acceptodds
Under review as a conference paper at ICLR 2027

Progress-Guided Client Participation in Federated Learning

Abstract

Communication is a primary bottleneck in federated learning (FL), yet standard pipelines typically fix the number of participating clients per round. A static participation budget may incur unnecessary communication when additional participation offers limited optimization benefit and, conversely, fail to capture sufficient client diversity under strong data heterogeneity. We propose Intelligent Selection of Participants (ISP), an adaptive mechanism that treats the per-round client count as a controllable optimization variable and adjusts it online to balance communication and optimization progress. ISP uses a lightweight client-count search and complements methods that select which clients participate. Across public vision and NLP benchmarks and a large-scale real-world ECG dataset, ISP reduces overall communication by up to 30% while maintaining the downstream performance of fixed-budget baselines. To assess generality, we evaluate ISP across a wide range of heterogeneity levels and system settings, demonstrating robust adaptation under noisy per-round estimates. We also provide a convergence analysis of the adaptive client-count rule underlying ISP.

Primary area:
optimization
Keywords:
Federated Learning, Adaptive Participation, Client Sampling, Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aPDgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studied client sampling for federated learning. More specifically, this paper proposed an algorithm, called Intelligent Selection of Participants (ISP), to determine the number of clients to be sampled in each round. To do this, the paper formulates the optimization problem in Eq. 2 and argues that minimizing it yields the optimal number of clients to sample.

Critical strengths
  • The formulation in Eq. 2 to determine the number of clients to be sampled looks novel.
  • Eq. 2 can not be efficiently solved. Thus, this paper proposes an algorithm to approximately solve Eq. 2, which looks novel and interesting.
Critical weaknesses
  • The problem formulation of Eq. 2 looks strange. Even in the non-distributed SGD, it is not guaranteed that f(xt+1)<f(xt)f(x_{t+1}) < f (x_t). I understand that this paper considers the deterministic setting, while it would be a severe drawback when we apply the proposed method to neural network training.
  • The convergence analysis seems insufficient. Specifically, there is no theorem that the proposed method converges to a stationary point. It is important to evaluate how many times we sample the clients in total until reaching the ϵ\epsilon-accurate point. Otherwise, it is unclear if the proposed methods superior to the standard uniform sampling. Furthermore, the bounds in Theorems 4.4 and 4.5 depend on ∥∇f(xτ)\| \nabla f (x^\tau) and $| \nabla f_i (x^\tau)| $, which makes them hard to interpret.
  • Is the expression $f(x^\tau - \gamma g^\tau(C^n)) - f(x^\tau) $ in Line 8 of Algorithm 2 a typo? If not, computing f(xτ)f(x^\tau) appears to require communication with all clients in every round, which would significantly increase the communication cost and make the method highly communication-inefficient.
  • Gradients are computed in line 4 of Algorithm 1 and again in line 5 of Algorithm 2. This seems inefficient from both communication and computational perspectives.
  • The proposed methods only determine the number of clients to be sampled, but is it possible to determine which clients to be sampled as well?
Questions for authors

See the weakness section.

4 comments
AuthorsOpenReview ↗

Expected Progress, Convergence, and Communication Efficiency (Part 1/2)

We thank the reviewer for the careful evaluation and for recognizing the novelty of the client-count formulation and its approximation. We address the five concerns below, following their order in the review. References refer to the revised manuscript; additions are highlighted in blue.

1. Expected progress and neural-network training.

Equation (2) seeks a negative expected loss change over the sampled client cohort. It does not require every realized update to decrease the loss. Thus, the possibility that an SGD step increases the loss does not contradict the formulation. The strict feasible set can also be empty; Section 3.1 explicitly discusses this case, and Algorithm 2 provides a fallback. Theorem 4.4 quantifies the estimation error, and Theorem 4.5 controls expected progress while the count is reused.

The practical search implements (2) through an estimated response, so it does not enforce monotone loss reduction at every round. We have added the paragraph “Full-population response evaluation” in Appendix G to illustrate this distinction: the estimated full-population loss change in (2) is nonnegative at the accepted count in 8 of 63 states (12.7%).

Stochastic client gradients and finite local mini-batches are addressed in Proposition F.15 and Corollary F.16. Empirically, the method is evaluated with neural networks, including ResNet, Swin-T, and LSTM, across image, ECG, and text tasks.

2. Stationarity, total client participation, and interpretation of the bounds.

Stationarity. Theorem F.8 in Appendix F.3 establishes

\leq \frac{f(x^0)-f_{\inf}}{\gamma c_\Delta T}.$$ For an iterate $x^R$ selected uniformly from these $T$ iterates, this gives $\mathbb{E}\\|\nabla f(x^R)\\|^2\leq\varepsilon^2$ after $$T_\varepsilon=\left\lceil\frac{f(x^0)-f_{\inf}}{\gamma c_\Delta\varepsilon^2}\right\rceil$$ rounds, i.e., $O(\varepsilon^{-2})$ for the theorem's fixed constants. For clarity, we have added an explicit reference to Theorem F.8 in the main theoretical analysis (Section 4). **Total participation.** Equation (49) counts the client-update exchanges of the reference algorithm, including auxiliary updates: $$C_{\rm updates}(T)=\sum_{t<T}m_{t+1}+\sum_{s\in\mathcal{T}_\Delta}b_{s+1/2},$$ where $\mathcal{T}_\Delta$ contains the adaptation rounds. In particular, $C_{\rm updates}(T_\varepsilon)\leq M(T_\varepsilon+\lceil T_\varepsilon/\Delta\rceil)$. The actual cost is given by the selected counts in the preceding sum, including the amortized auxiliary cost. To compare communication at the same predictive accuracy, we have extended **Appendix D.7, Table 22 and Figure 9**. At reference accuracy **0.836**, ISP uses **14,343** communications, versus **17,940 / 16,147 / 26,275 / 39,600 / 62,250** for fixed Uniform counts $m=15/20/25/40/50$. ISP therefore uses **11.2% fewer** communications than the most efficient fixed count reaching this accuracy; $m=5,10$ do not reach it. Intermediate updates are included. This comparison retains Uniform client selection and isolates the benefit of adapting its count. ISP also composes with other selectors (Section 5.2, Table 2). **Gradient norms.** These terms quantify how uncertainty in a sampled update translates into uncertainty in its loss response. In Theorem 4.4, the global and local gradient norms determine the sensitivity of the loss to changes in the update, while differences between client gradients measure heterogeneity. Greater heterogeneity increases the response-error bound; a larger auxiliary pool or more Monte Carlo samples reduces the corresponding error term. In Theorem 4.5, the gradient-dependent drift term also accounts for the model's evolution while the count is reused. The theory thus explains how auxiliary participation, estimation depth, and adaptation frequency control the reliability of the progress-based decision that ISP makes from evaluated candidate responses. We ablate these parameters, together with count-search resolution, in Appendix E.
AuthorsOpenReview ↗

Evaluation Overhead, Update Reuse, and Client Selection (Part 2/2)

We continue with the reviewer's third, fourth, and fifth concerns.

3. Objective evaluation and communication cost.

The expression in Algorithm 2 is not a typo: subtracting f(xτ)f(x^\tau) measures the loss change relative to the current model. The count is reconsidered only once every Δ\Delta rounds (RQ2), and the baseline objective is evaluated once per adaptation and reused across candidate evaluations (Appendix A).

ISP targets the client-update communication bottleneck through the uplink cost in Definition C.1, following Konečný et al., Federated Learning: Strategies for Improving Communication Efficiency. Objective evaluation requires client-side inference and returns scalar losses, whose payload is negligible relative to a model update. As discussed in RQ3, when public data provide a server-side surrogate f^\hat f, as in the ECG experiment, candidate evaluation requires no client communication.

These evaluation steps are part of the overall procedure assessed in our experiments. The reported update counts include intermediate communications, with savings across client selectors (Table 2), compression strategies (Table 18), and aggregation methods (Table 10). Tables 2 and 10 also report reductions in overall training time, accounting for the computational cost of the procedure.

When f^\hat f is unavailable, we empirically evaluate the GP add-on to reduce client-side objective queries and their associated model downlinks (Definition C.2). Candidate evaluation need not involve all clients: the add-on uses a proxy-coverage evaluator cohort (Appendix C.6, “Proxy-coverage evaluator cohort”) to estimate the baseline and candidate losses. In the experiments in Table 13, these cohorts contain only 5–6 of the 100 clients (5–6%) and approximate the full-client loss response.

Finally, the “Transferred updates” paragraph in Appendix D.2 accounts for both uplink and downlink traffic, including candidate evaluation. The added total-transfer comparisons cover optimization methods, client selection, and domains (Table 15), as well as different fixed client counts (Figure 5), demonstrating communication savings with evaluation traffic included.

4. Repeated gradient computations in Algorithms 1 and 2.

Algorithm 2 collects intermediate updates only when τ mod Δ=0\tau\bmod\Delta=0; at other rounds, it retains the count and the selected clients compute their updates once. At adaptation rounds, Algorithms 1–2 separate auxiliary estimation and training for the theoretical formulation, whereas the implementation reuses the updates from Iτ+1/2I_{\tau+1/2} for the server update and retains the selected count for subsequent regular rounds. We have made this reuse explicit in Algorithm 3, Appendix A, with caching specified in Algorithm 4. Intermediate communications remain included in the reported totals: Table 8 separately reports their contribution to the overall communication cost.

5. Selecting client identities as well as their number.

Yes. ISP is designed to work jointly with methods that determine which clients participate. ISP provides the count mm, and the client sampling strategy Sτ(A,m)\mathcal{S}_\tau(A,m) selects the identities from pool AA (Definition 3.1; Algorithms 1–3 and Appendix A). Section 5.2 and Table 2 demonstrate this composition with Uniform, Power-of-Choice, FedCor, FedCBS, DELTA, and Oort, showing that ISP's adaptive participation counts work with these statistical and system-aware client-selection mechanisms while preserving their selection rules.

We thank the reviewer again for the questions, which helped us clarify the convergence result, strengthen the matched-accuracy comparison, and make update reuse explicit in the practical pseudocode.

Thank you for your rebuttal and additional analyses. However, I believe my main concerns remain unresolved and would like to maintain my original score. In particular, the upper bound on total client participation M(T+T/Δ)M (T + T/ \Delta) is essentially a trivial worst-case bound. It does not establish a non-trivial communication-complexity advantage over fixed-count uniform sampling. While the additional experiments are useful, they do not resolve this theoretical limitation.

Thank you for your follow-up and for acknowledging the additional experiments.

We agree that the worst-case upper bound on total client participation does not, by itself, establish a non-trivial communication-complexity advantage over fixed-count uniform sampling. Our theoretical results establish convergence and characterize the behavior of the adaptive count-selection mechanism under the stated assumptions, rather than guaranteeing superior communication complexity in general.

We address the communication-efficiency question empirically, through comparisons at matched accuracy across different fixed participation budgets (Appendix D.7, Table 22 and Figure 9). At the target accuracy of 0.836, ISP requires 14,343 client-update exchanges, compared with 16,147 for the most communication-efficient fixed-count Uniform baseline reaching this target, corresponding to an 11.2% reduction. Importantly, these counts include the additional auxiliary communication introduced by ISP.

The experiments across different client-selection strategies (Table 2) further support the practical benefits of adaptive participation.

We acknowledge that these empirical improvements do not constitute a general theoretical guarantee. Nevertheless, we believe that the combination of convergence analysis and experimentally demonstrated communication savings provides meaningful support for the proposed approach.

We thank the reviewer again for the constructive discussion.

Reviewer_eGjGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ISP, which adapts the number of participating clients in federated learning. It uses updates from an auxiliary pool to search for the smallest count predicted to reduce loss. This count is reused for several rounds, with a new client group sampled each round. The paper analyzes estimation error and count reuse, and evaluates communication and accuracy across several tasks, client selection methods, and optimizers.

Critical strengths
  1. Equation (2) seeks the smallest client count with negative expected loss change.
  2. Theorems 4.4–4.5 address estimation error and count reuse for the reference algorithm.
  3. Fixed-count sweeps and ablations of pool size, sampling depth, reuse interval, and EMA complement tests across selectors and optimizers.
Critical weaknesses
  1. The benchmarks do not measure the expected loss change Φτ([M],m)\Phi_\tau([M],m) at the chosen mm. Table 12 evaluates loss on all clients, but its candidate updates still come from the auxiliary pool.
  2. The paper does not report download costs for direct ISP (Table 11). Clients must download candidate models to evaluate them on local data, but Comm. counts only update uploads.
  3. The experiments lack a comparison with baselines like FedTune by Huanle Zhang et al. [1], which adjust client counts using training progress and resource costs. This comparison is needed to assess whether candidate-model evaluation improves the accuracy–communication trade-off over existing feedback-based control.

[1] Zhang et al. “Federated Learning Hyperparameter Tuning From a System Perspective.” IEEE Internet of Things Journal, 2023.

Questions for authors
  1. At fixed checkpoints, how often does the chosen mm give Φτ([M],m)≥0\Phi_\tau([M],m)\geq 0, estimated using repeated samples from all MM clients?
  2. Does ISP still save communication at the same accuracy when both uploaded and downloaded bytes are counted?
  3. Under matched training settings and comparable tuning, how does ISP compare with FedTune in accuracy at the same total communication cost, including candidate-model evaluation?
3 comments
AuthorsOpenReview ↗

Full-Population Response Validation, Bidirectional Communication Costs, and FedTune Comparison

We thank the reviewer for the questions. We added experiments on all three points. References refer to the revised manuscript; additions are highlighted in blue.

1. Full-population response at the selected count.

The new “Full-population response evaluation” paragraph in Appendix G (Figure 23) retains Section 5.1's ISP training setup. We audit 63 states throughout training, every Δ=20\Delta=20 rounds. At each audit, we additionally collect updates from all M=100M=100 clients to compare Φ^τ([M],m)\widehat\Phi_\tau([M],m) using N=150N=150 full-population cohort samples with Φ^τ(Iτ+1/2,m)\widehat\Phi_\tau(I_{\tau+1/2},m) using N=10N=10. Let mI⋆m_I^\star and mM⋆m_M^\star be the first counts with negative auxiliary-pool and full-population estimates, respectively.

At the count mI⋆m_I^\star accepted by ISP, Φ^τ([M],mI⋆)≥0\widehat\Phi_\tau([M],m_I^\star)\geq0 in 8/63 states (12.7%). These changes are 0.096–0.140% of the baseline loss f(xτ)f(x^\tau). All eight estimates Φ^τ([M],mI⋆+1)\widehat\Phi_\tau([M],m_I^\star+1) are negative. The boundary error BE=∣mI⋆−mM⋆∣\mathrm{BE}=|m_I^\star-m_M^\star| averages 1.17 clients, with BE≤2\mathrm{BE}\leq2 in 59/63 states (93.7%).

2. Communication including downloads.

ISP seeks the smallest count predicted to decrease the loss. Its communication objective is the upload-based cost in Definition C.1, including auxiliary updates. Theorem F.12 establishes convergence and bounds this cost, including intermediate updates, in its reference regime. This objective addresses the upload bottleneck emphasized by Konečný et al., or settings where public data provide a server-side surrogate f^\hat f that eliminates candidate-evaluation traffic, as in our ECG experiment (RQ3).

When f^\hat f is unavailable, we propose the GP add-on to reduce evaluation downlinks (Definition C.2): response history and round context guide candidate queries, while proxy coverage selects a small evaluator cohort (Appendix C.6). Thus, the bidirectional experiments use ISP-GP, with k=50k=50 on CIFAR-10. The revised “Transferred updates” paragraph in Appendix D.2 provides the detailed rationale and accounting: total transfers combine uploaded and downloaded bytes, including intermediate updates and candidate evaluation, normalized by model size. For a common model payload, this is equivalent to comparing total bytes.

Table 15 reports total transfers:

SettingMatched AccuracyFixed-count baselineISP-GP
CIFAR-10 / FedAvg0.82629,70728,468
CIFAR-10 / FedNova0.84633,56033,688
CIFAR-10 / FedCBS0.82532,66731,005
Shakespeare / FedAvg0.53127,45025,710

Savings are 4.2%, 5.1%, and 6.3% for FedAvg, FedCBS, and Shakespeare; FedNova traffic is nearly unchanged (+0.4%). Figure 5 additionally compares ISP-GP at matched accuracy against the fixed client counts studied in Appendix D.7. In ECG, the public server-side surrogate permits direct ISP evaluation without candidate-model downloads: at matched G-mean 0.882, total transfers decrease by 52.7% (Table 16).

3. Comparison with FedTune.

We added Appendix D.5, Table 20, following Section 5.1's CIFAR-10 setup. FedTune-ΔM\Delta M adapts counts at one local epoch; FedTune-ΔMΔE\Delta M\Delta E adapts both counts and epochs. In the latter regime, ISP-GP uses the same epoch controller. FedTune's accuracy threshold is ε=0.01\varepsilon=0.01. Comm weights communication time and volume; Comp also weights computation time. GP settings follow D.2 with k=50k=50.

Table 20 reports matched-accuracy total transfers, including evaluation traffic, and runtime:

MethodMatched AccuracyTotal transfersTraining time (hh:mm:ss)
Fixed epochs: Comm
FedTune-ΔM\Delta M-Comm0.82631,64807:53:35
ISP-GP-FedAvg0.82628,46805:50:29
Adaptive epochs: Comm
FedTune-ΔMΔE\Delta M\Delta E-Comm0.82511,31003:16:32
ISP-GP-FedTune-ΔE\Delta E-Comm0.82510,95203:17:17
Adaptive epochs: Comp
FedTune-ΔMΔE\Delta M\Delta E-Comp0.83125,60605:09:51
ISP-GP-FedTune-ΔE\Delta E-Comp0.83115,93203:45:10

The ISP-based configurations use fewer total transfers at matched accuracy, with lower runtime for fixed epochs and adaptive-epoch Comp, and comparable runtime for adaptive-epoch Comm. Unlike FedTune's task-dependent accuracy trigger, ISP continues to reassess participation as accuracy gains diminish, with controls that directly specify evaluation resources, reassessment frequency, and responsiveness (Appendix D.5).

We thank the reviewer again for the questions, which helped us assess the selected counts against full-population responses, document communication costs in both directions, and extend the comparison with adaptive participation methods.

Thank you for the response. My first concern has been addressed. However, while the authors recommend ISP-GP to reduce communication costs, direct ISP appears to be the default method in Table 2. Its download costs should still be reported. In the revision, the authors should also include a Table 2-style evaluation of ISP-GP across the same client-selection baselines, covering both upload and download costs.

Given these remaining issues, the current experiments do not fully convince me of the advantages of either version. I will maintain my rating.

Thank you for confirming that the first concern has been addressed. To supplement our previous response, upload-based communication evaluation is also used in recent work: LoCoDL (ICLR 2025) explicitly focuses on uplink communication, while FedBAT (ICML 2024) motivates this choice by bandwidth asymmetry. Nevertheless, we agree that bidirectional communication is important and evaluate it using ISP-GP.

Additional ISP-GP experiments across client-selection strategies.

Following your suggestion for a Table 2-style comparison, we additionally evaluated ISP-GP with POW-D, DELTA, FedCor, and Oort, complementing the existing results for Uniform and FedCBS (Table 15). The table below consolidates these results across all six client-selection strategies, accounting for both uploads and downloads, including candidate evaluation. Due to limited experimental time, standard deviations are not yet available for the additional runs.

MethodAccuracyUploadsDownloadsTotal transfers
POW-D0.81210,34710,34720,694
ISP-GP-POW-D0.8388,9409,33318,273
DELTA0.72119,70019,70039,400
ISP-GP-DELTA0.71614,46216,01530,477
FedCor0.84819,36019,36038,720
ISP-GP-FedCor0.84312,17316,89229,065
Oort0.83419,50019,50039,000
ISP-GP-Oort0.85112,46013,38425,844

Total transfers include bidirectional model communication, including candidate evaluation.

ISP-GP reduces total transfers by 11.7% for POW-D, 22.6% for DELTA, 24.9% for FedCor, and 33.7% for Oort. Notably, POW-D and Oort also achieve higher accuracy (+2.6 and +1.7 percentage points, respectively). For DELTA and FedCor, the absolute accuracy differences are negligible (approximately 0.005), while communication reductions exceed 20%.

Together with the previously reported Uniform and FedCBS results, these experiments provide further evidence that ISP-GP's communication benefits extend across different client-selection strategies even when bidirectional traffic is accounted for.

We thank the reviewer again for the suggestion, which helped strengthen the practical evaluation.

Reviewer_3sbaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies adaptive client participation count selection in federated learning and proposes ISP. The method constructs an auxiliary client pool, estimates the next-round loss change for different candidate participation counts, and selects the first candidate count with a negative estimated loss change to adjust the communication budget. To reduce search overhead, ISP re-runs the search every few rounds and reuses the selected count in the remaining rounds. The paper provides a theoretical analysis and validates the performance of the proposed approach through experiments.

Critical strengths

The research problem is of practical significance. The paper focuses on the number of participating clients per round, rather than only on selecting which clients to participate under a fixed budget. ISP dynamically adjusts the number of participants using an optimization-progress signal, and can therefore be combined with various existing client-selection strategies. This design targets the practical trade-off between communication cost and optimization performance, and has certain application value.

Critical weaknesses

The paper's current experiments make it difficult to distinguish whether ISP's gains stem from its core loss-response-guided count selection or from the overall framework, including periodic search, the auxiliary pool, and count reuse. Meanwhile, the coverage of other dynamic client-count rules is also limited.

Questions for authors

If ISP’s periodic search, auxiliary-pool construction, and count-reuse mechanisms are kept unchanged (or at least held at the same auxiliary communication budget and reuse window), and only the loss-response-guided count-selection criterion is replaced with random selection, a loss- or gradient-based heuristic, or a predefined dynamic schedule, can ISP still achieve similar communication–performance advantages? Could the authors provide or discuss such mechanism-isolation ablation experiments to demonstrate that the main gains indeed come from the progress-guided criterion rather than the overall search and reuse framework?

2 comments
AuthorsOpenReview ↗

Isolating ISP’s Count-Selection Criterion and Comparing Dynamic Client-Count Methods

We thank the reviewer for recognizing the practical relevance of adaptive participation and its compatibility with client selection, and for suggesting experiments that isolate the count-selection criterion. We have added these comparisons. References refer to the revised manuscript; additions are highlighted in blue.

1. Does the gain come from the loss-response criterion or the surrounding procedure?

In Appendix E.7, we compare four alternative count rules within the same practical ISP framework, using Section 5.1's CIFAR-10 setup:

  • ISP-Random: sample the count uniformly from the admissible auxiliary-pool range.
  • ISP-Permuted: randomly reorder the counts from a reference ISP run, preserving their empirical frequencies and reuse window.
  • ISP-Scheduler: use a linearly increasing, AdaFL-style schedule over the minimum and maximum counts of that reference run, updated at the same adaptation events.
  • ISP-Gradient: choose the smallest mm satisfying Qτ(I,m)=EC∼Sτ(I,m)∣gτ(C)−gτ(I)∣2≤θ∣gτ(I)∣2Q_\tau(I,m)=\mathbb{E}_{C\sim\mathcal{S}_\tau(I,m)}\\|g^\tau(C)-g^\tau(I)\\|^2\leq\theta\\|g^\tau(I)\\|^2, with I=Iτ+1/2I=I_{\tau+1/2} and θ=1\theta=1.

Permuted and Scheduler deliberately use participation information from an ISP run: they test whether its count distribution or increasing trend can replace adaptation to the current optimization state.

The matched-accuracy block of Table 29 reports communications including intermediate client updates:

MethodCommunicationsTest lossMatched Accuracy
Uniform16,1470.4680.836
ISP-Uniform14,3430.4640.836
ISP-Random37,2740.5230.836
ISP-Permuted14,5020.4710.836
ISP-Scheduler16,2380.4950.836
ISP-Gradient21,1060.5570.836

ISP reaches the target with fewer communications than ISP-Scheduler, supporting adaptation to the optimization state over a predefined increasing trend even when the schedule uses ISP's participation range. ISP-Permuted has a small communication gap at matched accuracy, but exhibits repeated loss spikes and accuracy drops (Figure 16), highlighting the role of adaptive count dynamics beyond the empirical count distribution. ISP-Random frequently selects large cohorts and requires substantially more communication.

ISP-Gradient selects larger client cohorts (Figure 16): its criterion is more conservative because it controls update error in all directions. Under heterogeneity, client gradients may remain dispersed as the pool-gradient norm decreases, making its relative-error test more demanding. ISP instead evaluates the loss response directly and can accept a smaller cohort when its update still predicts progress. Appendix E.7 gives the definitions and detailed interpretation of these results. The additional ISP-RI objective is discussed in the same section (Table 30).

2. Coverage of other dynamic client-count methods.

We have added FedTune in Appendix D.5, Table 20, complementing the AdaFL comparison in Table 19. Under the same CIFAR-10 training setup, we examine two modes:

  • FedTune-ΔM\Delta M: adapt the number of clients with one local epoch; compare against ISP's count adaptation.
  • FedTune-ΔMΔE\Delta M\Delta E: adapt the number of clients and local epochs; compare against ISP coupled to the same FedTune epoch controller.

Without a server-side surrogate, evaluating candidate models requires client queries and model downloads. We therefore use ISP-GP, which reduces these queries through GP-guided candidate evaluation and a small proxy-coverage evaluator cohort, while retaining the loss-response criterion (Appendices C.6 and D.2). We report overall communications, including uploads, downloads, and candidate evaluation. Comm weights communication time and volume; Comp also weights computation time. Selected matched-accuracy results from Table 20 are:

MethodMatched AccuracyTotal transfersTraining time (hh:mm:ss)
Fixed epochs: Comm
FedTune-ΔM\Delta M-Comm0.82631,64807:53:35
ISP-GP-FedAvg0.82628,46805:50:29
Adaptive epochs: Comp
FedTune-ΔMΔE\Delta M\Delta E-Comp0.83125,60605:09:51
ISP-GP-FedTune-ΔE\Delta E-Comp0.83115,93203:45:10

ISP reduces communications and runtime in these comparisons; adaptive-epoch Comm also reduces communications with comparable runtime (Table 20). Unlike FedTune's task-dependent accuracy trigger, ISP continues to reassess participation as accuracy gains diminish, with controls that directly specify evaluation resources, reassessment frequency, and responsiveness (Appendix D.5).

We thank the reviewer again for the suggestion, which helped us distinguish the contribution of the progress-guided criterion from the auxiliary-pool and count-reuse procedure.

Thank you for the author response. After reviewing the feedback and the manuscript, my doubts have been resolved, and I will maintain my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.