Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Scheduler cuts federated edge learning rounds by 58 percent

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 22:20 UTC pith:WMH4KXFL

load-bearing objection Useful temporal-drift idea for wireless FL scheduling, but the L(c) estimator makes the algorithm unimplementable as described. the 4 major comments →

arxiv 2509.07342 v1 pith:WMH4KXFL submitted 2025-09-09 cs.LG cs.DC

FedTeddi: Temporal Drift and Divergence Aware Scheduling for Timely Federated Edge Learning

Yuxuan Bai , Yuxuan Sun , Tan Chen , Wei Chen , Sheng Zhou , Zhisheng Niu This is my paper
classification cs.LG cs.DC
keywords federated edge learningclient schedulingtemporal driftcollective divergenceclass distribution shiftcontinual learningbandwidth allocationEarth Mover's Distance
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.

FedTeddi addresses a gap in federated edge learning: most client-scheduling algorithms assume the data on each device are fixed, but real edge devices keep collecting data whose distribution shifts over time. The paper argues that the right scheduling objective for such streaming settings has three parts: schedule enough clients to reduce sampling noise, choose a set whose combined data mirrors the global class distribution, and early in each frame give priority to clients whose data changed most, with that priority decaying to protect old knowledge. It derives closed-form bounds for the first two quantities as weighted Earth Mover's Distances between class-proportion vectors, folds them with the drift term into a greedy client-selection and bandwidth-allocation algorithm, and shows on CIFAR-10 and CIFAR-100 that the result reaches target accuracy in up to 58% fewer rounds than random scheduling and in fewer rounds than seven baselines. A sympathetic reader would care because it offers a practical way to keep federated models timely under data evolution without exchanging raw data or forcing the model to choose between new knowledge and old.

Core claim

The central claim is that data evolution in federated edge learning can be scheduled explicitly, not just reacted to. Define temporal drift as the change in a client's local gradient between two frames, and collective divergence as the gap between the aggregated gradient of the scheduled clients and the global gradient; for classification, both are bounded by Earth Mover's Distance between class-proportion vectors weighted by per-class gradient norms. The paper's optimization objective minimizes sampling variance plus collective divergence while subtracting a decaying-weighted drift term, and a greedy algorithm adds clients one by one until the bandwidth budget is exhausted. The experiments

What carries the argument

Temporal drift and collective divergence, both expressed as Earth Mover's Distance over class-proportion vectors, are the two quantities that carry the argument. The decaying drift weight lambda_k = lambda_0(1 - k/K_l) shifts the scheduler from plasticity early in a frame to stability late in the frame, and the greedy marginal-gain rule decides when adding a client still reduces the joint objective. The Lambert-W bandwidth formula converts each client's channel state and the delay budget into a minimum bandwidth requirement, which is what lets scheduling and bandwidth allocation be optimized together.

Load-bearing premise

The central premise is that data evolution is fully captured by changes in class proportions: per-class expected gradient norms are treated as identical over time and across clients, so within-class feature drift is invisible to the scheduler.

What would settle it

Run the same experiment but keep class proportions fixed while rotating image styles per class (for example, different fonts or lighting) across frames; if FedTeddi and random scheduling then converge at the same rate, the drift term is measuring the wrong quantity. Alternatively, record the actual per-client gradient change between frames and check whether the clients FedTeddi selects for high temporal drift are truly the ones with the largest gradient changes.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If FedTeddi is right, federated systems facing class or concept drift can keep models current without replaying or sharing data, simply by choosing which clients to listen to at each round.
  • The decaying drift weight gives system designers a single lever for the plasticity-stability trade-off: raising or lowering lambda_0 and reshaping the decay schedule should directly trade early adaptivity against late forgetting.
  • The EMD formulation means the scheduler needs only class-proportion metadata, not raw data or gradient norms, to make its decisions, which is cheap to report and privacy-preserving to aggregate.
  • Under tight per-round latency deadlines the gains are largest, because selective scheduling matters most when only a few clients can participate in each aggregation.
  • The same joint objective can be reused with any replay or memory-selection strategy, since the paper treats representative-sample selection as orthogonal to scheduling.

Where Pith is reading between the lines

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

  • The method assumes per-class gradient norms are time-invariant, so under within-class domain shift (same class, new style or appearance) its drift measure would undercount the real change; tracking per-class gradient norms over time would be a natural extension.
  • The estimator for per-class gradient norms is derived for single-class clients and applied to multi-class clients heuristically; a testable extension is to compare this heuristic against direct per-class gradient estimation on multi-class clients.
  • The three-term objective could transfer to non-classification tasks by replacing EMD with a gradient-space distance, provided per-client gradient estimates are available.
  • The fixed linear decay of the drift weight might be suboptimal; an adaptive lambda based on measured drift magnitude could improve the plasticity-stability trade-off further.

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 FedTeddi, a client scheduling and bandwidth allocation algorithm for federated edge learning (FEEL) under temporally evolving, non-i.i.d. data. The authors define 'temporal drift' and 'collective divergence' via gradient-norm bounds, express these as Earth Mover's Distance between class proportions for classification tasks, and formulate a scheduling objective (P2/P3) that trades off sampling variance, collective divergence, and a decaying temporal-drift incentive. A greedy algorithm with per-round delay and bandwidth constraints is presented. Experiments on CIFAR-10 and CIFAR-100 report faster convergence and higher accuracy than several baselines, e.g., 58.4% fewer rounds to reach a target accuracy on CIFAR-10 compared with random scheduling.

Significance. The problem addressed is timely and relevant: FEEL with continuously evolving client data is an important practical setting, and a scheduling method that explicitly accounts for data drift while respecting wireless resource constraints would be valuable. The empirical study is substantial: it compares against seven baselines, considers three delay budgets, and reports both accuracy and convergence-round metrics. The paper also makes the bandwidth-allocation component explicit, which is a strength. However, the theoretical foundation is not established within the manuscript: Theorem 1 is stated without proof and deferred to a self-cited preprint, and the optimization objective is introduced heuristically rather than derived from the convergence bound. More seriously, the estimator for the per-class gradient norm L^(c) in Eq. (18) cannot be computed from the information the server receives in Algorithm 1. Since this estimator feeds directly into the scheduling objective P2/P3, the algorithm as presented is not implementable in the claimed federated setting. These are load-bearing issues, not presentation concerns.

major comments (4)
  1. [Section IV-C and Theorem 1 (Eq. (12))] Theorem 1 is asserted with 'We omit the proof due to space limit' and is deferred to a self-cited preprint [29]. The transformation from P1 to P2 (Eq. (13)) is not derived from the bound in Eq. (12); the temporal-drift term and the decaying weight lambda_k are introduced via Remark 3 as a heuristic plasticity-stability trade-off. Since P2/P3 is the core of the proposed method, the central claim that FedTeddi is 'convergence-aware' is not supported by a derivation in this paper.
  2. [Section IV-D3] The estimator of L^(c) in Eq. (18) requires per-class local and global gradients, i.e., \nabla F_{n,l}(w, D_{n,l}^{(c)}) and \nabla F_l(w, D_l^{(c)}). But in Algorithm 1, before scheduling, clients send only class proportions p_{n,l,k} and gradient variances (lines 9–10); local models are transmitted only by already-scheduled clients (lines 25–27). The server never receives D_{n,l}^{(c)}, D_l^{(c)}, or per-class gradients, so it cannot compute Eq. (18). The paper itself acknowledges in Section IV-D3 that 'we cannot collect all data in a single round to compute the global loss and gradient,' yet the estimator requires exactly that. Consequently, the scheduling objective P2/P3 and the selection condition (17) are undefined in the proposed federated protocol. If the experiments used a server-side oracle for global per-class gradients, this must be stated; otherwise, the reported gains rely
  3. [Section IV-A] The bound in Lemma 1 assumes that for any model w, the expected gradient on a given class c is identical across clients and across frames. The proof explicitly states: 'For any given model w, the expected gradient on class c is assumed to be identical.' This excludes within-class domain shift, i.e., changes in the class-conditional feature distribution over time. Such shifts are part of the temporal drift the paper claims to address. Thus the bound in Eq. (9) does not quantify the full temporal drift; it only captures shifts in class proportions. The paper's own wording in Lemma 1—'a realistic scenario where the data features of each class do not change over time'—contradicts the dynamic-data motivation of the paper.
  4. [Section IV-D3] The derivation of Eq. (18) assumes a client contains only a single class ('Assume that client n contains only class c of data'). The paper then states that when clients contain multiple classes, 'deriving L^(c) becomes difficult, so we adopt the same estimation approach.' This is a heuristic extension without justification. The experiments include clients with two classes (Section V-A, CIFAR-10 setting: '10 clients each has one class of data and 10 clients each has two classes'), so Eq. (18) is used in a regime where its derivation does not hold. Additionally, the denominator in Eq. (18) is zero whenever p_{n,l,k}^{(c)} = p_{l,k}^{(c)}, an unhandled case.
minor comments (5)
  1. [Section IV-A] The temporal drift bound xi_{n,l} defined in Definition 1 is not used anywhere in the formulation or algorithm; only the EMD-based expression in Lemma 1 appears in the scheduling objective. The role of xi_{n,l} should be clarified or removed.
  2. [Section IV-D3 / Algorithm 1] The timing of the L^(c) estimate is unclear: line 29 estimates L^(c) after aggregation, but the scheduling at round k appears to need L^(c) for the same round. If a stale estimate from the previous round is used, this should be stated and justified in terms of the convergence bound.
  3. [Section V-A] The choice of T_max = 30 s for CIFAR-100 is not motivated; for CIFAR-10, the text considers 0.8, 1.0, and 1.2 s. A brief justification would help.
  4. [Section V-B] The value of lambda_0 (2.0) and the linear decay schedule are introduced without sensitivity analysis or a selection criterion. Since these parameters directly control the plasticity-stability trade-off, some robustness discussion is needed.
  5. [General] There are minor typographical issues, e.g., 'federared' in Theorem 1 and inconsistent notation such as w_{l,K_l} vs. w_{l,K_l}. Figure 4's bar labels ('Total selected clients' / 'Clients with New Classes') are visually ambiguous and should be explained more clearly.

Circularity Check

2 steps flagged

Theoretical motivation rests on a self-cited omitted proof, and the L(c) estimator requires the very global gradient the paper says is unavailable, making the scheduling objective partially self-referential/undefined.

specific steps
  1. self citation load bearing [Section IV-B, Theorem 1 and Eq. (12)]
    "Following our previous work [29], the impact of δ l,k is derived in the following theorem. ... Theorem 1 can be obtained following similar derivations as our previous work [29]. We omit the proof due to space limit."

    The P2 objective's first two terms (sampling variance and collective divergence) are taken directly from Theorem 1's convergence bound, which is the only theoretical justification for the proposed scheduling rule. The theorem is not proved in this paper; it is imported from the authors' own prior FedCGD work [29]. Thus the claimed 'impact' of collective divergence on convergence is not independently established here, and the derivation of the central optimization objective reduces to a self-citation with an omitted proof.

  2. other [Section IV-D3, Eq. (18) and Algorithm 1 lines 9-10]
    "Since we cannot collect all data in a single round to compute the global loss and gradient, the precise values of gradient variance σ and gradient norm L(c) of class c cannot be obtained. ... Meanwhile, L(c) in round k of frame l can be estimated as ... ∇F_l(w,D^(c)_l) ..."

    The scheduling objective P2/P3 depends on L(c), the per-class gradient norm. The proposed estimator (18) uses ∇F_l(w,D^(c)_l), the global class-c gradient, in its numerator. But the paper explicitly states that the global loss/gradient cannot be computed in a single round, and Algorithm 1 only sends class proportions and gradient variances from clients to the server before scheduling (lines 9-10). Local models are sent only by scheduled clients after selection. Therefore the server cannot compute Eq. (18) from the information the algorithm actually collects, making the objective's key input undefined in a true federated setting. The reported results must rely on an unstated server-side oracle, which undermines the claimed federated/privacy setup and makes the objective self-referential to

full rationale

The paper's empirical comparisons are independent content and demonstrate strong performance, so this is not a case of pure fabrication. However, the theoretical derivation chain is partially circular. First, the convergence bound that motivates the entire scheduling objective is borrowed from the authors' own previous work [29] with the proof omitted; this is a load-bearing self-citation rather than an independently verified mathematical result. Second, the L(c) estimator (18) requires the global class-wise gradient, which the paper admits cannot be collected, and which Algorithm 1 never transmits to the server; thus the scheduling objective P2/P3 is undefined from the information the algorithm actually uses. These two issues mean the 'first-principles' justification of FedTeddi reduces to a self-citation plus an estimator that presupposes the global data the federated setting excludes. The temporal-drift heuristic and the experimental results retain independent empirical content, so the overall circularity is partial rather than total.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The approach rests on assumptions borrowed from prior work, a hand-picked drift weight, and an approximate estimator for class gradient norms. The most fragile items are the omitted proof of Theorem 1, the invariance of per-class expected gradients, and the unjustified single-class estimator applied to multi-class clients.

free parameters (2)
  • lambda_0 (initial temporal-drift weight) = 2.0
    Chosen by hand to balance drift and divergence in Eq. (13); no sensitivity analysis is reported, and it is tuned on the exact datasets used for evaluation.
  • lambda_k decay schedule = lambda_0 (1 - k/K_l)
    Ad hoc linear decay to trade plasticity versus stability; introduced in Eq. (13) without derivation or ablation.
axioms (4)
  • domain assumption The convergence bound of Theorem 1 from [29] applies to this setup
    Theorem 1 is stated in Section IV-B with 'We omit the proof due to space limit' and is deferred to the authors' own FedCGD preprint [29].
  • ad hoc to paper Per-class expected gradient norms L^(c) are identical across frames and clients
    Assumed in Lemma 1 ('for any given model w, the expected gradient on class c is assumed to be identical') to express temporal drift as EMD of class proportions; excludes within-class feature drift.
  • standard math Loss functions are convex, rho-Lipschitz, and beta-smooth with bounded gradient and sampling variances
    Assumed in Theorem 1 for the convergence bound, but the experiments use non-convex CNN and ResNet-18 models with Group Normalization.
  • domain assumption Computation delay follows a shifted exponential distribution (Eq. 4)
    Standard model for straggler behavior, cited from [34]-[36]; used to derive the bandwidth allocation in Eq. (14).
invented entities (1)
  • Temporal drift bound xi_{n,l} no independent evidence
    purpose: Quantifies data evolution per client to prioritize clients holding new classes
    Defined via gradient differences and EMD of class proportions in Eqs. (8)-(9); it is a construct of this paper with no external falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 15367 in / 10217 out tokens · 109411 ms · 2026-08-04T22:20:44.374053+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FedTeddi: Temporal Drift and Divergence Aware Scheduling for Timely Federated Edge Learning." pith.science (2026). https://pith.science/paper/WMH4KXFL

@misc{pith2026250907342,
  author       = {Pith},
  title        = {Pith review of: FedTeddi: Temporal Drift and Divergence Aware Scheduling for Timely Federated Edge Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WMH4KXFL}},
  note         = {Machine review of arXiv:2509.07342}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Federated edge learning (FEEL) enables collaborative model training across distributed clients over wireless networks without exposing raw data. While most existing studies assume static datasets, in real-world scenarios clients may continuously collect data with time-varying and non-independent and identically distributed (non-i.i.d.) characteristics. A critical challenge is how to adapt models in a timely yet efficient manner to such evolving data. In this paper, we propose FedTeddi, a temporal-drift-and-divergence-aware scheduling algorithm that facilitates fast convergence of FEEL under dynamic data evolution and communication resource limits. We first quantify the temporal dynamics and non-i.i.d. characteristics of data using temporal drift and collective divergence, respectively, and represent them as the Earth Mover's Distance (EMD) of class distributions for classification tasks. We then propose a novel optimization objective and develop a joint scheduling and bandwidth allocation algorithm, enabling the FEEL system to learn from new data quickly without forgetting previous knowledge. Experimental results show that our algorithm achieves higher test accuracy and faster convergence compared to benchmark methods, improving the rate of convergence by 58.4% on CIFAR-10 and 49.2% on CIFAR-100 compared to random scheduling.

Figures

Figures reproduced from arXiv: 2509.07342 by Sheng Zhou, Tan Chen, Wei Chen, Yuxuan Bai, Yuxuan Sun, Zhisheng Niu.

Figure 1
Figure 1. Figure 1: Schematic illustration of the FEEL system with heterogeneous data arrival across frames. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The workflow of the FedTeddi within a single frame [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Test accuracy of baselines with different delay constraints. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Client selection results with different delay constraints. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rounds to reach different accuracy levels with different delay constraints. The results are averaged over 5 independent runs, and the shadow areas [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Test accuracy of baselines on CIFAR-100. [PITH_FULL_IMAGE:figures/full_fig_p009_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

39 extracted references · 36 canonical work pages · 2 internal anchors

  1. [1]

    Federated learning in mobile edge networks: A comprehensive survey,

    W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y . Liang, Q. Yang, D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A comprehensive survey,”IEEE Commun. Surveys Tuts., vol. 22, no. 3, pp. 2031–2063, 3rd Quart. 2020

    work page 2031

  2. [2]

    Federated edge learning for 6G: Foundations, methodologies, and applications,

    M. Tao, Y . Zhou, Y . Shi, J. Lu, S. Cui, J. Lu, and K. B. Letaief, “Federated edge learning for 6G: Foundations, methodologies, and applications,”Proceedings of the IEEE, pp. 1–39, 2024

    work page 2024

  3. [3]

    A comprehensive survey of continual learning: Theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 8, pp. 5362–5383, Aug. 2024

    work page 2024

  4. [4]

    Federated continual learning for edge-AI: A compre- hensive survey,

    Z. Wang, et al., “Federated continual learning for edge-AI: A compre- hensive survey,” inarXiv preprint arXiv:2411.13740, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Federated continual learning: Concepts, challenges, and solutions,

    P. Hamedi, R. Razavi-Far, and E. Hallaji, “Federated continual learning: Concepts, challenges, and solutions,”arXiv preprintarXiv:2502.07059, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Federated learning: Challenges, methods, and future directions,

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,”IEEE Signal Process. Mag., vol. 37, no. 3, pp. 50–60, May 2020

    2020

  7. [7]

    A comprehensive survey of forgetting in deep learning beyond continual learning,

    Z. Wang, E. Yang, L. Shen, and H. Huang, “A comprehensive survey of forgetting in deep learning beyond continual learning,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 3, pp. 1464–1483, 2025

    work page 2025

  8. [8]

    On the convergence of FedAvg on non-i.i.d. data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of FedAvg on non-i.i.d. data,” inProc. Int. Conf. Learn. Represent. (ICLR), Addis Ababa, Ethiopia, Apr. 2020

    work page 2020

  9. [9]

    Rethinking data heterogeneity in federated learning: Introducing a new notion and standard benchmarks,

    S. Vahidian, M. Morafah, C. Chen, M. Shah, and B. Lin, “Rethinking data heterogeneity in federated learning: Introducing a new notion and standard benchmarks,”IEEE Trans. Artif. Intell., vol. 5, no. 3, pp. 1386–1397, 2024

    work page 2024

  10. [10]

    Recent advances on federated learning: A systematic survey,

    B. Liu, N. Lv, Y . Guo, and Y . Li, “Recent advances on federated learning: A systematic survey,”Neurocomputing, vol. 597, p. 128019, 2024

    work page 2024

  11. [11]

    Non-i.i.d. data and continual learning processes in federated learning: A long road ahead,

    M. F. Criado, F. E. Casado, R. Iglesias, C. V . Regueiro, and S. Barro, “Non-i.i.d. data and continual learning processes in federated learning: A long road ahead,”Inf. Fusion, vol. 88, pp. 263–280, 2022

    work page 2022

  12. [12]

    Advances in robust federated learning: A survey with heterogeneity considerations,

    C. Chen, T. Liao, X. Deng, Z. Wu, S. Huang, and Z. Zheng, “Advances in robust federated learning: A survey with heterogeneity considerations,” IEEE Trans. Big Data, pp. 1–20, 2025

    work page 2025

  13. [13]

    Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity,

    A. Reisizadeh, I. Tziotis, H. Hassani, A. Mokhtari, and R. Pedarsani, “Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity,”IEEE J. Sel. Areas Inf. Theory, vol. 3, no. 2, pp. 197–205, Jun. 2022

    work page 2022

  14. [14]

    Toward scalable wireless federated learning: Challenges and solutions,

    Y . Zhou, Y . Shi, H. Zhou, J. Wang, L. Fu, and Y . Yang, “Toward scalable wireless federated learning: Challenges and solutions,”IEEE Internet Things Mag., vol. 6, no. 4, pp. 10–16, Dec. 2023

    work page 2023

  15. [15]

    Communication- efficient edge AI: Algorithms and systems,

    Y . Shi, K. Yang, T. Jiang, J. Zhang, and K. B. Letaief, “Communication- efficient edge AI: Algorithms and systems,”IEEE Commun. Surveys Tuts., vol. 22, no. 4, pp. 2167–2191, 4th Quart., 2020

    work page 2020

  16. [16]

    Scheduling for cellular federated edge learning with importance and channel aware- ness,

    J. Ren, Y . He, D. Wen, G. Yu, K. Huang, and D. Guo, “Scheduling for cellular federated edge learning with importance and channel aware- ness,”IEEE Trans. Wireless Commun., vol. 20, no. 11, pp. 7690–7703, Nov. 2021

    work page 2021

  17. [17]

    Generative replay with meta-learning for efficient continual adaptation,

    Z. Zhang, Y . Li, and J. Liu, “Generative replay with meta-learning for efficient continual adaptation,”Neurocomputing, vol. 601, pp. 150–162, 2024

    work page 2024

  18. [18]

    Towards Efficient Replay in Federated Incremental Learning,

    Y . Li, Q. Li, H. Wang, R. Li, W. Zhong, and G. Zhang, “Towards Efficient Replay in Federated Incremental Learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle, W A, USA, Jun. 2024, pp. 12820–12829

    work page 2024

  19. [19]

    Adaptive importance weighting for continual learning,

    C. Sun, L. Wang, and T. Zhang, “Adaptive importance weighting for continual learning,”Pattern Recognit., vol. 142, p. 109673, 2023

    work page 2023

  20. [20]

    FedSSI: Rehearsal-Free Continual Fed- erated Learning with Synergistic Synaptic Intelligence,

    H. Kim, J. Park, and S. Lee, “FedSSI: Rehearsal-Free Continual Fed- erated Learning with Synergistic Synaptic Intelligence,” inProc. 42nd Int. Conf. Mach. Learn. (ICML), Vancouver, Canada, PMLR, 2025

    work page 2025

  21. [21]

    Mask-based continual learning for adaptive neural networks,

    Y . Liu, F. Chen, and P. Zhao, “Mask-based continual learning for adaptive neural networks,”Inf. Sci., vol. 642, p. 119200, 2023

    work page 2023

  22. [22]

    Asynchronous Federated Learning for Sensor Data with Concept Drift

    Y . Chen, Z. Chai, Y . Cheng, and H. Rangwala, “FedConD: Asyn- chronous Federated Learning for Sensor Data with Concept Drift,”arXiv preprintarXiv:2109.00151, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  23. [23]

    Federated Learning under Distributed Concept Drift (FedDrift),

    E. Jothimurugesan, K. Hsieh, J. Wang, G. Joshi, and P. B. Gibbons, “Federated Learning under Distributed Concept Drift (FedDrift),” in Proc. Int. Conf. Artif. Intell. Statist. (AISTATS), 2023

    work page 2023

  24. [24]

    FedINC: Federated Incremental Learning under Concept Drift,

    A. Singh, R. Patel, and M. Kumar, “FedINC: Federated Incremental Learning under Concept Drift,” inProc. IEEE Int. Conf. Big Data, 2023, pp. 4567–4576

    work page 2023

  25. [25]

    Flash: Fast learning for adaptive concept drift in federated environments,

    S. Panchal, P. Chawla, and V . Gupta, “Flash: Fast learning for adaptive concept drift in federated environments,” inProc. Int. Conf. Learn. Representations (ICLR), 2023

    work page 2023

  26. [26]

    Fed-CBS: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction,

    L. Zhang, B. He, S. Chen, X.-H. Sun, and Y . Wang, “Fed-CBS: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction,” inProc. Int. Conf. Mach. Learn. (ICML), Honolulu, HI, USA, Jul. 2023, pp. 41622–41633

    work page 2023

  27. [27]

    Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,

    T. Chen, J. Yan, Y . Sun, S. Zhou, D. Gunduz, and Z. Niu, “Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,”IEEE Trans. V eh. Technol., vol. 74, no. 1, pp. 1657–1673, Jan. 2025

    work page 2025

  28. [28]

    Dynamic scheduling for over-the-air federated edge learning with energy constraints,

    Y . Sun, S. Zhou, Z. Niu, and D. Gunduz, “Dynamic scheduling for over-the-air federated edge learning with energy constraints,”IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 227–242, Jan. 2022

    work page 2022

  29. [29]

    FedCGD: Collective Gradient Divergence Optimized Scheduling for Wireless Federated Learning

    T. Chen, J. Yan, Y . Sun, S. Zhou, and Z. Niu, “FedCGD: Collective gra- dient divergence optimized scheduling for wireless federated learning,” arXiv preprintarXiv:2506.07581, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Fairness-Aware Client Selection in Federated Learning with Heterogeneous Data and Resources,

    X. Zeng, H. Xu, and M. Chen, “Fairness-Aware Client Selection in Federated Learning with Heterogeneous Data and Resources,”IEEE Trans. Mobile Comput., vol. 22, no. 12, pp. 7312–7326, Dec. 2023

    work page 2023

  31. [31]

    Reinforcement learning-based client selection for resource-constrained federated learning,

    J. Wang, Y . Liu, and K. Xu, “Reinforcement learning-based client selection for resource-constrained federated learning,”Neurocomputing, vol. 563, pp. 126–139, 2025

    work page 2025

  32. [32]

    Adaptive Memory Replay for Continual Learning,

    J. S. Smith, L. Valkov, S. Halbe, V . Gutta, R. Feris, Z. Kira, and L. Karlinsky, “Adaptive Memory Replay for Continual Learning,” in Proc. 2024 IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), Los Alamitos, CA, USA, Jun. 2024, pp. 3605–3615

    work page 2024

  33. [33]

    Effec- tive Generative Replay with Strong Memory for Continual Learning,

    J. Yang, X. Zhou, Y . He, Q. Li, Z. Su, X. Ruan, and C. Zhang, “Effec- tive Generative Replay with Strong Memory for Continual Learning,” Knowledge-Based Systems, vol. 319, p. 113477, 2025

    work page 2025

  34. [34]

    Joint device scheduling and resource allocation for latency constrained wireless federated learning,

    W. Shi, S. Zhou, Z. Niu, M. Jiang, L. Geng, “Joint device scheduling and resource allocation for latency constrained wireless federated learning,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 453-467, Sep. 2020

    work page 2020

  35. [35]

    A unified coding framework for distributed computing with straggling servers,

    S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “A unified coding framework for distributed computing with straggling servers,” inProc. IEEE Globecom Workshops (GC Wkshps), Dec. 2016, pp. 1–6

    work page 2016

  36. [36]

    Speeding up distributed machine learning using codes,

    K. Lee, M. Lam, R. Pedarsani, et al., “Speeding up distributed machine learning using codes,” IEEE Trans. Inf. Theory, vol. 64, no. 3, pp. 1514– 1529, Mar. 2017

    work page 2017

  37. [37]

    Convergence of update aware client scheduling for federated learning at the wireless edge,

    M. M. Amiri, D. G ¨und¨uz, S. R. Kulkarni, and H. V . Poor, “Convergence of update aware client scheduling for federated learning at the wireless edge,”IEEE Trans. Wireless Commun., vol. 20, pp. 3643–3658, no. 6, Jun. 2021

    work page 2021

  38. [38]

    Towards understanding biased client selection in federated learning,

    Y . J. Cho, J. Wang, and G. Joshi, “Towards understanding biased client selection in federated learning,” in Proc. Artificial Intelligence and Statistics (AIStats), May 2022, pp. 10351–10375

    work page 2022

  39. [39]

    Learning Multiple Layers of Features from Tiny Im- ages,

    A. Krizhevsky, “Learning Multiple Layers of Features from Tiny Im- ages,” Technical Report, University of Toronto, 2009

    work page 2009