Pith. sign in

REVIEW 3 major objections 5 minor 53 references

FedSceneX: Time-to-Target Orchestration for Same-Scene Multimodal Federated Edge Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read One price per client decides who uploads, at what precision, and where, cutting federated round time to 3.31 hours against 4.85 to 9.78 for twelve baselines.

desk verdict A genuinely useful empirical study on time-based evaluation for multimodal FL, but the theoretical pricing story is not established; the deployed utility is hand-calibrated. read the letter →

arxiv 2608.07730 v1 pith:NTXQXT63 submitted 2026-08-07 cs.NI

Dhe Yeong Tchalla , Beining Wu , Jun Huang , Shuyang Gu , Qiang Duan This is my paper · ORCID
classification cs.NI
keywords federatedlearningmultimodaledgeorchestrationclientselectionvalue-per-hourpricingtime-to-targetresourceallocationautonomousdriving
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

FedSceneX claims that the right unit of progress for federated learning at the sensing edge is the active hour, not the communication round, because a round costs what its clients cost and on identical hardware the twelve baselines need 4.85 to 9.78 hours per round while FedSceneX needs 3.31. The paper argues that round composition should be the decision variable, and that one closed-form price per client can decide which clients upload, at what precision, and on which edge server. Counting by hours rather than rounds changes the ranking, and the paper reports that within a twenty-hour budget FedSceneX reaches the highest accuracy across all seeds while keeping all four sensing modalities in every round. The advantage is bounded: after about forty-five hours conventional methods overtake it, because the coarseness that makes rounds cheap also limits late-stage refinement.

What carries the argument

The central object is the per-client utility $U_t^k = \alpha P_t^k + \gamma G_t^k + \mu F_t^k - \delta X_t^k - \eta L_t^k$, produced by applying a parametric transform to the fractional objective $\max V/T$ and dualizing the uplink constraint, with $\delta$ equal to the shadow price of a byte and $\eta$ equal to the worth of an hour. This single scalar carries the whole orchestration: its sign is the participation threshold, its magnitude sets uplink rate and quantization level by reverse water-filling (giving larger updates more bits until every reconstruction reaches a common distortion floor), and it orders clients for least-loaded edge-server placement. The derivation also uses list scheduling to replace the makespan by a separable total with a bounded factor, so the coupled program collapses into one price per client.

What would settle it

A direct test is to apply the same calibrated utility weights to a different multimodal edge-learning benchmark without any retuning; if FedSceneX no longer completes rounds in roughly 3.3 active hours or no longer leads the same baseline set within a twenty-hour budget, the claimed time-to-target advantage is a fitted result rather than a property of the method.

Watch

Extended reading notes

Core claim

The central claim is that round composition, not compression, is the lever that determines how quickly a same-scene multimodal federation reaches a usable model, because the network cost of a round is only 1.99 seconds out of 3.52 hours while local training dominates. FedSceneX therefore maximizes learning value per active hour rather than per round. Value-per-Hour Pricing (VHP) transforms the fractional objective via a parametric transform of fractional programming and dualizes the uplink budget, producing a per-client scalar whose negative weights are the shadow price of a byte and the worth of an hour. That one scalar is read three ways: a positivity threshold selects clients subject to a modality-coverage floor, a reverse water-filling rule sets quantization to a common distortion floor, and a least-loaded rule places updates on edge servers with a bounded makespan loss. On the full fifteen-client benchmark with twelve baselines, a FedSceneX round costs 3.31 active hours, every seed leads within a twenty-hour budget, and the crossover happens between forty-five and fifty hours.

Load-bearing premise

The reported advantage rests on the assumption that the utility weights and the value-of-an-hour parameter, calibrated once rather than re-derived each round, are representative enough to preserve the speed and accuracy lead on other same-scene multimodal platforms and datasets.

Editorial extensions

If this is right

  • Round-based convergence curves misreport progress by a factor of 1.5 to 3 on identical hardware, so evaluations should be drawn against cumulative active time or time-to-target instead of round index.
  • A deployment with a fixed-hour budget should price clients by one utility rather than tune separate selection, compression, and placement thresholds.
  • All four sensing modalities can be held in every round while uploading about 2.9 MB per round, less than baselines that drop a modality.
  • The speed advantage is bounded: for training budgets beyond about forty-five hours, conventional heterogeneity-aware optimizers catch up and pass FedSceneX.

Reading between the lines

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

  • Because workload-pricing telemetry reveals modality with perfect accuracy, any orchestrator that prices clients by their workload inherits a privacy leak; this exposure should be budgeted as a cost independently of FedSceneX.
  • Updating the value-of-an-hour parameter by the parametric root-finding each round, instead of holding it constant, is a natural extension the paper does not test; it could either recover the late-stage accuracy or erase the speed advantage.
  • The reported crossover at forty-five hours is likely specific to the dataset and encoder choices; on tasks with stronger fine-grained refinement needs or higher quantization floors, the crossover should shift earlier.
  • Since network cost is about 0.016% of round time, communication-compression methods are roughly orthogonal to the bottleneck; the larger gains for same-scene multimodal edges lie in scheduling local compute, not in shrinking payloads.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. FedSceneX is an orchestration layer for same-scene multimodal federated edge learning. It selects clients, quantizes their updates, and places them on edge servers so as to maximize learning value per active hour, subject to an uplink budget and a modality-coverage floor. The claimed technical novelty, Value-per-Hour Pricing (VHP), converts the fractional program via Dinkelbach and dualization into a closed-form per-client utility whose weights are resource shadow prices. Experiments on the nuScenes benchmark with fifteen clients and twelve baselines report a per-round cost of 3.31 ± 0.13 active hours (baselines 4.85–9.78 h), highest Macro-F1 within a 20-hour budget, and a crossover beyond 45 hours; the paper also reports a modality-inference privacy leak from the server-visible scalars.

Significance. The paper addresses a real and under-appreciated issue (rounds are not equal work) and proposes a concrete budget-aware objective with coverage constraints. The empirical evaluation is broad (12 baselines, full nuScenes, five seeds for the proposed method) and unusually honest about the crossover region and the privacy cost of orchestration. If the central derivation were sound and the calibrated constants transparently reported, the paper would be a useful contribution to client selection and resource allocation in federated edge learning. As it stands, however, the derivation of the utility is not what is implemented, the makespan surrogate is not equivalent to the stated objective, and the headline comparison lacks baseline variance, so the core claims are not yet established.

major comments (3)
  1. [Section IV-C, Eq. (4), Proposition 1] The derivation presented for the utility U_k is not what the implementation uses. In Eq. (4) the positive terms αP+γG+µF are asserted as 'expected worth' without any derivation from V(Γ) in Eq. (2); the proposition therefore defines V rather than deriving a price from it. The paper then states that κ is 'calibrated once' and 'the weights appear as constants' (Section IV-C). Consequently the implementation is a hand-weighted linear score with five constants (plus κ, λ, ρ) whose values are never reported. This contradicts the abstract's claim of a closed-form price and leaves the central claim that one utility replaces hand-tuned thresholds unsupported. Please either solve/update the Dinkelbach and dual parameters per round and report them, or present Eq. (4) as a learned/calibrated score and provide the calibration procedure, values, and a sensitivity analysis.
  2. [Section IV-B/C, Eq. (2)/(3)] The reduction from T to total work via Graham's list-scheduling bound is invalid as stated. Graham's bound bounds makespan of a given assignment by a factor times the optimal makespan; it does not imply that the ratio V / sum_k L_k is within a constant of V / makespan. With two servers and two one-hour jobs on separate servers, sum L = 2 h and makespan = 1 h, while a single two-hour job also has sum L = 2 h but makespan = 2 h; the surrogate and the true objective disagree by a factor of 2 on equal total work. Therefore the term η L_k in Eq. (4) is a price on total work, not a shadow price of the round's critical path, and the claim that η is 'the worth of an hour' is not established. This is load-bearing because the entire VHP decomposition depends on separability of T.
  3. [Section V-B, Table IV, Fig. 3(c)] All twelve baselines are reported as single runs while FedSceneX has five seeds. At the 20-hour budget the weakest FedSceneX seed is 0.6486 vs. the strongest baseline 0.6458; the margin 0.0028 is almost 70% of FedSceneX's own between-seed standard deviation at 20 h (0.0040). Without baseline variance it is impossible to know whether the claimed lead across random seeds is significant. Similarly the crossover at 42.5–45 h and the Table V ablation comparisons are based on single-run baselines/variants. Please provide multiple seeds for the main baselines (or at least error bars on the budget curves) and statistically test the 20-hour claim.
minor comments (5)
  1. [Section IV-D, Eqs. (5)–(6)] The symbol b_t^k is used both for the precision/rate variable in the contract (Eq. (5)) and for the uplink rate allocation in Eq. (6); rename one of them to avoid confusion.
  2. [Section IV-C, Table III] Table III lists κ, λ, ρ, and D^* but no values are reported anywhere; without these values (or the calibration protocol) the method is not reproducible.
  3. [Section V-A] FedSceneX grants local epochs n_t^k in proportion to utility, while the baselines are said to use the same local budget; justify that this does not bias the per-round time comparison.
  4. [Section V-E] The modality-inference attackers are described only as 'logistic regression' and 'random forest'; specify the exact scalar features, normalization, and hyperparameters used so the measurement can be reproduced.
  5. [Section IV-D, Eq. (6)] In Eq. (6), if all U_j^+ are zero the denominator is zero; state the fallback allocation for this edge case.

Circularity Check

2 steps flagged · score 6.0 of 10

VHP's price is not derived: α,γ,μ are an assumed linear 'worth' and κ is calibrated once, so the headline 20–45 h lead is a fitted result rather than a closed-form shadow price.

  1. self definitional [Section IV-C, Proposition 1 and Eq. (4)]
    "Proof: Collect what one selected client contributes to V−κT: its expected worth αP_t_k + γG_t_k + µF_t_k and the time κL_t_k it adds to the round. ... U_t_k = αP_t_k + γG_t_k + µF_t_k − δX_t_k − ηL_t_k, where δ=λ and η=κ."

    The 'expected worth' of a client is simply declared to be the same linear combination αP+γG+μF; V is never given any other definition anywhere in the paper. The per-client Lagrangian term in Proposition 1 is therefore the initial weighted-sum ansatz restated, with only the δX and ηL terms coming from the dualization and the Dinkelbach parameter. The paper then asserts that Eq. (4) 'is therefore not a weighted sum chosen by hand,' but its own proof shows that α, γ, and μ are free weights, not dual variables or quantities solved from V. Thus the claimed 'closed-form client price' is, by construction, the hand-weighted score that was put into the derivation, and the claim that one utility replaces hand-tuned thresholds is not established by the optimization.

  2. fitted input called prediction [Section IV-C, paragraph after Eq. (4)]
    "The implementation holds κ at a value calibrated once instead of running a Dinkelbach update every round, which is why the weights appear as constants."

    κ is the η weight in the utility, and α, γ, μ, δ are the other weights in Eq. (4); the paper reports neither the calibrated values nor a separate calibration protocol or dataset split. The headline results—3.31 active hours per round and the highest Macro-F1 within a 10–45 h budget—are produced by this U with calibrated constants, so the 'prediction' of the time-budget lead is statistically indistinguishable from an in-sample fit if the calibration used the nuScenes setting. The paper presents these constants as shadow prices, but a value calibrated once off-line and never reported is a fitted parameter whose fitted nature is then read back as the experimental advantage of the pricing mechanism.

full rationale

The central VHP derivation is partially circular at the point where the per-client price is formed. Proposition 1 defines a client's contribution to V−κT as αP+γG+μF, and Eq. (4) is exactly that weighted sum plus the dual terms; the positive weights are never derived from V, so the 'closed-form price' reduces to the hand-weighted score the paper claims to avoid. In addition, κ (the η weight) is 'calibrated once' with no reported value or calibration split, so the headline 20–45 h accuracy lead may be a fitted result rather than a demonstration of the pricing mechanism. I do not count the many self-citations as circular: they appear in related work and background, and none is load-bearing for the VHP derivation. The list-scheduling replacement of makespan by total work is a separate correctness concern (the two objectives can disagree on the same total work), but it is not a circularity and is not scored here. Because the central one-number-price claim reduces, by the paper's own equations and calibration statement, to a fitted linear score, the partial-circularity score is 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a parametrized scoring rule with unreported calibrated constants, on standard but unverified quantization and loss-sensitivity models, and on the assumption that telemetry scalars are privacy-safe, which the paper's own attack contradicts.

free parameters (4)
  • Utility weights alpha, gamma, mu = not reported
    Equation (4) uses three positive weights for local progress, update magnitude, and fairness; the text says weights carry relative importance but no values or fitting procedure are given, and they are held constant in the implementation.
  • kappa (value of an hour) = not reported, calibrated once
    Section IV-C: The implementation holds kappa at a value calibrated once instead of running a Dinkelbach update every round. This constant sets the trade-off between value and time and directly affects which clients are selected.
  • lambda (shadow price of a byte) = not reported
    Lambda is the Lagrange multiplier on the uplink budget and controls delta in Equation (4) and the distortion floor D* in Proposition 3; its value is not reported.
  • rho (loss sensitivity) = not reported
    Proposition 3 assumes reconstruction error raises the expected loss at a constant rate rho and sets D* = lambda / (16 rho ln 2); rho is not measured or reported.
assumptions (5)
  • standard math Dinkelbach parametric transform converts the fractional objective max V/T into a sequence of linear problems (Equation 3)
    Section IV-B cites Dinkelbach [52]; this is a standard result in fractional programming.
  • standard math List scheduling makespan is within 2 - 1/|E| of the optimum, so T can be replaced by total time at bounded loss
    Section IV-B cites Graham [53]; used to make the makespan separable per client.
  • domain assumption Quantization obeys the high-resolution law E||Delta_theta_hat - Delta_theta||^2 = d_s sigma^2 2^{-2z}
    Section IV-D, Proposition 3; standard in quantization theory but not validated on the actual model updates.
  • domain assumption Reconstruction error raises expected global loss at a constant rate rho
    Section IV-D, Proposition 3; no empirical support for the linear loss-sensitivity model.
  • domain assumption Server-visible telemetry scalars P, G, F, X, L are privacy-safe
    Section III-A states data locality restricts the controller to scalars; Section V-E then shows logistic regression recovers client modality with 1.00 accuracy from these scalars, contradicting the privacy-safety assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedSceneX: Time-to-Target Orchestration for Same-Scene Multimodal Federated Edge Learning." pith.science (2026). https://pith.science/paper/NTXQXT63

@misc{pith2026260807730,
  author       = {Pith},
  title        = {Pith review of: FedSceneX: Time-to-Target Orchestration for Same-Scene Multimodal Federated Edge Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTXQXT63}},
  note         = {Machine review of arXiv:2608.07730}
}
read the original abstract

Federated learning at the sensing edge is typically evaluated by communication rounds, yet a round does not represent a fixed amount of work. Even on identical hardware, the methods we compare require 3.3 to 9.8 hours per round, which makes round-based comparisons misleading. The problem is more obvious for same-scene multimodal clients, since camera, video, LiDAR, and radar workloads differ substantially in training and communication cost, while existing methods treat the modality composition of each round as fixed. To address it, we introduce FedSceneX, an orchestrator that jointly determines round composition to maximize learning value per active hour. The optimization method, Value-per-Hour Pricing (VHP), converts the fractional objective through a parametric transformation and dualizes the uplink constraint, yielding a closed-form client price whose weights capture resource shadow costs. Based on these prices, FedSceneX selects clients subject to a modality coverage constraint, allocates precision through reverse water filling, and assigns updates to edge servers. On the full nuScenes benchmark with fifteen clients and twelve baselines, FedSceneX reduces the active time per round to 3.31 hours, compared with 4.85 to 9.78 hours for the baselines. Across all random seeds, it achieves the highest accuracy within a twenty-hour budget while preserving all four modalities. Its advantage persists from ten to forty-five hours, after which conventional methods overtake it.

Figures

Figures reproduced from arXiv: 2608.07730 by the authors.

Figure 1
Figure 1. (a) Active hours per round, FedSceneX over five seeds. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of FedSceneX. Private encoders stay on the device; one utility over privacy-safe scalars fixes participation, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The same trajectories read four ways. (c) shades one standard deviation over five seeds and marks the crossover; (d) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 32 canonical work pages

  1. [1]

    Communication-Efficient Learning of Deep Networks from Decentralized Data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. Ag ¨uera y Arcas, “Communication-Efficient Learning of Deep Networks from Decentralized Data,” inProc. AISTATS, 2017

    work page 2017

  2. [2]

    Advances and Open Problems in Federated Learning,

    P. Kairouzet al., “Advances and Open Problems in Federated Learning,” Foundations and Trends in Machine Learning, vol. 14, no. 1–2, pp. 1– 210, 2021

    work page 2021

  3. [3]

    Combining Fed- erated Learning and Edge Computing Toward Ubiquitous Intelligence in 6G Network: Challenges, Recent Advances, and Future Directions,

    Q. Duan, J. Huang, S. Hu, R. Deng, Z. Lu, and S. Yu, “Combining Fed- erated Learning and Edge Computing Toward Ubiquitous Intelligence in 6G Network: Challenges, Recent Advances, and Future Directions,” IEEE Communications Surveys & Tutorials, vol. 25, no. 4, pp. 2892– 2950, 2023

    work page 2023

  4. [4]

    FedTD3: An Accelerated Learning Approach for UA V Trajectory Planning,

    B. Wu, J. Huang, and Q. Duan, “FedTD3: An Accelerated Learning Approach for UA V Trajectory Planning,” inProc. WASA, 2025, pp. 13– 24

    work page 2025

  5. [5]

    Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory

    B. Wu, Z. Ding, J. Huang, and Y . Zhao, “Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory,” arXiv preprint arXiv:2606.25115, 2026

    work page Pith review arXiv 2026

  6. [6]

    CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization

    B. Wu and J. Huang, “CrystalMem: Elastic Memory for Self- Evolving LLM Agents via Knowledge Crystallization,” arXiv preprint arXiv:2608.00303, 2026

    work page Pith review arXiv 2026

  7. [7]

    FedMultimodal: A Benchmark for Multimodal Federated Learning,

    T. Feng, D. Bose, T. Zhang, R. Hebbar, A. Ramakrishna, R. Gupta, M. Zhang, S. Avestimehr, and S. Narayanan, “FedMultimodal: A Benchmark for Multimodal Federated Learning,” arXiv preprint arXiv:2306.09486, 2023

    arXiv 2023

  8. [8]

    Towards Multi-Modal Transformers in Federated Learning,

    G. Sun, Y . Cong, J. Dong, Q. Wang, and J. Liu, “Towards Multi-Modal Transformers in Federated Learning,” inProc. ECCV, 2024

    work page 2024

Show all 53 references
  1. [9]

    R- ACP: Real-Time Adaptive Collaborative Perception Leveraging Robust Task-Oriented Communications,

    Z. Fang, J. Wang, Y . Ma, Y . Tao, Y . Deng, X. Chen, and Y . Fang, “R- ACP: Real-Time Adaptive Collaborative Perception Leveraging Robust Task-Oriented Communications,”IEEE Journal on Selected Areas in Communications, 2025

    2025

  2. [10]

    Shared Spatial Memory Through Predictive Coding,

    Z. Fang, Y . Guo, J. Wang, Y . Zhang, H. An, Y . Wang, and Y . Fang, “Shared Spatial Memory Through Predictive Coding,” arXiv preprint arXiv:2511.04235, 2025

    2025

  3. [11]

    ST-Hybrid: Dynamic Graph Learning with Multi- Scale Spatio-Temporal Attention for Traffic Forecasting,

    D. Y . Tchalla, “ST-Hybrid: Dynamic Graph Learning with Multi- Scale Spatio-Temporal Attention for Traffic Forecasting,”ACM SIGAPP Applied Computing Review, vol. 25, no. 4, pp. 35–52, 2026

    2026

  4. [12]

    PRISM: Exposing and Resolving Spurious Isolation in Federated Multimodal Continual Learning,

    B. Wu, Z. Ding, and J. Huang, “PRISM: Exposing and Resolving Spurious Isolation in Federated Multimodal Continual Learning,” arXiv preprint arXiv:2605.01061, 2026

    2026 arXiv

  5. [13]

    Federated Learning Based on Dynamic Regularization,

    D. A. E. Acar, Y . Zhao, R. Matas, M. Mattina, P. Whatmough, and V . Saligrama, “Federated Learning Based on Dynamic Regularization,” inProc. ICLR, 2021

    2021

  6. [14]

    FedADMM: A Robust Federated Deep Learning Framework with Adaptivity to System Heterogeneity,

    Y . Gong, Y . Li, and N. M. Freris, “FedADMM: A Robust Federated Deep Learning Framework with Adaptivity to System Heterogeneity,” inProc. IEEE ICDE, 2022, pp. 2575–2587

    2022

  7. [15]

    Adaptive Federated Optimization,

    S. J. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Kone ˇcn´y, S. Kumar, and H. B. McMahan, “Adaptive Federated Optimization,” in Proc. ICLR, 2021

    2021

  8. [16]

    FedBABU: Toward Enhanced Represen- tation for Federated Image Classification,

    J. Oh, S. Kim, and S.-Y . Yun, “FedBABU: Toward Enhanced Represen- tation for Federated Image Classification,” inProc. ICLR, 2022

    2022

  9. [17]

    Model-Contrastive Federated Learning,

    Q. Li, B. He, and D. Song, “Model-Contrastive Federated Learning,” in Proc. IEEE/CVF CVPR, 2021, pp. 10 713–10 722

    2021

  10. [18]

    A Dual- Level Game-Theoretic Approach for Collaborative Learning in UA V- Assisted Heterogeneous Vehicle Networks,

    Z. Ding, J. Huang, Q. Duan, C. Zhang, Y . Zhao, and S. Gu, “A Dual- Level Game-Theoretic Approach for Collaborative Learning in UA V- Assisted Heterogeneous Vehicle Networks,” inProc. IEEE IPCCC, 2025, pp. 1–8

    2025

  11. [19]

    Combating Knowledge Diversity and Catastrophic Forgetting in UA V-Assisted Collaborative Vehicular Learning: A Game-Theoretic Approach,

    Z. Ding, J. Huang, Y . Zhao, and Z. Cai, “Combating Knowledge Diversity and Catastrophic Forgetting in UA V-Assisted Collaborative Vehicular Learning: A Game-Theoretic Approach,”ACM Transactions on Autonomous and Adaptive Systems, 2026

    2026

  12. [20]

    Learning to Defend: A Multi-Agent Reinforcement Learning Framework for Stackelberg Security Game in Mobile Edge Computing,

    Z. Ding, J. Huang, and J. Qi, “Learning to Defend: A Multi-Agent Reinforcement Learning Framework for Stackelberg Security Game in Mobile Edge Computing,” inProc. IEEE ICNC, 2026

    2026

  13. [21]

    Model-Free Cooperative Optimal Output Regulation for Linear Discrete-Time Multi-Agent Systems Using Reinforcement Learning,

    B. Wu and W. Wu, “Model-Free Cooperative Optimal Output Regulation for Linear Discrete-Time Multi-Agent Systems Using Reinforcement Learning,”Mathematical Problems in Engineering, vol. 2023, no. 1, p. 6350647, 2023

    2023

  14. [22]

    FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization,

    A. Reisizadeh, A. Mokhtari, H. Hassani, A. Jadbabaie, and R. Pedarsani, “FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization,” inProc. AISTATS, 2020

    2020

  15. [23]

    Federated Learning: Strategies for Improving Communica- tion Efficiency,

    J. Kone ˇcn´y, H. B. McMahan, F. X. Yu, P. Richt ´arik, A. T. Suresh, and D. Bacon, “Federated Learning: Strategies for Improving Communica- tion Efficiency,” arXiv preprint arXiv:1610.05492, 2016

    2016 arXiv

  16. [24]

    Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge,

    T. Nishio and R. Yonetani, “Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge,” inProc. IEEE ICC, 2019

    2019

  17. [25]

    Enhancing Vehic- ular Platooning With Wireless Federated Learning: A Resource-Aware Control Framework,

    B. Wu, J. Huang, Q. Duan, L. Dong, and Z. Cai, “Enhancing Vehic- ular Platooning With Wireless Federated Learning: A Resource-Aware Control Framework,”IEEE/ACM Transactions on Networking, 2025

    2025

  18. [26]

    A Fast UA V Tra- jectory Planning Framework in RIS-Assisted Communication Systems With Accelerated Learning via Multithreading and Federating,

    J. Huang, B. Wu, Q. Duan, L. Dong, and S. Yu, “A Fast UA V Tra- jectory Planning Framework in RIS-Assisted Communication Systems With Accelerated Learning via Multithreading and Federating,”IEEE Transactions on Mobile Computing, pp. 1–16, 2025

    2025

  19. [27]

    ‘X of Information’ Continuum: A Survey on AI-Driven Multi-Dimensional Metrics for Next-Generation Net- worked Systems,

    B. Wu, J. Huang, and S. Yu, “‘X of Information’ Continuum: A Survey on AI-Driven Multi-Dimensional Metrics for Next-Generation Net- worked Systems,”IEEE Communications Surveys & Tutorials, vol. 28, pp. 5307–5344, 2026

    2026

  20. [28]

    AoI-Aware Resource Management for Smart Health via Deep Reinforcement Learning,

    B. Wu, Z. Cai, W. Wu, and X. Yin, “AoI-Aware Resource Management for Smart Health via Deep Reinforcement Learning,”IEEE Access, 2023

    2023

  21. [29]

    Transformer-Based Dynamic Resource Allocation for Multi-Carrier NOMA Systems,

    L. Dong, J. Huang, and R. W. Heath, “Transformer-Based Dynamic Resource Allocation for Multi-Carrier NOMA Systems,”IEEE Transac- tions on Cognitive Communications and Networking, vol. 12, pp. 4926– 4941, 2026

    2026

  22. [30]

    Securing Smart Agriculture with Communication-Efficient Federated Unlearning,

    U. Pudasaini, Z. Ding, and J. Huang, “Securing Smart Agriculture with Communication-Efficient Federated Unlearning,” inProc. IEEE HPSR, 2026, pp. 1–8

    2026

  23. [31]

    Reinforcement Learning- Based Energy-Aware Coverage Path Planning for Precision Agriculture,

    B. Wu, Z. Ding, L. Ostigaard, and J. Huang, “Reinforcement Learning- Based Energy-Aware Coverage Path Planning for Precision Agriculture,” inProc. ACM RACS, 2025, pp. 1–8

    2025

  24. [32]

    A Stochastic Geometry-Based Analysis of SWIPT-Assisted Underlaid Device-to-Device Energy Har- vesting,

    C.-C. Xing, Z. Ding, and J. Huang, “A Stochastic Geometry-Based Analysis of SWIPT-Assisted Underlaid Device-to-Device Energy Har- vesting,”ACM SIGAPP Applied Computing Review, vol. 25, no. 4, pp. 18–34, 2026

    2026

  25. [33]

    A Fault-Tolerant and Energy-Efficient Design of a Network Switch Based on a Quantum- Based Nano-Communication Technique,

    D. Pan, B.-N. Wu, Y .-L. Sun, and Y .-P. Xu, “A Fault-Tolerant and Energy-Efficient Design of a Network Switch Based on a Quantum- Based Nano-Communication Technique,”Sustainable Computing: In- formatics and Systems, vol. 37, p. 100827, 2023

    2023

  26. [34]

    Inference-Time Budget Control for LLM Search Agents,

    Z. Fang, S. F. Hu, Z. Chang, Y . Guo, Y . Tao, H. Liu, M. Ruan, J. Huang, and Y . Fang, “Inference-Time Budget Control for LLM Search Agents,” arXiv preprint arXiv:2605.05701, 2026

    2026 arXiv

  27. [35]

    FedMFS: Federated Multimodal Fusion Learning with Selective Modal- ity Communication,

    L. Yuan, D.-J. Han, V . P. Chellapandi, S. H. ˙Zak, and C. G. Brinton, “FedMFS: Federated Multimodal Fusion Learning with Selective Modal- ity Communication,” arXiv preprint arXiv:2310.07048, 2023

    2023 arXiv

  28. [36]

    RELIEF: Turning Missing Modalities into Training Acceleration for Federated Learning on Heterogeneous IoT Edge,

    B. Wu, Z. Ding, and J. Huang, “RELIEF: Turning Missing Modalities into Training Acceleration for Federated Learning on Heterogeneous IoT Edge,” arXiv preprint arXiv:2604.04243, 2026

    2026 arXiv

  29. [37]

    EASE: Federated Multimodal Un- learning via Entanglement-Aware Anchor Closure,

    Z. Ding, B. Wu, and J. Huang, “EASE: Federated Multimodal Un- learning via Entanglement-Aware Anchor Closure,” arXiv preprint arXiv:2605.00733, 2026

    2026 arXiv

  30. [38]

    Oort: Efficient Federated Learning via Guided Participant Selection,

    F. Lai, X. Zhu, H. V . Madhyastha, and M. Chowdhury, “Oort: Efficient Federated Learning via Guided Participant Selection,” inProc. USENIX OSDI, 2021

    2021

  31. [39]

    Federated Optimization in Heterogeneous Networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated Optimization in Heterogeneous Networks,” inProc. MLSys, 2020

    2020

  32. [40]

    Elastic Aggregation for Federated Optimization,

    D. Chen, J. Hu, and V . J. Tan, “Elastic Aggregation for Federated Optimization,” inProc. IEEE/CVF CVPR, 2023, pp. 12 187–12 197

    2023

  33. [41]

    FedBN: Federated Learning on Non-IID Features via Local Batch Normalization,

    X. Li, M. Jiang, X. Zhang, M. Kamp, and Q. Dou, “FedBN: Federated Learning on Non-IID Features via Local Batch Normalization,” inProc. ICLR, 2021

    2021

  34. [42]

    Exploiting Shared Representations for Personalized Federated Learning,

    L. Collins, H. Hassani, A. Mokhtari, and S. Shakkottai, “Exploiting Shared Representations for Personalized Federated Learning,” inProc. ICML, 2021, pp. 2089–2099

    2021

  35. [43]

    Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relation- ships,

    Y . Guo, K. Guo, X. Cao, T. Wu, and Y . Chang, “Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relation- ships,” inProc. ICML, 2023, pp. 11 905–11 933

    2023

  36. [44]

    FedSR: A Simple and Effec- tive Domain Generalization Method for Federated Learning,

    A. T. Nguyen, P. Torr, and S. N. Lim, “FedSR: A Simple and Effec- tive Domain Generalization Method for Federated Learning,” inProc. NeurIPS, 2022

    2022

  37. [45]

    A Review of Continual Learning in Edge AI,

    B. Wu, Z. Ding, and J. Huang, “A Review of Continual Learning in Edge AI,”IEEE Transactions on Network Science and Engineering, vol. 13, pp. 6571–6588, 2026

    2026

  38. [46]

    Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systems,

    B. Wu and J. Huang, “Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systems,” arXiv preprint arXiv:2604.20745, 2026

    2026 arXiv

  39. [47]

    From Alpha to Omega: Lifecycle- Aware Forgetting Defense in Federated Continual Learning for Planetary Exploration,

    B. Wu, J. Huang, and Y . Zhao, “From Alpha to Omega: Lifecycle- Aware Forgetting Defense in Federated Continual Learning for Planetary Exploration,” inProc. IEEE ICDCS, 2026

    2026

  40. [48]

    nuScenes: A Multimodal Dataset for Autonomous Driving,

    H. Caesaret al., “nuScenes: A Multimodal Dataset for Autonomous Driving,” inProc. IEEE/CVF CVPR, 2020

    2020

  41. [49]

    Real-Time Intelligent Healthcare Enabled by Federated Digital Twins With AoI Optimization,

    B. Wu, J. Huang, and Q. Duan, “Real-Time Intelligent Healthcare Enabled by Federated Digital Twins With AoI Optimization,”IEEE Network, vol. 40, no. 2, pp. 184–191, 2026

    2026

  42. [50]

    Application-Aware Twin-in- the-Loop Planning for Federated Split Learning over Wireless Edge Networks,

    Z. Ding, B. Wu, J. Huang, and S. Mao, “Application-Aware Twin-in- the-Loop Planning for Federated Split Learning over Wireless Edge Networks,” arXiv preprint arXiv:2604.26105, 2026

    2026 arXiv

  43. [51]

    SCALE: Sensitivity-Aware Federated Unlearning with Information Freshness Optimization for Mobile Edge Computing,

    Z. Ding, B. Wu, and J. Huang, “SCALE: Sensitivity-Aware Federated Unlearning with Information Freshness Optimization for Mobile Edge Computing,” inProc. IEEE ICDCS, 2026

    2026

  44. [52]

    On Nonlinear Fractional Programming,

    W. Dinkelbach, “On Nonlinear Fractional Programming,”Management Science, vol. 13, no. 7, pp. 492–498, 1967

    1967

  45. [53]

    Bounds for Certain Multiprocessing Anomalies,

    R. L. Graham, “Bounds for Certain Multiprocessing Anomalies,”Bell System Technical Journal, vol. 45, no. 9, pp. 1563–1581, 1966

    1966

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.