Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Model-heterogeneous federated learning rankings change when evaluated under measured real-device constraints rather than proportional model splits; depth-level methods and SHeteroFL perform best, and memory limits reverse several results.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

PracMHBench evaluates eight model-heterogeneous federated learning algorithms under practical edge device constraints and finds that depth-level heterogeneity wins under compute/communication limits while memory limits change the ranking.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful benchmark that rightly questions proportional model splits in MHFL, but its central rankings rest on an under-specified model-assignment procedure that needs full disclosure before the specific conclusions can be trusted. the 4 major comments →

arxiv 2509.08750 v1 pith:CH5VI2BD submitted 2025-09-04 cs.LG cs.DC

PracMHBench: Re-evaluating Model-Heterogeneous Federated Learning Based on Practical Edge Device Constraints

Yuanchun Guo , Bingyan Liu , Yulong Sha , Zhensheng Xian This is my paper
classification cs.LG cs.DC
keywords model-heterogeneous federated learningedge device constraintsbenchmarking platformresource-aware model assignmentdepth heterogeneitySHeteroFLtime-to-accuracynon-IID federated learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that existing comparisons of model-heterogeneous federated learning (MHFL) are unfair because they assign devices models by fixed proportions of a reference model, ignoring what those splits actually cost on real edge hardware. It builds PracMHBench, a platform that assigns models according to measured device constraints—computation, communication, and memory—and re-runs eight MHFL algorithms across vision, language, and activity-recognition tasks. Under computation-limited and communication-limited settings, depth-level methods (FeDepth, InclusiveFL, DepthFL) generally dominate on accuracy and training speed; memory-limited settings weaken DepthFL, strengthen FeDepth, and change stability conclusions; when constraints are combined, SHeteroFL is the most consistently strong method. If correct, the paper provides the first quantitative, deployment-oriented ranking of MHFL algorithms and a template for benchmarking them fairly.

Core claim

The central claim is that the right unit for comparing MHFL methods is not model proportion but measured resource behavior on representative devices. The paper reports that under computation- and communication-limited MHFL, depth-level heterogeneity consistently outperforms width- and topology-level methods across data tasks, while Fjord and FedProto are weakest. Memory-limited settings change this picture: DepthFL's large memory footprint forces it to shrink and lose accuracy, whereas FeDepth's small footprint lets it fit larger models and rise. Combining memory with communication, or all three constraints, favors SHeteroFL, whose slimmable partition method controls computation, memory, and

What carries the argument

PracMHBench, the platform: a model pool in which candidate sub-models from each method are measured for parameters, FLOPs, training time, and memory usage on Jetson Orin NX, Jetson TX2 NX, and Raspberry Pi; plus three constraint cases—computation-limited (equalize per-round training time using device computing-power statistics), communication-limited (equalize upload/download time to a fixed per-round budget), and memory-limited (assign the largest trainable model per device memory class). It functions as the fair assignment mechanism that replaces proportional splitting, and its evaluation metrics are global accuracy, time-to-accuracy, stability, and effectiveness.

Load-bearing premise

All three constraint scenarios assume the model-pool assignment rule—matching each client's measured speed, bandwidth, or memory to exactly one sub-model per method—is fair and that device measurements transfer to the simulated client population; if that mapping is biased, every ranking changes.

What would settle it

Re-run PracMHBench's protocol after recomputing each algorithm's sub-model assignment from fresh measurements on a larger, more diverse device fleet or on a different architecture family such as pure Transformers. If depth-level methods no longer dominate in computation- and communication-limited settings, or if DepthFL's memory degradation disappears, the paper's rankings are artifacts of its assignment measurements rather than properties of the algorithms.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners targeting compute-limited or bandwidth-limited federated deployments should prefer depth-level heterogeneous algorithms, which held the accuracy and training-speed advantage across every data task tested.
  • In memory-constrained deployments, algorithm choice should be driven by actual memory footprint: FeDepth benefits while DepthFL degrades sharply.
  • When devices face multiple constraints at once, SHeteroFL's slimmable width partition appears to be the safest default among the tested methods.
  • Stability and effectiveness should be reported separately in MHFL benchmarks; the experiments found no single method delivers both across datasets.
  • Benchmarking practice for MHFL should report measured device constraints rather than model proportions alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If proportional splitting systematically misrepresents real cost, earlier MHFL results that rank methods by accuracy under such splits may be artefactual; re-running them under PracMHBench-style constraints could change not just rankings but which heterogeneity level researchers focus on.
  • A testable extension is a cost-accuracy frontier: for each method, plot accuracy against measured device-seconds and memory at several constraint levels, letting a deployer choose from a Pareto front instead of a single ranking.
  • The finding that memory footprint, not parameter count, drives memory-constrained behavior suggests FL systems could dynamically re-select model assignments as device memory availability changes—a regime the paper does not explore.
  • The constraint-assignment protocol could be applied to other model families and broader hardware fleets to see whether depth-level dominance is a general property of residual architectures or specific to the tested pool.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PracMHBench, a benchmark platform for model-heterogeneous federated learning (MHFL) under practical edge-device constraints. It classifies existing MHFL methods into width, depth, and topology heterogeneity levels, and evaluates them on CV, NLP, and HAR tasks using four metrics: global accuracy, time-to-accuracy, stability, and effectiveness. The main methodological contribution is the construction of three constraint cases—computation-limited, communication-limited, and memory-limited MHFL—based on real device statistics (IMA dataset, AI Benchmark, memory distributions). The central empirical claims are: (1) under computation and communication limits, depth-level methods (FeDepth, InclusiveFL, DepthFL) are generally superior; (2) memory limits reverse part of this picture, with DepthFL degrading and FeDepth improving; (3) SHeteroFL is the most robust method when constraints are combined (Figure 7). The paper also includes non-IID and scalability analyses.

Significance. If the results are reliable, this is a valuable contribution: it challenges the field's default of using parameter-count proportions to model device heterogeneity, and it provides the first systematic, multi-task, multi-metric comparison of MHFL algorithms under measured device constraints. The platform and code release are a strength, as is the breadth of tasks and methods. The paper's core message—that evaluation should be grounded in real device capabilities—is well motivated and timely. However, the validity of the rankings rests on constraint-case construction details that are currently underspecified, so the significance is conditional on those details being made precise and validated.

major comments (4)
  1. The central ranking under computation-limited MHFL depends on an unspecified and unvalidated mapping from device computing power to a method-specific sub-model. The text says models are selected from a model pool 'with the help of the statistics of the training time for various models, which is provided by an AI Benchmark,' but AI Benchmark does not contain the method-specific sub-models used here (e.g., a 0.4-width SHeteroFL model or a 0.6-depth FeDepth model). No exact rule, tolerance for 'roughly equivalent training times,' or validation that the assigned models actually equalize per-round training times is given. Table I shows that at the same nominal proportion, DepthFL requires 515.93 s vs. SHeteroFL's 430.24 s on Jetson Nano; an assignment based only on parameter count could therefore systematically favor depth-level methods by giving them larger effective compute. Since all subse
  2. The memory-limited setting is defined by assigning 'the largest trainable model' to each client, but the rule is not reproducible. It is not stated whether memory usage is measured as peak CUDA memory, PyTorch allocator memory, or an estimate from parameter counts, nor how 'trainable' is determined across Jetson Orin NX (16GB), Jetson TX2 NX (4GB), and Raspberry Pi 4B (no GPU). The actual assigned model proportions for each method are never reported. This matters because the memory-limited results reverse the computation-limited ranking (DepthFL degrades, FeDepth improves); without the assignment details, this reversal cannot be audited. Additionally, Section V states that memory-limited experiments only use ResNet-101 and ALBERT and exclude small models for HAR; this scope restriction should be stated in Section IV.C and taken into account when drawing general memory-constrained conclus
  3. Time-to-accuracy is a load-bearing metric in the paper, but the 'pre-set accuracy' threshold is never reported, nor is it stated whether the threshold varies by dataset or task. Without this value, the time-to-accuracy comparisons in Figures 4-6 cannot be reproduced or interpreted. Furthermore, all experiments are run three times and averaged (Section V), but no variance, confidence interval, or statistical test is reported for any result. For an empirical benchmark whose contribution is a method ranking, this is insufficient to establish that observed differences are not noise. Please report per-run values or error bars, and state the accuracy target used for time-to-accuracy.
  4. The communication-limited setting is described only as controlling per-round communication time 'to a certain time (e.g., 200s)' based on IMA bandwidth data, followed by selecting 'appropriate models and corresponding quantities from the model pool.' The paper does not specify the bandwidth distribution, how model sizes are converted to communication times (e.g., bytes/parameter, protocol overhead), or what 'quantities' refers to (number of clients? number of sub-models?). Since Figure 5's conclusion—that the ranking under communication limits mirrors the computation-limited ranking—depends on this construction, the procedure needs to be fully specified and validated, including a sensitivity analysis of the 200s cap.
minor comments (5)
  1. Figure 1's radar charts are described as 'just for demonstration' but appear in the evaluation track diagram; please either label them as illustrative or remove them to avoid confusion.
  2. Typo: 'Jeston Nano' and 'JESTON NANO' should be 'Jetson Nano.'
  3. The AI Benchmark reference is a webpage without a version number or access date; please add the accessed date and, if available, the benchmark version.
  4. The non-IID analysis is performed only under computation-limited MHFL (Figure 8), but the text states 'the conclusion is not changed' in general. Please qualify this claim to the computation-limited setting, or add non-IID results for the other constraint cases.
  5. The effectiveness baseline is described as 'training the smallest homogeneous model across all heterogeneous devices.' Please specify which model that is for each dataset/heterogeneity level and how it is trained, since the improvement values depend on this baseline.

Circularity Check

0 steps flagged

No circular derivation: PracMHBench is an empirical benchmark whose rankings are externally anchored; minor caveats about constraint-case construction are fairness risks, not circularity.

full rationale

This paper is an empirical measurement study, not a formal derivation, so the standard circularity patterns (self-definitional equivalence, fitted-input-called-prediction, uniqueness imported from authors, ansatz smuggled via citation) do not apply. The central claims—that depth-level methods dominate under computation/communication limits, that memory limits reverse parts of that picture, and that SHeteroFL is most robust under combined constraints—are obtained by running existing MHFL algorithms under three constraint cases built from external data sources: the IMA device dataset [21], AI Benchmark device statistics [37], and the paper's own measured Table I statistics on Jetson/Raspberry Pi hardware. The model-pool assignment is based on measured training time, memory, and communication cost, not on the accuracy or ranking outcomes; therefore the reported accuracy and time-to-accuracy rankings are not constructed to match the assignment rule. The 'effectiveness' metric is defined against an independent homogeneous baseline, providing an external reference rather than a self-referential target. The only self-citations (DistFL [3], FedTuning [10]) are background references to the corresponding author's prior federated-learning work and are not load-bearing for the benchmark conclusions. The main weakness—that the mapping from AI Benchmark scores to method-specific submodels is not fully validated and could affect assignment fairness—is a legitimate correctness/validity concern about the experimental protocol, but it is not a circularity: the paper does not fit its conclusions into the assignment procedure, and the rankings remain empirically falsifiable. Hence no circular step can be identified, and the appropriate circularity score is very low.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The ledger for this paper is about experimental construction, not mathematical derivation. The main unfree parameters are the undisclosed time-to-accuracy threshold and the unspecified model-assignment tolerance. The axioms are domain assumptions about device statistics, model-pool transferability, dataset representativeness, and synchronous aggregation. No invented theoretical entities are introduced.

free parameters (2)
  • pre-set accuracy target for time-to-accuracy
    Section III defines time-to-accuracy as wall-clock time to reach a pre-set accuracy, but the target value is never reported. Every time-to-accuracy comparison depends on this hand-chosen, undisclosed threshold.
  • constraint-matching tolerance for model assignment
    Section IV constructs each constraint case by selecting models from a pool to satisfy a resource limit. The tolerance for matching training speed (computation/communication) or the rule for choosing the largest trainable model (memory) is not specified, and all rankings inherit this choice.
axioms (4)
  • domain assumption IMA dataset device capabilities and AI Benchmark training-time statistics accurately represent practical edge devices.
    Invoked in Section IV.A and IV.B to set per-client computational and communication limits; if these measurements are not representative, the constructed constraint cases do not reflect real edges.
  • domain assumption The model-pool measurements (training time, memory, parameters) for a few devices transfer to the full client population.
    Table I and Figure 3 provide measurements on Jetson Orin NX and Jetson Nano; these are used to select models for all clients, including Raspberry Pi-class devices in the memory case.
  • domain assumption The chosen datasets and model architectures are representative of CV, NLP, and HAR tasks.
    Section III-B selects six datasets and several architectures and then generalizes conclusions to 'all data tasks'; representativeness is assumed, not demonstrated.
  • domain assumption Synchronous aggregation with roughly equalized per-client training time is the correct framing for computation- and communication-limited MHFL.
    Definitions IV.1 and IV.2 define the scenarios by equalizing training or communication speed; if real deployments use asynchronous aggregation, these scenarios may not be representative.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PracMHBench: Re-evaluating Model-Heterogeneous Federated Learning Based on Practical Edge Device Constraints." pith.science (2026). https://pith.science/paper/CH5VI2BD

@misc{pith2026250908750,
  author       = {Pith},
  title        = {Pith review of: PracMHBench: Re-evaluating Model-Heterogeneous Federated Learning Based on Practical Edge Device Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CH5VI2BD}},
  note         = {Machine review of arXiv:2509.08750}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Federating heterogeneous models on edge devices with diverse resource constraints has been a notable trend in recent years. Compared to traditional federated learning (FL) that assumes an identical model architecture to cooperate, model-heterogeneous FL is more practical and flexible since the model can be customized to satisfy the deployment requirement. Unfortunately, no prior work ever dives into the existing model-heterogeneous FL algorithms under the practical edge device constraints and provides quantitative analysis on various data scenarios and metrics, which motivates us to rethink and re-evaluate this paradigm. In our work, we construct the first system platform \textbf{PracMHBench} to evaluate model-heterogeneous FL on practical constraints of edge devices, where diverse model heterogeneity algorithms are classified and tested on multiple data tasks and metrics. Based on the platform, we perform extensive experiments on these algorithms under the different edge constraints to observe their applicability and the corresponding heterogeneity pattern.

Figures

Figures reproduced from arXiv: 2509.08750 by Bingyan Liu, Yuanchun Guo, Yulong Sha, Zhensheng Xian.

Figure 1
Figure 1. Figure 1: Evaluation track of our platform PracMHBench. Here the information in radar charts is just for demonstration. The concrete performance will be [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Demonstration of different heterogeneity levels. Here ‘skip’ denotes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of our constructed model pool. We use the tested system [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results on computation-limited MHFL. The top row refers to the comprehensive performance of global accuracy and time-to-accuracy. The bottom row [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results on communication-limited MHFL. better better [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results on memory-limited MHFL. Observation: In computation-limited MHFL, the superiority in terms of global accuracy and training speed remains a consistent pattern regardless of changes in data tasks. Among them, the methods belonging to the depth-level heterogeneity are relatively better than others. However, there is no one￾size-fits-all algorithm that can achieve a desirable trade-off between the stab… view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of constraint combinations. iid niid-0.5 niid-5 CIFAR100 0.4 0.6 Accuracy iid niid-0.5 niid-5 CIFAR10 Non-IID Performance iid niid-0.5 niid-5 Ag-news Fjord SHeteroFL FedRolex FeDepth InclusiveFL DepthFL Fed-ET FedProto [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Analysis of scalability. C. Results on Memory-Limited MHFL Global accuracy and time-to-accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 6
Figure 6. Figure 6: From the figure, we find the following interesting observation: [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 29 canonical work pages · 2 internal anchors

  1. [1]

    Towards federated learning at scale: System design

    Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloe Kiddon, Jakub Kone ˇcn`y, Stefano Mazzocchi, H Brendan McMahan, et al. Towards federated learning at scale: System design. arXiv preprint arXiv:1902.01046 , 2019

    Pith/arXiv arXiv 1902

  2. [2]

    Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space

    Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1013–1023, 2021

    work page 2021

  3. [3]

    Distfl: Distribution-aware federated learning for mobile scenarios

    Bingyan Liu, Yifeng Cai, Ziqi Zhang, Yuanchun Li, Leye Wang, Ding Li, Yao Guo, and Xiangqun Chen. Distfl: Distribution-aware federated learning for mobile scenarios. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , 5(4):1–26, 2021

    work page 2021

  4. [4]

    Pyramidfl: A fine-grained client selection framework for efficient federated learning

    Chenning Li, Xiao Zeng, Mi Zhang, and Zhichao Cao. Pyramidfl: A fine-grained client selection framework for efficient federated learning. In Proceedings of the 28th Annual International Conference on Mobile Computing And Networking , pages 158–171, 2022

    work page 2022

  5. [5]

    Heterofl: Computation and communication efficient federated learning for heterogeneous clients

    Enmao Diao, Jie Ding, and Vahid Tarokh. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. ICLR, 2021

    work page 2021

  6. [6]

    Fedrolex: Model- heterogeneous federated learning with rolling sub-model extraction

    Samiul Alam, Luyang Liu, Ming Yan, and Mi Zhang. Fedrolex: Model- heterogeneous federated learning with rolling sub-model extraction. NeurIPS, 2022

    work page 2022

  7. [7]

    Fedproto: Federated prototype learning across heterogeneous clients

    Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8432–8440, 2022

    work page 2022

  8. [8]

    Fedbalancer: data and pace control for efficient federated learning on heterogeneous clients

    Jaemin Shin, Yuanchun Li, Yunxin Liu, and Sung-Ju Lee. Fedbalancer: data and pace control for efficient federated learning on heterogeneous clients. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services , pages 436–449, 2022

    work page 2022

  9. [9]

    Federated few-shot learning for mobile nlp

    Dongqi Cai, Shangguang Wang, Yaozong Wu, Felix Xiaozhu Lin, and Mengwei Xu. Federated few-shot learning for mobile nlp. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–17, 2023

    work page 2023

  10. [10]

    Beyond fine-tuning: Efficient and effective fed- tuning for mobile/web users

    Bingyan Liu, Yifeng Cai, Hongzhe Bi, Ziqi Zhang, Ding Li, Yao Guo, and Xiangqun Chen. Beyond fine-tuning: Efficient and effective fed- tuning for mobile/web users. In Proceedings of the ACM Web Conference 2023, pages 2863–2873, 2023

    work page 2023

  11. [11]

    Tailorfl: Dual-personalized federated learning under system and data heterogeneity

    Yongheng Deng, Weining Chen, Ju Ren, Feng Lyu, Yang Liu, Yunxin Liu, and Yaoxue Zhang. Tailorfl: Dual-personalized federated learning under system and data heterogeneity. In Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems , pages 592–606, 2022

    work page 2022

  12. [12]

    Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout

    Samuel Horvath, Stefanos Laskaridis, Mario Almeida, Ilias Leontiadis, Stylianos Venieris, and Nicholas Lane. Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout. Advances in Neural Information Processing Systems , 34:12876–12889, 2021

    work page 2021

  13. [13]

    Group knowledge transfer: Federated learning of large cnns at the edge.Advances in Neural Information Processing Systems , 33:14068–14080, 2020

    Chaoyang He, Murali Annavaram, and Salman Avestimehr. Group knowledge transfer: Federated learning of large cnns at the edge.Advances in Neural Information Processing Systems , 33:14068–14080, 2020

    work page 2020

  14. [14]

    Heterogeneous ensemble knowledge transfer for training large models in federated learning

    Yae Jee Cho, Andre Manoel, Gauri Joshi, Robert Sim, and Dimitrios Dimitriadis. Heterogeneous ensemble knowledge transfer for training large models in federated learning. IJCAI, 2022

    work page 2022

  15. [15]

    Fedduap: Federated learning with dynamic update and adaptive pruning using shared data on the server

    Hong Zhang, Ji Liu, Juncheng Jia, Yang Zhou, Huaiyu Dai, and Dejing Dou. Fedduap: Federated learning with dynamic update and adaptive pruning using shared data on the server. IJCAI, 2022

    work page 2022

  16. [16]

    Leaf: A benchmark for federated settings

    Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Kone ˇcn`y, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018

    Pith/arXiv arXiv 2018

  17. [17]

    FedEval: A Holistic Evaluation Framework for Federated Learning

    Di Chai, Leye Wang, Liu Yang, Junxue Zhang, Kai Chen, and Qiang Yang. Fedeval: A holistic evaluation framework for federated learning. arXiv preprint arXiv:2011.09655 , 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  18. [18]

    End-to-end evaluation of federated learning and split learning for internet of things

    Yansong Gao, Minki Kim, Sharif Abuadbba, Yeonjae Kim, Chandra Thapa, Kyuyeon Kim, Seyit A Camtepe, Hyoungshick Kim, and Surya Nepal. End-to-end evaluation of federated learning and split learning for internet of things. International Symposium on Reliable Distributed Systems (SRDS), 2020

    work page 2020

  19. [19]

    Fedscale: Benchmarking model and system performance of federated learning at scale

    Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xiangfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. Fedscale: Benchmarking model and system performance of federated learning at scale. In International Conference on Machine Learning , pages 11814–11827. PMLR, 2022

    work page 2022

  20. [20]

    Federated learning on non-iid data silos: An experimental study

    Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE) , pages 965–978. IEEE, 2022

    work page 2022

  21. [21]

    Characterizing impacts of heterogeneity in federated learning upon large-scale smartphone data

    Chengxu Yang, Qipeng Wang, Mengwei Xu, Zhenpeng Chen, Kaigui Bian, Yunxin Liu, and Xuanzhe Liu. Characterizing impacts of heterogeneity in federated learning upon large-scale smartphone data. In Proceedings of the Web Conference 2021 , pages 935–946, 2021

    work page 2021

  22. [22]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770–778, 2016

    2016

  23. [23]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  24. [24]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 , 2017

    Pith/arXiv arXiv 2017

  25. [25]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1–9, 2015

    2015

  26. [26]

    Memory-adaptive Depth-wise Heterogeneous Federated Learning

    Kai Zhang, Yutong Dai, Hongyi Wang, Eric Xing, Xun Chen, and Lichao Sun. Memory-adaptive depth-wise heterogenous federated learning. arXiv preprint arXiv:2303.04887, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    No one left behind: Inclusive federated learning over heterogeneous devices

    Ruixuan Liu, Fangzhao Wu, Chuhan Wu, Yanlin Wang, Lingjuan Lyu, Hong Chen, and Xing Xie. No one left behind: Inclusive federated learning over heterogeneous devices. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 3398–3406, 2022

    work page 2022

  28. [28]

    Depthfl: Depthwise federated learning for heterogeneous clients

    Minjae Kim, Sangyoon Yu, Suhyun Kim, and Soo-Mook Moon. Depthfl: Depthwise federated learning for heterogeneous clients. In The Eleventh International Conference on Learning Representations , 2022

    work page 2022

  29. [29]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

    2009

  30. [30]

    Character-level convolu- tional networks for text classification

    Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolu- tional networks for text classification. Advances in neural information processing systems, 28, 2015

    work page 2015

  31. [31]

    Tensorflow federated stack overflow dataset, 2019

    tff. Tensorflow federated stack overflow dataset, 2019

    work page 2019

  32. [32]

    Clusterfl: a similarity-aware federated learning system for human activity recognition

    Xiaomin Ouyang, Zhiyuan Xie, Jiayu Zhou, Jianwei Huang, and Guoliang Xing. Clusterfl: a similarity-aware federated learning system for human activity recognition. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services , pages 54–66, 2021

    work page 2021

  33. [33]

    A public domain dataset for human activity recognition using smartphones

    Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, Jorge Luis Reyes-Ortiz, et al. A public domain dataset for human activity recognition using smartphones. In Esann, volume 3, page 3, 2013

    work page 2013

  34. [34]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  35. [35]

    Albert: A lite bert for self-supervised learning of language representations

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942 , 2019

    Pith/arXiv arXiv 1909

  36. [36]

    Evaluating federated learning for human activity recognition

    Sannara Ek, François Portet, Philippe Lalanda, and German Eduardo Vega Baez. Evaluating federated learning for human activity recognition. In Workshop AI for Internet of Things, in conjunction with IJCAI-PRICAI 2020, 2021

    work page 2020

  37. [37]

    https://ai-benchmark.com/ranking_ deeplearning_detailed.html, 2021

    Ai benchmark for different devices. https://ai-benchmark.com/ranking_ deeplearning_detailed.html, 2021

    work page 2021

  38. [38]

    How much ram is in smartphones

    ScientiaMobile. How much ram is in smartphones. https://www. scientiamobile.com/how-much-ram-is-in-smartphones/, 2022. Accessed: 2025-08-29

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.