Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

Carbon-aware scheduling in federated learning can cut emissions substantially without sacrificing accuracy, provided the planner's selection probabilities match reality.

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 ↗

A carbon-aware federated learning scheduler that adds slack time, fair client selection, and fine-tuning beats a full-participation baseline on MNIST under tight carbon budgets.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid green-FL scheduling formulation with a real evaluation-protocol flaw: the headline accuracy gains come from test-set-chosen fine-tuning placement, not the scheduler as specified. the 4 major comments →

arxiv 2509.08980 v1 pith:NJNHAD5Q submitted 2025-09-10 cs.LG

Green Federated Learning via Carbon-Aware Client and Time Slot Scheduling

Daniel Richards Arputharaj , Charlotte Rodriguez , Angelo Rodio , Giovanni Neglia This is my paper
classification cs.LG
keywords carbon-aware schedulingfederated learningslack timecarbon intensityclient selectionfairnessunbiased aggregationfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Federated learning spreads training over many devices, and those devices run on electricity grids with very different carbon intensity at different hours. This paper tries to show that a carbon-aware scheduler can exploit that variability — by adding slack time so training gets deferred to low-carbon periods, by allocating carbon fairly across clients, and by correcting selection bias — and that it can reduce emissions substantially without meaningful accuracy loss. The paper supports this with simulations on real-world carbon-intensity traces, reporting accuracy gains over a slack-agnostic baseline that are largest exactly when the carbon budget is tight. A sympathetic reader would take away that carbon efficiency in FL is not a sacrifice but a scheduling problem with identifiable biases to manage.

Core claim

The paper's central claim is that federated learning's carbon footprint can be cut by jointly choosing which clients train and when, rather than treating all rounds and regions as interchangeable. It shows that simply allowing training to extend past its minimum duration (slack time) lets clients pick their lowest-carbon hours, yielding 20–60% CO2e savings per client on variable grids and aggregate savings that stay above 20% even when all clients participate. It then identifies three learning biases that a carbon-greedy schedule introduces — statistical heterogeneity from excluding high-emission clients, selection bias from unequal sampling frequencies, and last-iterate bias from temporally

What carries the argument

The load-bearing object is the scheduling optimization problem (Eq. 8): a carbon-budgeted integer program that maximizes an alpha-fair utility over client-time-slot assignments, with a movable fine-tuning window and a slack window. The alpha-fair utility interpolates between carbon-greedy (alpha=1) and equal-carbon allocation (alpha→0), maintaining fair representation. Around it, two corrections do the work: U-FedAvg, which replaces simple averaging with inverse-frequency weighting so the global update is unbiased under unequal selection, and the fine-tuning window, in which all clients are selected to counteract correlated participation and last-iterate bias. The problem is NP-hard but solv

Load-bearing premise

The load-bearing premise is that the planner's selection probabilities match reality: the unbiased aggregation (U-FedAvg) is only unbiased if each client actually participates at each scheduled slot with the planned probability π_c, and if real devices are unavailable or the schedule is not followed exactly, the accuracy gains claimed here do not follow.

What would settle it

Run the scheduler's exact schedule in an environment where each scheduled client participates with probability 0.9 instead of 1, and compare final accuracy to the slack-agnostic baseline under the same carbon budget; if the accuracy advantage disappears or reverses, the unbiasedness assumption is the fragile link. More directly, compute realized participation frequencies and compare them with the planned π_c — any divergence larger than the scale of 1/T breaks the unbiasedness of the aggregation rule.

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

If this is right

  • With slack time alone, federated training can cut emissions by 20–60% per client on grids with variable carbon intensity, before any aggregation changes.
  • A carbon-greedy scheduler that ignores fairness excludes high-emission clients and hurts accuracy; using an alpha-fair allocation improves accuracy by up to 8 percentage points at medium/large budgets.
  • Unbiased FedAvg removes the non-vanishing selection-bias error, at the cost of slower convergence when selection is highly heterogeneous.
  • A short all-client fine-tuning phase at the end recovers up to 3.8 percentage points of accuracy under temporally and spatially correlated carbon intensity.
  • The advantage over a slack-agnostic baseline is highest under tight carbon budgets (up to +4.36 pp) and shrinks to about +0.2 pp when budgets are large.

Where Pith is reading between the lines

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

  • Editorial extension: the same scheduling formulation applies to any per-client, time-varying cost, not just carbon intensity; replacing CI with electricity price or device energy would turn the method into a general cost-aware scheduler.
  • Editorial extension: the scheduler is computed offline from carbon-intensity forecasts; a live version that re-solves the problem as forecasts update could capture additional savings when forecasts are wrong, because slack time gives room to shift.
  • Editorial extension: the planned-vs-realized assignment assumption in U-FedAvg suggests a practical deployment should track realized participation frequencies and either reweight online or fall back to standard FedAvg when schedules are disrupted.
  • Editorial extension: the reported gains use a 7-client MNIST experiment; transferring to large-scale cross-device FL would require accounting for communication energy and client dropout, but the scheduling math is independent of these.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper addresses carbon-efficient federated learning by scheduling client participation and training time slots according to regional and temporal carbon intensity (CI). It first quantifies potential CO2e savings from slack time, then studies statistical and temporal biases induced by carbon-aware client selection, and proposes a scheduler that combines slack time, α-fair carbon allocation, unbiased aggregation (U-FedAvg), and a fine-tuning phase. The method is formulated as a non-convex integer program (Problem (8)) and evaluated with real-world Electricity Maps CI traces on federated MNIST. The main claimed result is that the proposed scheduler outperforms a slack-agnostic all-clients baseline in model accuracy across carbon budgets, with especially strong gains at tight budgets.

Significance. If the accuracy gains were obtained by the algorithm as specified, the paper would make a useful contribution to green FL: it integrates carbon awareness with statistical fairness and temporal bias correction, and the carbon-savings analysis in Figures 2 and 3 is a straightforward, reproducible computation on public CI traces. The paper also ships code, which strengthens reproducibility. However, the central accuracy claim is currently supported by an evaluation protocol that selects fine-tuning placement and hyperparameters after observing test accuracy, so the reported gains are likely an oracle upper bound rather than a property of the proposed scheduler. The underlying carbon-savings derivations and the unbiasedness mechanism are plausible, but the empirical comparison needs to be rerun under a non-oracle protocol before the main claim can be accepted.

major comments (4)
  1. [Section VII, Fig. 7/Table II] The headline accuracy gains (+4.36 pp at 5.73% budget, +3.24 pp at 7.65%) are obtained by choosing, after seeing test accuracy, the best fine-tuning end time s and duration t_ft for each budget (Table II), while α=0.1 is fixed from Figure 5 and learning rates are grid-searched per experiment. Algorithm 1 / Problem (8) chooses s from the carbon utility objective, not from test accuracy. The reported numbers are therefore an oracle upper bound over test-set-selected hyperparameters, not a property of the proposed scheduler. Please evaluate the algorithm as specified (solve (8) for s, or define an a priori selection rule) and report accuracy without test-set peeking.
  2. [Section VI, Problem (8)] The objective (8a) depends on A only through the carbon-utility sum; s enters only via constraints (8c)-(8f). Hence any feasible s is equally optimal for (8), and the statement that the scheduler 'optimizes' fine-tuning placement is not supported by the optimization problem. Consequently Algorithm 1's line 2 is under-specified: CarbonScheduler must return an s, but no criterion selects it. A secondary objective or an accuracy-aware tie-breaking rule is needed before the simulation can be claimed to evaluate the proposed method.
  3. [Section V-B, Eq. (7)] Unbiasedness of U-FedAvg requires that the realized participation probabilities equal the planned π_c. In Section V-B the argument is explicitly conditional on independent selection with probability π_c. The experiments execute the planned schedule deterministically, so Eq. (7) is internally consistent in simulation, but the real cross-device claim depends on availability being orthogonal to the schedule. No experiment or bound addresses deviation from planned π_c. Either add a robustness experiment with stochastic availability and resampling, or clearly scope the unbiasedness claim to exact schedule execution.
  4. [Section V/Experimental Setup and Section VII] The hyperparameter protocol is not valid for comparing algorithms: learning rates are 'tuned for each experiment via grid search' (Section V), α=0.1 is selected from Figure 5 on test accuracy, and baselines are not given the same degree of fine-tuning/slack optimization. This biases comparisons in favor of the proposed method. Please separate model selection (on validation data) from evaluation (on test data), and use the same selection protocol for all baselines.
minor comments (3)
  1. [Section II] The sentence 'To our knowledge, we are the first to address temporal bias in carbon-aware scheduling...' appears twice, once at the end of the related-work section and once earlier in the same paragraph. Please remove the duplicate.
  2. [Section V-B, Eq. (7)] The reweighting in Eq. (7) divides by π_c. If a client is never selected (π_c=0), the update is undefined. The paper should state how zero selection frequencies are handled, especially under very tight carbon budgets where the feasible set may exclude some clients despite the α-fair objective.
  3. [Table II] Table II is dense and the reader cannot easily see which (s, t_ft) configuration is selected for each budget. Consider marking the chosen configuration in each row, or moving the full sweep to an appendix and reporting only the selected configuration in the main text.

Circularity Check

1 steps flagged

Reported accuracy gains are oracle best-of-grid selections over test-tuned fine-tuning placement, not the output of the proposed scheduler.

specific steps
  1. fitted input called prediction [Section VII, Figure 7 caption / Table II; Section VI, Algorithm 1 line 2 and Problem (8)]
    "Each point corresponds to the best end time for fine-tuning configuration (from Table II) for a given budget level and is annotated with its associated budget. We fix the fairness parameter to α=0.1, identified in Figure 5 as providing an effective trade-off between test accuracy and carbon efficiency."

    Algorithm 1 line 2 delegates the choice of s to CarbonScheduler solving Problem (8), whose objective (8a) is the α-fair carbon utility and contains no model-accuracy term. The reported benchmark points are not the output of that optimization: each point is the maximum test accuracy over the grid s∈{1,3,...,150}, t_ft∈{1,3} in Table II, and α is selected from the test-accuracy curve in Figure 5. Hence the +4.36 pp and +3.24 pp gains are best-of-grid values chosen after observing the test set, not predictions generated by the proposed scheduler. The performance claim is therefore forced by construction: it compares the baseline to an oracle-tuned version of the method rather than to Algorithm 1 as specified.

full rationale

The carbon-savings quantification in Section IV is self-contained and not circular: Eqs. (4)-(5) compare the T lowest-CI slots to the first T slots using Electricity Maps data. Problem (6)/(8) defines the scheduler independently of test accuracy, and Eq. (7) is a standard importance-weighting correction; the unbiasedness and convergence bounds are imported from [31], a co-author's published theorem, but as external mathematical support rather than a result fitted to the present data, so I do not count that as circularity under the hard rules. The central empirical claim, however, is not evaluated under Algorithm 1 as specified: Figure 7 uses the best s per budget chosen from Table II after observing test accuracy, and α is fixed from Figure 5's test-accuracy curves. Since the scheduler's objective (8a) contains no accuracy term, the reported +4.36 and +3.24 pp gains are maxima over a test-set-tuned grid, i.e. an oracle upper bound. This is a fitted input (s, t_ft, α) presented as the method's predicted performance, reducing the headline comparison to construction. Score 6 reflects this partial circularity; the underlying scheduling derivation itself is not definitionally circular.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, but its reported accuracy depends on four tuned decision parameters and on strong assumptions about constant per-round energy and exact execution of the planned schedule.

free parameters (4)
  • alpha (fairness exponent) = 0.1
    Fixed in Section VII from the accuracy sweep in Figure 5; test accuracy is used to pick it.
  • learning rate eta = not reported (grid 0.1, 0.0316, 0.01)
    Tuned per experiment by grid search (Section V), so results depend on test-set-informed hyperparameter choice.
  • fine-tuning end time s = varies 1..150, best per budget
    Section VII selects the best s from Table II for each carbon budget after observing test accuracy; effectively a free parameter.
  • fine-tuning duration t_ft = 1 or 3, best per budget
    Chosen post hoc per budget from Table II; longer fine-tuning consumes carbon budget and changes accuracy.
axioms (5)
  • domain assumption Each client consumes fixed energy E_c per one-hour round, so carbon cost g_c^t = E_c * CI_c^t.
    Section III-B; ignores variable round durations, communication costs, and device-level power dynamics.
  • domain assumption Selection frequencies pi_c are known and equal to realized participation probabilities for U-FedAvg unbiasedness.
    Section V-B, Eq. (7); unbiasedness depends on this, but the scheduler is computed offline and cannot control client availability in cross-device settings.
  • standard math Global objective is the unweighted average of client losses F(theta) = (1/K) sum_c F_c(theta).
    Equation (1); standard FL assumption.
  • domain assumption Smooth strongly convex objectives and diminishing learning rates for the cited U-FedAvg convergence bound O(rho_H/T).
    Section V-B cites [31]; the experiments use a non-convex CNN, so the theoretical bound does not directly apply to the reported accuracy.
  • standard math Objective (8) is monotone submodular under a single-knapsack constraint, enabling the greedy (1-1/e) approximation.
    Section VI, citing [34]; used only for large-scale instances, not the experiments, which use exact MOSEK solving.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Green Federated Learning via Carbon-Aware Client and Time Slot Scheduling." pith.science (2026). https://pith.science/paper/NJNHAD5Q

@misc{pith2026250908980,
  author       = {Pith},
  title        = {Pith review of: Green Federated Learning via Carbon-Aware Client and Time Slot Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJNHAD5Q}},
  note         = {Machine review of arXiv:2509.08980}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Training large-scale machine learning models incurs substantial carbon emissions. Federated Learning (FL), by distributing computation across geographically dispersed clients, offers a natural framework to leverage regional and temporal variations in Carbon Intensity (CI). This paper investigates how to reduce emissions in FL through carbon-aware client selection and training scheduling. We first quantify the emission savings of a carbon-aware scheduling policy that leverages slack time -- permitting a modest extension of the training duration so that clients can defer local training rounds to lower-carbon periods. We then examine the performance trade-offs of such scheduling which stem from statistical heterogeneity among clients, selection bias in participation, and temporal correlation in model updates. To leverage these trade-offs, we construct a carbon-aware scheduler that integrates slack time, $\alpha$-fair carbon allocation, and a global fine-tuning phase. Experiments on real-world CI data show that our scheduler outperforms slack-agnostic baselines, achieving higher model accuracy across a wide range of carbon budgets, with especially strong gains under tight carbon constraints.

Figures

Figures reproduced from arXiv: 2509.08980 by Angelo Rodio, Charlotte Rodriguez, Daniel Richards Arputharaj, Giovanni Neglia.

Figure 3
Figure 3. Figure 3: CO2e savings from slack-aware versus slack-agnostic client selection as a function of the number of selected clients (with lowest average carbon emissions) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: CO2e savings per individual client as a function of slack time. Each row corresponds to a different region. and up to 60% reductions with tsl = 236 hours. Moreover, with tsl = 236 hours, approximately 80% of clients reduce their training carbon intensity by at least 10%, and 50% of them achieve reductions by at least 20%. B. Impact of Slack Time and Client Selection We extend the analysis to jointly optimi… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of fairness parameter α on allocation (green: selected, red: excluded): from a carbon-greedy (Fig. 4a) to a carbon-fair (Fig. 4b) allocation. with learning rates tuned for each experiment via grid search over η ∈ {10−1 , 10−1.5 , 10−2}. All results are averaged over three independent runs. A. Fair Carbon Allocation under Statistical Heterogeneity In real-world FL systems, time slots are not intercha… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of heterogeneity and correlation on test accuracy, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Final test accuracy for our carbon-aware scheduler (y [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Standardized Methods and Recommendations for Green Federated Learning

    cs.DC 2026-01 conditional novelty 4.0

    A phase-aware carbon-accounting method for federated learning shows that client efficiency tiers and coordination idle time can inflate total CO2e by 8-22x, while GPU choice changes runtime more than energy.

Reference graph

Works this paper leans on

34 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Energy and Policy Consid- erations for Modern Deep Learning Research,

    E. Strubell, A. Ganesh, and A. McCallum, “Energy and Policy Consid- erations for Modern Deep Learning Research,”Proceedings of the AAAI Conference on Artificial Intelligence, 2020

    2020

  2. [2]

    Estimating Power Consumption and Carbon Impact of Deep- Learning on Nef,

    M. Vesin, “Estimating Power Consumption and Carbon Impact of Deep- Learning on Nef,” Inria SED SAM Blog, 2020

    2020

  3. [3]

    eco2AI: Carbon Emissions Tracking of Machine Learning Models as the First Step Towards Sustainable AI,

    S. A. Budennyy, V . D. Lazarevet al., “eco2AI: Carbon Emissions Tracking of Machine Learning Models as the First Step Towards Sustainable AI,”Doklady Mathematics, 2022

    2022

  4. [4]

    Understanding Higher Order Impacts of Green ICT,

    A. Hankel, “Understanding Higher Order Impacts of Green ICT,” inICT for Sustainability 2014. Atlantis Press, 2014

    2014

  5. [5]

    Green AI,

    R. Schwartz, J. Dodgeet al., “Green AI,”Communications of the ACM, 2020

    2020

  6. [6]

    Federated Learning: Strategies for Improving Communication Efficiency,

    J. Kone ˇcn´y, H. B. McMahanet al., “Federated Learning: Strategies for Improving Communication Efficiency,” inNIPS Workshop on Private Multi-Party Machine Learning, 2016

    2016

  7. [7]

    Communication-Efficient Learning of Deep Networks from Decentralized Data,

    B. McMahan, E. Mooreet al., “Communication-Efficient Learning of Deep Networks from Decentralized Data,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics. PMLR, 2017

    2017

  8. [8]

    Advances and Open Problems in Federated Learning,

    P. Kairouz, H. B. McMahanet al., “Advances and Open Problems in Federated Learning,”Foundations and Trends® in Machine Learning, 2021

    2021

  9. [9]

    Electricity Maps Datasets,

    “Electricity Maps Datasets,” https://portal.electricitymaps.com/datasets

  10. [10]

    Multiply-and-Fire: An Event-Driven Sparse Neural Network Accelerator,

    M. Yu, T. Xianget al., “Multiply-and-Fire: An Event-Driven Sparse Neural Network Accelerator,”ACM Transactions on Architecture and Code Optimization, 2023

    2023

  11. [11]

    StarSPA: Stride-Aware Sparsity Compres- sion for Efficient CNN Acceleration,

    N.-S. Pham, S. Shinet al., “StarSPA: Stride-Aware Sparsity Compres- sion for Efficient CNN Acceleration,”IEEE Access, 2024

    2024

  12. [12]

    Carbon Explorer: A Holistic Framework for Designing Carbon Aware Datacenters,

    B. Acun, B. Leeet al., “Carbon Explorer: A Holistic Framework for Designing Carbon Aware Datacenters,” inProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, 2023

    2023

  13. [13]

    Enabling Sustainable Clouds: The Case for Virtualizing the Energy System,

    N. Bashir, T. Guoet al., “Enabling Sustainable Clouds: The Case for Virtualizing the Energy System,” inProceedings of the ACM Symposium on Cloud Computing, 2021

    2021

  14. [14]

    SqueezeNext: Hardware-Aware Neural Network Design,

    A. Gholami, K. Kwonet al., “SqueezeNext: Hardware-Aware Neural Network Design,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018

    2018

  15. [15]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter,

    V . Sanh, L. Debutet al., “DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter,” in5th Workshop on Energy Efficient Machine Learning and Cognitive Computing, 2019

    2019

  16. [16]

    A Simple and Effective Pruning Approach for Large Language Models,

    M. Sun, Z. Liuet al., “A Simple and Effective Pruning Approach for Large Language Models,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  17. [17]

    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,” in2019 IEEE International Conference on Communications, 2019

    2019

  18. [18]

    REFL: Resource-Efficient Fed- erated Learning,

    A. M. Abdelmoniem, A. N. Sahuet al., “REFL: Resource-Efficient Fed- erated Learning,” inProceedings of the 18th ACM European Conference on Computer Systems, 2023

    2023

  19. [19]

    TiFL: A Tier-based Federated Learning System,

    Z. Chai, A. Aliet al., “TiFL: A Tier-based Federated Learning System,” inProceedings of the 29th ACM International Symposium on High- Performance Parallel and Distributed Computing, 2020

    2020

  20. [20]

    Optimal Client Sampling for Federated Learning,

    W. Chen, S. Horv ´ath, and P. Richt ´arik, “Optimal Client Sampling for Federated Learning,”Transactions on Machine Learning Research, 2022

    2022

  21. [21]

    Client Selection in Federated Learn- ing: Convergence Analysis and Power-of-Choice Selection Strategies,

    Y . J. Cho, J. Wang, and G. Joshi, “Client Selection in Federated Learn- ing: Convergence Analysis and Power-of-Choice Selection Strategies,” no. arXiv:2010.01243, 2020

    Pith/arXiv arXiv 2010

  22. [22]

    Fair Resource Allocation in Federated Learn- ing,

    T. Li, M. Sanjabiet al., “Fair Resource Allocation in Federated Learn- ing,” inInternational Conference on Learning Representations, 2019

    2019

  23. [23]

    Oort: Efficient Federated Learning via Guided Par- ticipant Selection,

    F. Lai, X. Zhuet al., “Oort: Efficient Federated Learning via Guided Par- ticipant Selection,” in15th USENIX Symposium on Operating Systems Design and Implementation, 2021

    2021

  24. [24]

    A First Look into the Carbon Footprint of Federated Learning,

    X. Qiu, T. Parcolletet al., “A First Look into the Carbon Footprint of Federated Learning,”Journal of Machine Learning Research, 2023

    2023

  25. [25]

    Measuring the Carbon Intensity of AI in Cloud Instances,

    J. Dodge, T. Prewittet al., “Measuring the Carbon Intensity of AI in Cloud Instances,” inProceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 2022

    2022

  26. [26]

    FedZero: Leveraging Renewable Excess Energy in Federated Learning,

    P. Wiesner, R. Khaliliet al., “FedZero: Leveraging Renewable Excess Energy in Federated Learning,” inProceedings of the 15th ACM Inter- national Conference on Future and Sustainable Energy Systems, 2024

    2024

  27. [27]

    FedCarbon: Carbon-Efficient Federated Learn- ing with Double Flexible Controls for Green Edge AI,

    Y . Li, T. Ouyanget al., “FedCarbon: Carbon-Efficient Federated Learn- ing with Double Flexible Controls for Green Edge AI,” in2024 IEEE/ACM 32nd International Symposium on Quality of Service, 2024

    2024

  28. [28]

    CAFE: Carbon-Aware Federated Learning in Geographically Distributed Data Centers,

    J. Bian, L. Wanget al., “CAFE: Carbon-Aware Federated Learning in Geographically Distributed Data Centers,” inProceedings of the 15th ACM International Conference on Future and Sustainable Energy Systems, 2024

    2024

  29. [29]

    Green Federated Learning via Energy-Aware Client Selection,

    R. Albelaihi, L. Yuet al., “Green Federated Learning via Energy-Aware Client Selection,” inIEEE Global Communications Conference, 2022

    2022

  30. [30]

    A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer Systems,

    R. Jain, D. Chiu, and W. Hawe, “A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer Systems,” Digital Equipment Corporation, Tech. Rep. TR-301, 1984

    1984

  31. [31]

    Federated Learning Under Heterogeneous and Correlated Client Availability,

    A. Rodio, F. Faticantiet al., “Federated Learning Under Heterogeneous and Correlated Client Availability,”IEEE/ACM Transactions on Net- working, 2023

    2023

  32. [32]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanuet al., “Overcoming catastrophic forgetting in neural networks,”Proceedings of the National Academy of Sciences, 2017

    2017

  33. [33]

    Last iterate convergence of SGD for Least-Squares in the Interpolation regime

    A. V . Varre, L. Pillaud-Vivien, and N. Flammarion, “Last iterate convergence of SGD for Least-Squares in the Interpolation regime.” inAdvances in Neural Information Processing Systems. Curran Associates, Inc., 2021

    2021

  34. [34]

    Krause and D

    A. Krause and D. Golovin,Submodular Function Maximization. Cam- bridge University Press, 2014. 9

    2014

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.