Pith. sign in

REVIEW 4 major objections 5 minor 49 references

ML Mule: Mobile-Driven Context-Aware Collaborative Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read ML Mule shows that models can evolve through physical spaces alone, carried by mobile devices.

desk verdict Novel space-only coupling idea, but the headline accuracy comparisons are likely apples-to-oranges and need fixing. read the letter →

arxiv 2501.07536 v2 pith:GNJJARO4 submitted 2025-01-13 cs.LG cs.HC

classification cs.LGcs.HC
keywords decentralizedlearningfederatedcontext-awaremachinemobilecomputingmodelmulesopportunistichumanactivityrecognitionsmartenvironments
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

ML Mule claims that machine-learning models can be trained collaboratively without a central server and without requiring devices to be present at the same time. Instead, the phones people carry act as mules that transport model snapshots between fixed devices in different physical spaces, so collaboration is coupled only in space and decoupled in time. The paper argues that this creates implicit affinity groups among devices whose users share spaces, letting a model trained in one room usefully seed a model in another room. Across simulated mobility patterns and two tasks, image classification and human activity recognition, the authors report that ML Mule converges faster and reaches higher accuracy than federated, clustered, gossip, and opportunistic baselines, and that it still works with little or no internet access. The significance, if true, is a learning paradigm whose collaboration metric is physical space rather than network reachability or simultaneity.

What carries the argument

The central mechanism is the mule cycle: a mobile device acts as a mule, carrying a model snapshot from one fixed device to another, while fixed devices host, aggregate, and train the model during an in-house phase. A dynamic freshness filter, based on the median absolute deviation of model update times, prevents stale snapshots from contaminating aggregation, and weighted averaging combines the exchanged models. The fixed device serves as a stable anchor that captures space-specific features over many visitors, which the paper credits for the method's advantage over purely encounter-based approaches.

What would settle it

Run the same CIFAR-100 simulation with data splits shuffled across spaces so that co-location carries no class similarity, then compare ML Mule with Local Only; if ML Mule still wins, spatial affinity is not the operative mechanism, and if it loses, the method's core premise is confirmed.

Watch

Extended reading notes

Core claim

The paper's central claim is that spatial coupling alone is sufficient for collaborative model evolution. A mobile device entering a space discovers a fixed device, exchanges model weights with it, performs or receives local training, and then carries the updated snapshot to the next space; a fixed device aggregates snapshots from many visitors and thereby accumulates space-specific knowledge. The paper states that this process consistently outperforms or matches FedAvg, Clustered Federated Learning, FedAS, Gossip Learning, OppCL, and Local Only learning across Dirichlet and Shards data distributions under diverse mobility patterns, and that it remains effective when devices are disconnected from the internet.

Load-bearing premise

The claim stands on the premise that people who occupy the same physical space behave or look similar enough that carrying a model from one space to another helps rather than hurts; the experiments build this correlation into the data splits, so it is not tested against real-world co-location noise.

Editorial extensions

If this is right

  • In environments without reliable internet, ML Mule still evolves a model through peer-to-peer encounters with fixed devices, so low-cost gateways can be deployed without network configuration.
  • Because a mobile device can deposit a model in a space it visits and later collect an updated one, collaboration no longer requires devices to be present at the same time, only in the same space.
  • Aggregation is intentionally local: fixed devices accumulate space-specific knowledge over many visitors, which the paper argues is why ML Mule beats global or clustered baselines under non-i.i.d. data.
  • The method's accuracy is stable across mobility patterns, with higher crossing probabilities giving more diverse models at the cost of early-training fluctuation.

Reading between the lines

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

  • If spatial affinity is weak, ML Mule's advantage could invert; a natural hybrid would weight mule-carried models by measured similarity between the source and destination spaces rather than by fixed weighted averaging.
  • The paper's freshness threshold and dwell-time weighting suggest a testable design rule: systems should increase aggregation weight for longer dwell times and decay stale snapshots faster in high-turnover spaces.
  • Since only weights travel, not data, differential privacy or secure aggregation could be added to the exchange step; the paper lists this as future work but does not quantify the privacy-accuracy trade-off.
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. The paper proposes ML Mule, a decentralized learning framework in which mobile devices ('mules') physically carry model snapshots between fixed devices in different spaces, enabling asynchronous, server-free collaboration. The system has two phases (in-house and mule), a freshness threshold for accepting carried models, and weighted averaging at fixed devices. The authors evaluate two training modes on CIFAR-100 and EgoExo4D against FedAvg, FedAS, CFL, Gossip Learning, OppCL, and Local Only baselines, and they validate feasibility with a small prototype. The central claim is that ML Mule converges faster and consistently outperforms or matches these baselines across data distributions and mobility patterns.

Significance. The contribution is conceptually appealing: it replaces time-coupled device-to-device or server-client communication with space-coupled but time-decoupled mule transport, which is a genuinely different communication topology for distributed learning. The paper has several concrete strengths: it includes a real mobility trace (Foursquare) alongside synthetic random walks; it evaluates on two modalities (image classification and IMU-based activity recognition); it reports time-based convergence curves in Figures 6-9; and it provides a physical prototype with measured operation times. If the comparison-protocol issues are resolved, the work would be a useful addition to the decentralized and context-aware learning literature. At present, however, the headline empirical claim is not yet supported because the main comparison table appears to mix pre-fine-tuning baselines with a post-fine-tuning ML Mule.

major comments (4)
  1. [§4.2.1, Table 1] The evaluation in Table 1 is not a like-for-like comparison. Section 4.2.1 states that 'the performance has been evaluated after the model returned to the fixed devices, and was retrained for 1 epoch with local training data as a fine-tuning step,' and then immediately adds that 'We report the accuracy of the model before local training on the baseline methods.' Since ML Mule has only a single accuracy entry per mobility setting while the baselines have both Pre-Local and Post-Local columns, the reported ML Mule numbers appear to include the one-epoch fine-tuning, whereas the headline baseline numbers do not. For Dirichlet alpha=0.001, the relevant comparison is ML Mule's 91.18% against FedAvg's Post-Local 84.50%, not against its Pre-Local 32.45%. The claim in Contribution 3 that ML Mule 'consistently outperforms or matches' baselines therefore needs a re-run or re-reporting in which Pre-Local and Post-Local values are given for ML Mule too, and all comparisons are conducted on the same evaluation protocol.
  2. [§4.2.1, §4.3.2] The convergence comparisons mix incompatible round definitions. For ML Mule 'one round of model evolution' is defined as 20 successful peer-to-peer model exchanges, while for federated baselines one round is a single server aggregation and for Local Only one round is one local epoch. This can make ML Mule appear to converge in fewer rounds without consuming less communication or computation. The table header includes a Round column, but no convergence-round values are reported for any method. The authors should report convergence on a common time-step or wall-clock basis for the fixed-device experiment, as is already done in Figures 6-9 for the mobile-device experiments; otherwise the abstract's convergence claim is not supported by Table 1.
  3. [§1, §4.2.1, §4.3.1] The spatial-affinity premise is untested and is partly enforced by construction. The paper motivates the method with the statement that 'users who share physical spaces are likely to exhibit similar characteristics' (Section 1), but in Sections 4.2.1 and 4.3.1 the data distributions assign classes or super-classes to specific areas and spaces, so the correlation between space and label is built into the experimental setup. Since the claimed advantage of mule-based transport depends on this correlation, the evaluation needs a sensitivity analysis with weakened or absent space-label correlation, or the claims should be explicitly scoped to applications where that correlation is known to hold.
  4. [§4.2.2, Table 1] The 'consistently outperforms' statement is based on single point estimates. ML Mule rows and baseline rows in Table 1 have no variance or seed count; the text mentions that 'we conducted experiments with various random seeds and obtained similar results' but does not report those results. Please include confidence intervals or standard deviations, or at least seed-level results, for all entries in Table 1 and for the convergence curves in Figures 6-9.
minor comments (5)
  1. [§4.2.1, §4.2.2] The text gives contradictory descriptions of the Dirichlet parameter. Section 4.2.1 says 'smaller alpha values typically yield a distribution closer to iid setting,' while Section 4.2.2 says 'a larger value of Dirichlet alpha implies a more non-IID distribution.' In the standard Dirichlet partition, smaller alpha means more non-IID; please correct both statements and make the caption consistent.
  2. [§3] There are several typos in the notation: 'donate' should be 'denote' and 'donated by' should be 'denoted by'; also, 'overleap' should be 'overlap' in Section 4.1.
  3. [§3.1, §4.2.1] The symbol alpha is used both for the freshness-threshold smoothing parameter in the dynamic threshold equation and for the Dirichlet concentration parameter in the evaluation; please disambiguate these symbols to avoid confusion.
  4. [Table 1] The formatting of Table 1 is difficult to read: the Local Only row appears to have fewer values than the other baseline rows, the Round column is not populated, and the ML Mule rows would benefit from explicit Pre-Local/Post-Local subcolumns. Please reformat the table so each method and each mobility setting has clearly labeled entries.
  5. [Figures 6-9] The captions for Figures 6-9 do not consistently state the smoothing procedure or the y-axis scale; for example, the caption for Figure 7 says a 100-time-step moving average was applied, but the other figure captions do not. Please state these details in every relevant caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: ML Mule's claims are empirical comparisons against external baselines, with only a minor evaluation asymmetry that does not force the reported advantage.

full rationale

ML Mule is an empirical systems paper and does not derive its headline result from a fitted parameter, a uniqueness argument, or a self-citation chain. The central claims are benchmark comparisons against FedAvg, CFL, FedAS, Gossip Learning, OppCL, and Local Only on CIFAR-100 and EgoExo4D. The spatial-affinity premise ('users who share physical spaces are likely to exhibit similar characteristics') is an input assumption rather than a derived conclusion; the same data splits are used for all compared methods, so the comparison is not circular. The self-citations in the paper ([21], [44]) are background attributions and are not load-bearing. The one protocol asymmetry worth flagging is in Section 4.2.1: ML Mule's reported accuracy appears to include one epoch of local fine-tuning, while the baseline comparison values are reported 'before local training.' This could inflate the apparent margin, but it does not manufacture the result, because ML Mule's Table 1 values also exceed the baselines' Post-Local columns (e.g., 91.18% vs FedAvg's 84.50% at alpha=0.001), and the time-based plots in Figures 6-9 provide an independent view of convergence. The untested real-world validity of the spatial-affinity assumption is a validity concern, not a circularity concern. No step in the paper reduces to its own inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on the spatial-affinity axiom about the data distribution and on several hand-chosen simulation parameters (alpha, beta, delay, sharing time, round definition) that are not fully reported. The paper introduces no new physical or conceptual entities in the sense of the ledger; the mule is a role borrowed from delay-tolerant networking.

free parameters (6)
  • alpha (freshness threshold smoothing)
    Exponential moving average weight in the dynamic freshness threshold (Section 3.1); value not reported, affects which models are accepted.
  • beta (freshness threshold variability)
    Scales the median absolute deviation in the threshold update (Section 3.1); value not reported.
  • d (delay between training cycles)
    Constant wait between cycles during co-location (Section 3.1); not specified, affects aggregation weighting.
  • P_cross = 0, 0.1, 0.5
    Mobility crossing probability in the random-walk model (Section 4.1); tested values, not fitted, but a free simulation parameter that changes results.
  • P2P model sharing time = 3 time steps
    Assumed duration of peer-to-peer model transfer (Sections 4.1 and 4.3), affects convergence speed comparison.
  • ML Mule round length = 20 P2P exchanges
    Custom definition of one round for ML Mule (Section 4.2.1), different from FL baselines' round, potentially biasing round-based comparisons.
assumptions (5)
  • domain assumption Users who share physical spaces are likely to exhibit similar data characteristics.
    Stated in the Introduction as the basis for implicit affinity groups; the evaluation enforces this by assigning classes/activities to spaces, so real-world validity is untested.
  • domain assumption Weighted averaging of models from different spaces is safe under non-IID data.
    Aggregation uses weighted averaging [40] with no convergence or negative-transfer analysis, especially for heterogeneous space-specific models.
  • ad hoc to paper Random-walk mobility with P_cross captures the structure of real human mobility relevant to learning.
    The mobility model in Section 4.1 is designed for this paper and validated only qualitatively via an ICA plot of Foursquare data, not quantitatively.
  • domain assumption Short-range discovery and model exchange work whenever a mobile device enters a space.
    Assumed throughout the design; demonstrated only in a small prototype with two fixed devices and one mule.
  • domain assumption The constructed spatial data splits reflect real space-centric tasks.
    CIFAR-100 and EgoExo4D are real datasets, but the authors artificially partition classes/activities across spaces, which may not match natural spatial correlations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ML Mule: Mobile-Driven Context-Aware Collaborative Learning." pith.science (2026). https://pith.science/paper/GNJJARO4

@misc{pith2026250107536,
  author       = {Pith},
  title        = {Pith review of: ML Mule: Mobile-Driven Context-Aware Collaborative Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GNJJARO4}},
  note         = {Machine review of arXiv:2501.07536}
}
read the original abstract

Artificial intelligence has been integrated into nearly every aspect of daily life, powering applications from object detection with computer vision to large language models for writing emails and compact models for use in smart homes. These machine learning models at times cater to the needs of individual users but are often detached from them, as they are typically stored and processed in centralized data centers. This centralized approach raises privacy concerns, incurs high infrastructure costs, and struggles to provide real time, personalized experiences. Federated and fully decentralized learning methods have been proposed to address these issues, but they still depend on centralized servers or face slow convergence due to communication constraints. We propose ML Mule, an approach that utilizes individual mobile devices as 'mules' to train and transport model snapshots as the mules move through physical spaces, sharing these models with the physical 'spaces' the mules inhabit. This method implicitly forms affinity groups among devices associated with users who share particular spaces, enabling collaborative model evolution and protecting users' privacy. Our approach addresses several major shortcomings of traditional, federated, and fully decentralized learning systems. ML Mule represents a new class of machine learning methods that are more robust, distributed, and personalized, bringing the field closer to realizing the original vision of intelligent, adaptive, and genuinely context-aware smart environments. Our results show that ML Mule converges faster and achieves higher model accuracy compared to other existing methods.

Figures

Figures reproduced from arXiv: 2501.07536 by the authors.

Figure 1
Figure 1. Example of ML Mule sharing process 1 Introduction AI-driven technologies are transforming modern life, stream￾lining processes, and embedding themselves into everyday routines. Individuals interact with machine learning models constantly—whether turning on lights with a virtual assis￾tant in the morning, composing emails using ChatGPT at work, tracking calories burned with a smartwatch, or let￾ting a smart thermosta… view at source ↗
Figure 2
Figure 2. Illustration of the two main training modes. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Example random-walk trajectories under three different crossing probabilities [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: CIFAR-100 Data distributions across different partitioning methods. The first subplot show IID distribution. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Accuracy over time for image classification with different methods and crossing probabilities. A higher [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Accuracy over time for image classification with different crossing probabilities and methods. To reduce [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Accuracy over time for human activity recognition with different methods and crossing probabilities. A [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Accuracy over time for human activity recognition with different crossing probabilities and methods. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Timeline of operations for the Mule device [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

  1. [1]

    The Insecurity of Home Digital Voice Assistants -- Amazon Alexa as a Case Study

    X. Lei et al. 2017. The insecurity of home digital voice assistants– amazon alexa as a case study. arXiv preprint arXiv:1712.03327

  2. [2]

    Cho et al

    E. Cho et al. 2020. Will deleting history make alexa more trustworthy? effects of privacy and content customization on user experience of smart speakers. In Proceedings of the 2020 CHI conference on human factors in computing systems , 1–13

  3. [3]

    Huang et al

    X. Huang et al. 2024. Federated learning-empowered ai-generated content in wireless networks. IEEE Network, 38, 5, 304–313

  4. [4]

    Hard et al

    A. Hard et al. 2018. Federated learning for mobile keyboard predic- tion. arXiv preprint arXiv:1811.03604

  5. [5]

    Hegedűs et al

    I. Hegedűs et al. 2019. Gossip learning as a decentralized alternative to federated learning. In Proc. of DAIS, 74–90

  6. [6]

    Lee et al

    S. Lee et al. 2021. Opportunistic federated learning: an exploration of egocentric collaboration for pervasive computing applications. In Proc. of PerCom, 1–8

  7. [7]

    Medjiah et al

    S. Medjiah et al. 2014. Sailing over data mules in delay-tolerant networks. IEEE Transactions on Wireless Communications , 13, 1

  8. [8]

    Krizhevsky et al

    A. Krizhevsky et al. 2009. Learning multiple layers of features from tiny images. Tech. rep. 0. University of Toronto, Toronto, Ontario. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf

Show all 49 references
  1. [9]

    Grauman et al

    K. Grauman et al. 2024. Ego-exo4d: understanding skilled human activity from first-and third-person perspectives. In Proc. of CVPR

  2. [10]

    McMahan et al

    B. McMahan et al. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statis- tics. PMLR, 1273–1282

  3. [11]

    Sattler et al

    F. Sattler et al. 2019. Clustered federated learning: model-agnostic distributed multi-task optimization under privacy constraints. (2019). https://arxiv.org/abs/1910.01991 arXiv: 1910.01991 [cs.LG]

  4. [12]

    Yang et al

    X. Yang et al. 2024. Fedas: bridging inconsistency in personalized federated learning. In Proc. of CVPR. (June 2024), 11986–11995

  5. [13]

    T.-M. H. Hsu et al. 2019. Measuring the effects of non-identical data distribution for federated visual classification. (2019). https://arxiv.o rg/abs/1909.06335 arXiv: 1909.06335 [cs.LG]

  6. [14]

    Fallah et al

    A. Fallah et al. 2020. Personalized federated learning: a meta-learning approach. arXiv preprint arXiv:2002.07948

  7. [15]

    Yi et al

    L. Yi et al. 2023. Pfedes: model heterogeneous personalized federated learning with feature extractor sharing.arXiv preprint arXiv:2311.06879

  8. [16]

    Hanzely et al

    F. Hanzely et al. 2020. Federated learning of a mixture of global and local models. arXiv preprint arXiv:2002.05516

  9. [17]

    Fallah et al

    A. Fallah et al. 2020. Personalized federated learning with theoretical guarantees: a model-agnostic meta-learning approach. In Advances in Neural Information Processing Systems . H. Larochelle et al., (Eds.) Vol. 33. Curran Associates, Inc., 3557–3568

  10. [18]

    J. H. Lim et al. 2024. Metavers: meta-learned versatile representations for personalized federated learning. In Proc of W ACV

  11. [19]

    Kundu et al

    A. Kundu et al. 2022. Robustness and personalization in federated learning: a unified approach via regularization. In Proc. of EDGE

  12. [20]

    Wu et al

    X. Wu et al. 2024. Decoupling general and personalized knowledge in federated learning via additive and low-rank decomposition. In Proc. of MM, 7172–7181

  13. [21]

    Yu et al

    H. Yu et al. 2022. Prototyping opportunistic learning in resource constrained mobile devices. In Proc. of PerCom Workshops, 521–526

  14. [22]

    Messmer et al

    L.-M. Messmer et al. 2024. Context-aware machine learning: a survey. In Proc. of FTC. K. Arai, (Ed.), 252–272

  15. [23]

    M. B. Harries et al. 1998. Extracting hidden context.Machine learning, 32, 2, 101–126

  16. [24]

    I. H. Sarker et al. 2021. Introduction to context-aware machine learn- ing and mobile data analytics. In Context-A ware Machine Learning and Mobile Data Analytics: Automated Rule-based Services with Intel- ligent Decision-Making, 3–13

  17. [25]

    Wang et al

    X. Wang et al. 2023. Context understanding in computer vision: a survey. Computer Vision and Image Understanding , 229

  18. [26]

    Miranda et al

    L. Miranda et al. 2022. A survey on the use of machine learning meth- ods in context-aware middlewares for human activity recognition. Artificial Intelligence Review, 55, 4, 3369–3400

  19. [27]

    Yu et al

    T. Yu et al. 2020. Learning context-aware policies from multiple smart homes via federated multi-task learning. In Proc. of IoTDI, 104–115

  20. [28]

    Huang et al

    G.-L. Huang et al. 2023. Context-aware machine learning for intelli- gent transportation systems: a survey.IEEE Transactions on Intelligent Transportation Systems, 24, 1, 17–36

  21. [29]

    Nascimento et al

    N. Nascimento et al. 2018. A context-aware machine learning-based approach. In Proceedings of the 28th Annual International Conference on Computer Science and Software Engineering , 40–47

  22. [30]

    G. H. Sim et al. 2018. An online context-aware machine learning algorithm for 5g mmwave vehicular communications. IEEE/ACM Transactions on Networking, 26, 6, 2487–2500

  23. [31]

    Liu et al

    R. Liu et al. 2017. Context aware machine learning approaches for modeling elastic localization in three-dimensional composite mi- crostructures. Integrating Materials and Manufacturing Innovation

  24. [32]

    Menik et al

    S. Menik et al. 2023. Towards modular machine learning solution development: benefits and trade-offs. (2023). https://arxiv.org/abs/23 01.09753 arXiv: 2301.09753 [cs.LG]

  25. [33]

    I. H. Sarker et al. 2020. Abc-ruleminer: user behavioral rule-based machine learning method for context-aware intelligent services. Journal of Network and Computer Applications , 168

  26. [34]

    K. Bayoudh. 2024. A survey of multimodal hybrid deep learning for computer vision: architectures, applications, trends, and challenges. Information Fusion, 105

  27. [35]

    Omolaja et al

    A. Omolaja et al. 2022. Context-aware complex human activity recog- nition using hybrid deep learning models. Applied Sciences, 12, 18

  28. [36]

    M. M. Bejani et al. 2018. A context aware system for driving style evaluation by an ensemble learning on smartphone sensors data. Transportation Research Part C: Emerging Technologies , 89, 303–320

  29. [37]

    Wu et al

    C. Wu et al. 2022. Knowledge graph-based multi-context-aware rec- ommendation algorithm. Information Sciences, 595, 179–194. https: //www.sciencedirect.com/science/article/pii/S0020025522001967

  30. [38]

    Brdiczka

    O. Brdiczka. 2019. (Apr. 2019). https://business.adobe.com/blog/pers pectives/contextual-ai-the-next-frontier-of-artificial-intelligence

  31. [39]

    King et al

    E. King et al. 2023. Candor: continuous adaptive neighbor discovery. In Proc. of MASS, 336–342

  32. [40]

    McMahan et al

    B. McMahan et al. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proc. of AISTATS. A. Singh et al., (Eds.) Vol. 54. (Apr. 2017), 1273–1282

  33. [41]

    D. A. E. Acar et al. 2021. Federated learning based on dynamic regu- larization. In Proc. of ICLR

  34. [42]

    S. P. Karimireddy et al. 2020. SCAFFOLD: stochastic controlled aver- aging for federated learning. In Proc. of ICML, 5132–5143

  35. [43]

    Li et al

    T. Li et al. 2020. Federated optimization in heterogeneous networks. (2020). https://arxiv.org/abs/1812.06127 arXiv: 1812.06127 [cs.LG]

  36. [44]

    Yu et al

    H. Yu et al. 2023. Idml: incentivized decentralized machine learning. arXiv preprint arXiv:2304.05354

  37. [45]

    d.] Foursquare

    [n. d.] Foursquare. https://foursquare.com/products/visits/. ()

  38. [46]

    Izmailov et al

    P. Izmailov et al. 2018. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407

  39. [47]

    Xia et al

    K. Xia et al. 2020. Lstm-cnn architecture for human activity recogni- tion. IEEE Access, 8

  40. [48]

    d.] Jetson orin nano

    [n. d.] Jetson orin nano. https://www.nvidia.com/en-us/autonomou s-machines/embedded-systems/jetson-orin/. ()

  41. [49]

    d.] Raspberry pi 5

    [n. d.] Raspberry pi 5. https://www.raspberrypi.com/products/raspb erry-pi-5/. ()

Pith tools

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