence predictions do not transfer to real conditions.

Watch

Extended reading notes

Core claim

The central claim is that a round-robin head aggregation scheme, implemented as a custom Flower strategy, reproduces the accuracy and convergence behavior of centralized and federated approaches in fully decentralized settings. In each communication round the designated head agent trains locally for E epochs, pulls current weights from its neighbors, and sets its model to a weighted average of those weights. The paper shows that with sufficient connectivity, roughly a double-ring topology, agents without any local data converge to the same accuracy as data-holding neighbors, while disconnected agents learn only from their own data and perform worse. The authors conclude that GLow is competitive with vanilla FedAvg and centralized learning, despite using only a subset of the data and including data-less agents in the network.

Load-bearing premise

The simulation assumes communication is synchronous, reliable, and free of Byzantine or packet-loss behavior, so accuracy measured in GLow may not hold in real decentralized deployments where these failures occur.

Editorial extensions

If this is right

  • Researchers can use GLow to estimate the convergence and accuracy of custom gossip topologies before physical deployment, under the simulator's reliable-communication assumptions.
  • Adding connectivity beyond a double ring yields little accuracy gain but can speed up early convergence, suggesting a sparse-connectivity design point for decentralized systems.
  • Including data-less agents exposes whether information truly propagates through the network; with sufficient connectivity these agents converge to neighbor-level accuracy.
  • In the 8+2 agent CIFAR10 scenario GLow reaches 0.754 accuracy versus 0.791 for federated averaging, a small gap given the removed server bottleneck; scaling to 16+4 agents lowers CIFAR10 accuracy to 0.683, attributed to data splitting and disconnected agents.
  • GLow outperforms self-learning in every connected topology tested, showing that decentralized parameter averaging provides a measurable benefit over isolated training.

Reading between the lines

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

  • Because GLow's simulation assumes synchronous rounds, reliable message delivery, and no Byzantine or packet-loss behavior, the strong accuracy results may not transfer to real deployments; extending the simulator to inject message drops, delays, or malicious agents would test this directly.
  • The saturation of accuracy beyond a double-ring topology suggests a general principle: sparse connectivity may suffice for gossip-style decentralized learning, which could guide communication budget choices in real systems.
  • The round-robin scheduling scheme can be seen as a deterministic traversal of the agent graph; comparing it with random head selection or multiple simultaneous heads would isolate the effect of scheduling on convergence and fairness.
  • The special-agent methodology (empty and disconnected agents as controls) could be reused in other decentralized learning simulators as a lightweight way to diagnose whether learned information actually propagates or merely reflects local data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents GLow, a simulation-oriented Gossip Learning (GL) strategy built on the Flower framework. GLow designates one agent as a head in each iteration via a round-robin schedule (Algorithm 1); the head trains locally and aggregates the current weights of its neighbors. A topology generator and special-agent configurations (disconnected, empty, or both) are introduced. The system is evaluated on MNIST and CIFAR10 with 8+2 and 16+4 agents and compared against centralized learning (CNL) and FedAvg (FL). The authors report that GLow achieves competitive accuracy in the 8+2 scenario and in the MNIST 16+4 scenario, but acknowledge lower accuracy for CIFAR10 with 16+4 agents.

Significance. If the claims are validated, a flexible, open-source, Flower-based GL simulator would be a useful community tool for studying decentralized learning before deployment. The authors provide a public repository and visualize per-agent behavior, which is a strength. However, the paper's central contribution is weakened by a mismatch between the simulated scheduling and the asynchronous gossip model introduced in the paper, by an overstatement in the abstract relative to Table II, and by a communication-round comparison that is not controlled across systems. The special-agent methodology is a useful experimental idea, though it is used more as a sanity check than as a substantive comparison.

major comments (4)
  1. [Section II.C and Algorithm 1] The paper defines Gossip Learning as a process where agents exchange parameters asynchronously without a server. However, Algorithm 1 implements a globally synchronized loop that selects k = iteration mod K as the single head, aggregates neighbor weights, and repeats. This is a synchronous, single-head, round-robin schedule, not an asynchronous peer-to-peer gossip process. Section VI confirms that the current version 'is designed to sequentially select an agent as head' and that a parallel version is only projected. Consequently, the abstract's claim that GLow 'will allow researchers to train and assess scalability and convergence of devices ... before making a physical deployment' is not supported for the asynchronous gossip systems the paper motivates. The authors should either implement asynchronous event-driven simulation or explicitly limit the claims to the synchronous variant.
  2. [Table II and Abstract] The abstract states that GLow 'performs similarly in terms of accuracy and convergence to its analogous Centralized and Federated approaches in all designed experiments.' Table II contradicts this: for CIFAR10 with 16+4 agents, GLow achieves 0.683, while FL achieves 0.778 and CNL achieves 0.789, a gap of roughly 0.1. The conclusions soften this to 'slightly lower accuracies,' but 0.1 is substantial. The abstract and conclusions should be revised to accurately reflect the 16+4 CIFAR10 result, or the experiment should be re-run with more communication rounds or different hyperparameters to close the gap.
  3. [Section IV.C and Table II] The communication-round comparison is not apples-to-apples. In FL, a communication round is a global round in which all clients train and the server aggregates. In GLow, Section IV.C defines a communication round as the number of times each agent is designated as head, so the total number of head aggregations is agents × rounds (e.g., 8 × 24 = 192 for MNIST 8+2). Thus GLow performs many more aggregation events than FL for the same reported 'Communication Rounds' value, while the local-epoch count per agent is similar. The convergence comparison should be controlled for the number of aggregation rounds or message exchanges, or the discrepancy should be explicitly stated and justified.
  4. [Section V] The empirical evaluation only compares GLow against centralized learning and vanilla FedAvg. Since the paper positions GLow as a GL simulator, it should also compare against existing decentralized baselines such as the gossip learning of Hegedus et al. [2], GossipFL [14], BrainTorrent [11], or Fedstellar [12]. Without such baselines, the claim that GLow is a useful GL strategy is not strongly established. At minimum, the authors should discuss how the synchronous round-robin schedule differs from these baselines and why the comparison to FedAvg alone is sufficient for the paper's claims.
minor comments (6)
  1. [Section V.C] The text contains a typo: 'Aditionally' should be 'Additionally'.
  2. [Figure 1] The subfigure label '(d) Topo 7 - F. connected' should be spelled out as 'Fully connected' for clarity.
  3. [Section IV.A] The model is attributed to 'Lecun et Al.' in the text, but the standard spelling is 'LeCun et al.' The paper should use consistent naming.
  4. [Section III] The text states that GLow maintains modularity by modifying only upper-level classes, but the footnote says a 'slight modification in the internal aggregate class of server' was made. This contradiction should be clarified.
  5. [Table II] The table reports 'Average Accuracy' for GLow only for topology 4 (double ring), while figures show results for many topologies. The selection criterion for reporting topology 4 should be stated explicitly in the table caption or text.
  6. [Section V.B] The expected random-guessing behavior of ED agents is used as a validation of the system, but this expectation is trivially true for agents with no data and no connections. The paper should not present this as a strong evidence of correctness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GLow's claims rest on direct empirical comparison against external CNL/FL baselines on public datasets; the special-agent expectations are construction-level control probes, not derived predictions.

full rationale

The paper's central claim is that a Flower-based gossip-style simulation strategy, GLow, achieves accuracy and convergence comparable to centralized and federated learning. This is supported by running Algorithm 1 on MNIST and CIFAR10 and comparing the resulting accuracies with externally defined baselines (centralized training and FedAvg); no parameter is fitted to the reported outcomes and no derived quantity is defined in terms of the data used to produce it. The only definitional-looking element is the special-agents methodology in Section IV.B: ED agents are defined as having no local data and no connectivity, and are 'expected' to behave as random guessers, with Section V.C noting this expectation is fulfilled. That is a control/sanity-check mechanism rather than a load-bearing empirical prediction, and the paper's convergence and accuracy conclusions do not depend on it. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The skeptical concerns about the synchronous round-robin head selection versus the paper's asynchronous gossip motivation, and about the 16+4 CIFAR10 accuracy gap, are correctness and scope limitations, not circular derivation.

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

The paper introduces no new theoretical entities. The free parameters are experimental design choices, not fitted constants. The key axioms are the modeling assumption that synchronous round-robin aggregation approximates true asynchronous gossip, the IID data split used in all experiments, and the repurposability of Flower's simulation engine for decentralized learning.

free parameters (3)
  • communication rounds = 24 (MNIST), 101 (CIFAR10)
    Chosen by hand to allow convergence; not derived from theory.
  • local epochs per head selection = 2,4,8,16,32 (32 for headline results)
    Varied across runs; 32 used for reported averages in Table II.
  • topology degree for reported average = Topology 4 (double ring)
    Selected post hoc as the first topology where E and R agents converge; other topologies shown only in figures.
assumptions (3)
  • domain assumption Round-robin synchronous head selection with reliable message passing adequately models asynchronous decentralized gossip learning.
    Algorithm 1 uses 'iteration mod K' and assumes agents can retrieve neighbor weights without modeling message loss, delays, or Byzantine behavior.
  • domain assumption Local data is IID across agents.
    Section IV-A splits MNIST and CIFAR10 in an IID way; the paper defers non-IID to future work in Section VI.
  • ad hoc to paper The Flower simulation engine can be repurposed for decentralized learning via upper-level class modifications without semantic distortion.
    Section III: 'modifying only upper-level classes client, server, model, dataset and creating a custom strategy', with one tweak to the aggregate class; this tooling assumption is central to the contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GLow -- A Novel, Flower-Based Simulated Gossip Learning Strategy." pith.science (2026). https://pith.science/paper/EWCPWRKE

@misc{pith2026250110463,
  author       = {Pith},
  title        = {Pith review of: GLow -- A Novel, Flower-Based Simulated Gossip Learning Strategy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWCPWRKE}},
  note         = {Machine review of arXiv:2501.10463}
}
read the original abstract

Fully decentralized learning algorithms are still in an early stage of development. Creating modular Gossip Learning strategies is not trivial due to convergence challenges and Byzantine faults intrinsic in systems of decentralized nature. Our contribution provides a novel means to simulate custom Gossip Learning systems by leveraging the state-of-the-art Flower Framework. Specifically, we introduce GLow, which will allow researchers to train and assess scalability and convergence of devices, across custom network topologies, before making a physical deployment. The Flower Framework is selected for being a simulation featured library with a very active community on Federated Learning research. However, Flower exclusively includes vanilla Federated Learning strategies and, thus, is not originally designed to perform simulations without a centralized authority. GLow is presented to fill this gap and make simulation of Gossip Learning systems possible. Results achieved by GLow in the MNIST and CIFAR10 datasets, show accuracies over 0.98 and 0.75 respectively. More importantly, GLow performs similarly in terms of accuracy and convergence to its analogous Centralized and Federated approaches in all designed experiments.

Figures

Figures reproduced from arXiv: 2501.10463 by the authors.

Figure 1
Figure 1. Topologies in 8+2 agent configuration, from a fully [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Loss evolution of CNL (left) and FL (right) systems in the MNIST dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Loss evolution of CNL (left) and FL (right) systems in the CIFAR10 dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Per-agent accuracy obtained in GLow simulation, MNIST dataset with 8+2 (left) and 16+4 agents (right). [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Loss evolution of each 8+2 agents during 24 communication rounds (32 local epochs) for topologies 0, 2, 4 and 7 in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Per-agent accuracy obtained in GLow simulation, CIFAR10 dataset with 8+2 (left) and 16+4 (right) agents. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Loss evolution of each 8+2 agents during 101 communication rounds (32 local epochs) for topologies 0, 2, 4 and 7 in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

  1. [2]

    Decentralized learning works: An empirical comparison of gossip learning and federated learning,

    I. Heged ˝us, G. Danner, and M. Jelasity, “Decentralized learning works: An empirical comparison of gossip learning and federated learning,” Journal of Parallel and Distributed Computing , vol. 148, pp. 109–124, 2 2021

    work page 2021

  2. [14]

    Gossipfl: A decentralized federated learning framework with sparsified and adaptive communication,

    Z. Tang, S. Shi, B. Li, and X. Chu, “Gossipfl: A decentralized federated learning framework with sparsified and adaptive communication,” IEEE Transactions on Parallel and Distributed Systems, vol. 34, pp. 909–922, 3 2023

    work page 2023

  3. [11]

    Braintorrent: A peer-to-peer environment for decentralized federated learning,

    A. G. Roy, S. Siddiqui, S. P ¨olsterl, N. Navab, and C. Wachinger, “Braintorrent: A peer-to-peer environment for decentralized federated learning,” 5 2019. [Online]. Available: http://arxiv.org/abs/1905.06731

    arXiv 2019

  4. [12]

    Fedstellar: A platform for decentralized federated learning,

    E. T. M. Beltr ´an, ´Angel Luis Perales G ´omez, C. Feng, P. M. S. S´anchez, S. L. Bernal, G. Bovet, M. G. P ´erez, G. M. P ´erez, and A. H. Celdr ´an, “Fedstellar: A platform for decentralized federated learning,” Expert Systems with Applications , vol. 242, 5 2024

    work page 2024

  5. [1]

    Communication-efficient learning of deep networks from decentralized data,

    H. Brendan McMahan, E. Moore, D. Ramage, S. Hampson, and B. Ag¨uera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” Proceedings of the 20th International Confer- ence on Artificial Intelligence and Statistics, AISTATS 2017 , vol. 54, 2017

    work page 2017

  6. [3]

    Flower: A friendly federated learning research framework,

    D. J. Beutel, T. Topal, A. Mathur, X. Qiu, T. Parcollet, and N. D. Lane, “Flower: A friendly federated learning research framework,” CoRR, vol. abs/2007.14390, 2020. [Online]. Available: https://arxiv.org/abs/2007.14390

    arXiv 2007

  7. [4]

    Survey of network intrusion detection methods from the perspective of the knowledge discovery in databases process,

    B. Molina-Coronado, U. Mori, A. Mendiburu, and J. Miguel-Alonso, “Survey of network intrusion detection methods from the perspective of the knowledge discovery in databases process,” IEEE Transactions on Network and Service Management, vol. 17, no. 4, pp. 2451–2479, 2020

    work page 2020

  8. [5]

    A Survey on Federated Learning Systems: Vision, Hype and Reality for Data Privacy and Protection,

    Q. Li, Z. Wen, Z. Wu, S. Hu, N. Wang, Y . Li, X. Liu, and B. He, “A Survey on Federated Learning Systems: Vision, Hype and Reality for Data Privacy and Protection,” pp. 1–44, 2019. [Online]. Available: http://arxiv.org/abs/1907.09693

    arXiv 2019

Show all 20 references
  1. [6]

    Calibrating noise to sensitivity in private data analysis,

    C. Dwork, F. McSherry, K. Nissim, and A. Smith, “Calibrating noise to sensitivity in private data analysis,” vol. V ol. 3876, 01 2006, pp. 265– 284

    2006

  2. [7]

    Deep Anomaly Detection for Time-Series Data in Industrial IoT: A Communication-Efficient On-Device Federated Learning Approach,

    Y . Liu, S. Garg, J. Nie, Y . Zhang, Z. Xiong, J. Kang, and M. S. Hossain, “Deep Anomaly Detection for Time-Series Data in Industrial IoT: A Communication-Efficient On-Device Federated Learning Approach,” IEEE Internet of Things Journal , vol. 8, no. 8, pp. 6348–6358, 2021

    2021

  3. [8]

    Gossip learning with linear models on fully distributed data,

    R. Orm ´andi, I. Hegedus, and M. Jelasity, “Gossip learning with linear models on fully distributed data,” in Concurrency and Computation: Practice and Experience , vol. 25. John Wiley and Sons Ltd, 2 2013, pp. 556–571

    2013

  4. [9]

    Decentralized federated learning: A survey and perspective,

    L. Yuan, Z. Wang, L. Sun, P. S. Yu, and C. G. Brinton, “Decentralized federated learning: A survey and perspective,” 6 2023. [Online]. Available: http://arxiv.org/abs/2306.01603

    2023 arXiv

  5. [10]

    Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and challenges,

    E. T. Mart ´ınez Beltr ´an, M. Q. P ´erez, P. M. S. S ´anchez, S. L. Bernal, G. Bovet, M. G. P ´erez, G. M. P ´erez, and A. H. Celdr ´an, “Decentralized federated learning: Fundamentals, state of the art, frameworks, trends, and challenges,” IEEE Communications Surveys & Tutor...

    2023

  6. [13]

    Communication-efficient de- sign for quantized decentralized federated learning,

    L. Chen, W. Liu, Y . Chen, and W. Wang, “Communication-efficient de- sign for quantized decentralized federated learning,” IEEE Transactions on Signal Processing , vol. 72, pp. 1175–1188, 2024

    2024

  7. [15]

    An asynchronous federated learning focusing on updated models for decentralized systems with a practical framework,

    Y . Kanamori, Y . Yamasaki, S. Hosoai, H. Nakamura, and H. Takase, “An asynchronous federated learning focusing on updated models for decentralized systems with a practical framework,” in Proceedings - International Computer Software and Applications Conference , vol. 2023-Jun...

    2023

  8. [16]

    MNIST handwritten digit database,

    Y . LeCun and C. Cortes, “MNIST handwritten digit database,” 2010. [Online]. Available: http://yann.lecun.com/exdb/mnist/

    2010

  9. [17]

    Cifar-10 (canadian institute for advanced research)

    A. Krizhevsky, V . Nair, and G. Hinton, “Cifar-10 (canadian institute for advanced research).” [Online]. Available: http://www.cs.toronto.edu/ ∼kriz/cifar.html

  10. [18]

    Gradient-based learning applied to document recognition,

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

    1998

  11. [19]

    Clustered federated learning based on client’s prototypes,

    W. Lai, Z. Xu, and Q. Yan, “Clustered federated learning based on client’s prototypes,” in 2024 27th International Conference on Computer Supported Cooperative Work in Design (CSCWD) , 2024, pp. 909–914

    2024

  12. [20]

    A survey on quantization methods for optimization of deep neural networks,

    U. Kulkarni, A. S. Hosamani, A. S. Masur, S. Hegde, G. R. Vernekar, and K. Siri Chandana, “A survey on quantization methods for optimization of deep neural networks,” in 2022 International Conference on Automation, Computing and Renewable Systems (ICACRS) , 2022, pp. 827–834

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.