acceptodds
Under review as a conference paper at ICLR 2027

When Do Stale Updates Help? Forecasting, Participation Weighting, and Step Scale in Federated Learning

Abstract

Stale client updates can reduce the variance of federated aggregation, but accuracy gains do not by themselves identify the value of forecasting: participation weighting and aggregate scale can change simultaneously. We analyze these effects through a forecast–residual estimator and use OPTICAST, which learns client-specific linear forecasts, as a case study. Exact conditional moments expose the coupling between propensity error and forecast residuals. A complementary decomposition separates aggregate weight mass from relative client weighting, and an aggregation-aware damping criterion identifies the optimal scalar correction at a frozen history. Across the four main image-classification settings, OPTICAST achieves higher mean accuracy than FedAU, MIFA, U-FedVARP, and FedStale. The control experiments qualify this result. Across three CIFAR-10 participation regimes, last-update reuse yields paired gains of 0.14, 1.06, and 3.51 percentage points over zero forecasting as participation decreases, whereas a separate five-seed test does not establish an advantage of learned gains over simple reuse. On FEMNIST, a one-seed oracle scale intervention recovers 14.34 points of a 16.54-point zero-control-to-participating-mean gap. Learned and damped configurations also collapse on Shakespeare. These findings distinguish configured-method performance from evidence for predictor complexity, and motivate evaluating forecasting jointly with participation weights, aggregate scale, implementation consistency, and stability.

Primary area:
learning theory
Keywords:
Federated learning, partial client participation, missing-update forecasting, Horvitz–Thompson estimator, propensity estimation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hQB2OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper analyzes the aggregate mass and relative weighting, connect the moments and a forcast-error-dependent convergence bound to frozen-state checks. Empriical results shows the stability of the method.

Critical strengths
  • Extensive results
Critical weaknesses
  • The paper is difficult to read. The problem definitions, methodology and design insights are unclear to readers.
  • The connections across Equations are unclear. It is hard to understand the insights behind these theoretical formulas and analysis, e.g., why did the author study the Eq. (1) and how it relates to Eq. (2)?
  • The structure of the paper is very difficult to follow. The author should provide a big picture about the literature review and the high-level definitions of the problems before going deeper into the derivations.
Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer hQB2

We thank the reviewer for the time spent on our paper. We revised part of the paper's presentation as you suggest. All changes are in the uploaded revision and are printed in blue.

PointWhat we didWhere in the revision
W3: literature big pictureNew "Background" paragraph: reweighting and caching methods before any equationOpening of §2
W1, W3: big picture before derivationsNew "Roadmap" paragraph: the three questions and where each is answeredEnd of §1
W2: why Eq. (1), and how it relates to Eq. (2)One sentence after each equation§2
W2: insight behind the formulasA plain-language "In words" sentence before Proposition 1, Proposition 2 and Corollary 1§3
W1: definitionsNotation tableTable 6, App. A

On the summary. One correction, because it bears on soundness: the paper does not claim that the results show stability. §5.5 and App. D.6 report that OPTICAST collapses on Shakespeare, and the revision adds an 18-run study of these collapses (App. D.6, Tables 21–22). The paper's conclusion (§6) is that simple historical forecasts help under the evaluated sparse CIFAR-10 protocol, that additional learned complexity is not established, and that normalization and scale are central to interpreting large gains. The abstract now states that we do not recommend the learned configuration over last-update reuse.

W2, Eq. (1) and Eq. (2).

  • Eq. (1) is the object of study. The server update is a forecast for every client plus inverse-propensity-weighted corrections from the clients that report. Every method in Table 1 has this form with its own forecast and correction weight (Eq. (1) uses the weight 1/q), so it lets us change the forecast while holding the weighting fixed.
  • Eq. (2) is the rule that produces the weights q used inside Eq. (1).
  • §3 asks what the forecast changes when q comes from Eq. (2) instead of equalling the true probability p. The answer is that it changes the mean of the update (Eq. 3), not only its variance.

W1 and W3, structure.

  • Roadmap. The new paragraph states the three questions: does forecasting help at fixed weighting, do learned forecasts beat reusing the last update, and do the answers survive normalization and scale controls. It names the section that answers each.
  • "In words" sentences. Each gives the message of a result before its formula. For Proposition 2: multiplying the update by a constant can fix its size but not which clients it over-represents.
  • Background. A new paragraph opening §2 places participation-aware reweighting (FedAU) and update caching, without reweighting (MIFA) or with inverse-propensity weights (U-FedVARP, FedStale, FedSteer), before any equation and states the question the paper isolates; Table 1 then writes these methods in one form.

If a specific soundness concern remains, we would be grateful to have it named so that we can address it directly. Do these changes address your concerns? Please leave any further comments; we will respond during the discussion period.

AuthorsOpenReview ↗

Follow up on our response

Thank you again for your suggestions on the presentation. We wanted to follow up on the revised roadmap and the explanation of how Equations (1) and (2) fit together. If you have had a chance to look at the revision, we would appreciate hearing whether these changes make the motivation and derivations easier to follow. If any specific concern about the analysis remains, we would be glad to address it directly.

Thank you for your time.

Reviewer_3gUgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates when stale-update forecasting benefits federated learning by distinguishing the effects of forecasting, participation weighting, and aggregate scale. It presents a forecast-residual aggregation framework and OPTICAST, which learns client-specific linear forecasts from historical updates. Conditional moment analysis, aggregation-weight decomposition, and convergence analysis are provided. Experiments on four tasks reveal both potential benefits of historical updates and limitations of learned forecasting.

Critical strengths
  1. The paper investigates an interesting problem concerning the effectiveness of stale-update reuse under heterogeneous client participation.

  2. The theoretical analysis provides useful insights into the interactions between forecasting errors, participation weighting, and aggregation scale.

  3. The experiments include multiple datasets, participation settings, and relatively detailed control studies. The paper carefully reports negative results and acknowledges limitations in experimental attribution and stability.

Critical weaknesses
  1. The main concern is novelty. The forecast-residual estimator follows the established augmented Horvitz–Thompson framework, while several theoretical results are relatively direct conditional-moment or algebraic decompositions. The additional technical contribution beyond existing approaches remains limited.

  2. The theoretical guarantees do not directly establish the effectiveness of OPTICAST. The convergence bound depends on forecast residuals and propensity errors, without proving the prediction quality or stability of the proposed learned forecasting method.

  3. The necessity of learned forecasting is insufficiently demonstrated. In Table 4, simple last-update reuse slightly outperforms OPTICAST on CIFAR-10, and Table 5 shows no significant improvement from learned gains. The additional complexity of OPTICAST therefore remains questionable.

  4. The experimental comparisons have unresolved implementation inconsistencies. Several configurations were evaluated using different source revisions. This weakens the reliability of attributing performance differences to individual components.

  5. The stability and computational efficiency are concerning. OPTICAST collapses in two of three Shakespeare runs and incurs approximately 62% longer training time than FedAU on FEMNIST. These limitations substantially constrain its practical applicability.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3gUg

We thank the reviewer for a careful review and for noting that we report negative results. All changes are in the uploaded revision and are printed in blue.

PointWhat we didWhere in the revision
W1, noveltyNew identity for forecasts that depend on the current reportsApp. A.7
W2, theory vs. OPTICASTBounded-harm remark§3, "Connection to optimization"
W3, need for learned forecastingCold-start-matched control; "not recommended" in the abstractApp. G.9, Table 37; §5.2; abstract; §6
W4, source revisionsRevision audits; new cells at one revisionApp. G.8; App. G.4–G.5
W5, stabilityNew 18-run Shakespeare panelApp. D.6, Tables 21–22; §5.5
W5, costDecomposition of the 62%§5.5, Table 23

W1, novelty. We agree the estimator is the augmented Horvitz–Thompson form; §2 says so. The paper analyses when stale-update forecasting helps; it is not a new-estimator paper. Not in the cited work:

  1. Proposition 2. Rescaling restores the full-participation mean for every set of updates if and only if p/q is constant on clients with positive weight. The two-client example shows a displaced stationary point that no server learning rate fixes.
  2. Corollary 1. A closed-form, aggregation-aware damping criterion. Its optimum is interior, and along training it ranges from 0.34 (early sparse training, where full reuse is worse than no forecast) to 0.94 (App. G.6, Table 36).
  3. The bias term of the convergence bound, which need not vanish under clipped propensities.
  4. New (App. A.7): a bias identity for forecasts that depend on the current reports. The bias is −Σᵢ (wᵢ/pᵢ)·Cov(Iᵢ, Fᵢ given H). Applied to FedSteer's released code, it gives a bias term that is generally nonzero even at known propensities (assuming mean independence), while its Algorithm 1 has none (mean only; size not measured). Exact enumeration confirms the identity to 7×10⁻¹⁶.

Empirically, FEMNIST forecast gains of +2.0 to +29.7 pp under HT become −0.71 to +1.00 pp under residual normalization (App. G.5): mostly compensation for lost aggregate mass.

W2, theory vs. OPTICAST. Agreed; §3 states that the bound does not establish OPTICAST's stability. We added one statement that does hold: OPTICAST's forecast is norm-clipped, so its squared residual is at most 2·E‖Δ‖² + 18·E[ν̄²], and forecasting cannot inflate the variance term without bound. This is not a stability guarantee, as Shakespeare shows. The paper measures the quantities the bound depends on instead (App. G.6–G.7).

W3, necessity of learned forecasting. We agree; this is the paper's finding. New evidence:

  • Cold-start control (App. G.9, Table 37). At the archived FEMNIST setting, OPTICAST led reuse by +1.75 [+0.17, +3.33] pp. When reuse gets the learned rule's cold start, the lead falls to +0.10 [−0.14, +0.34] pp, so, as predeclared, it is attributed to the cold-start rule. Leads at higher floors (Table 34) were not cold-start-matched.
  • Response surface (App. G.4). Learned − reuse shows no clear advantage at any interval once dense participation has five seeds, and is −2.37 and −8.37 pp at the two sparsest.
  • The abstract and §6 now state that "we do not recommend the learned configuration over last-update reuse".

W4, implementation inconsistencies. Two audits (App. G.8):

  • All 122 central archived comparisons are implementation-equivalent for the claims they support; 6 use different random-number streams.
  • The one unexplained discrepancy (App. F.2) first diverges at a commit that changed random-number consumption.

The new participation intervals and floors (App. G.4–G.5) ran at one frozen revision with all their arms; e.g. reuse − zero on sparse CIFAR-10 is +3.90 and +5.00 pp at the new floors 0.02 and 0.10 (Table 32).

W5, stability (App. D.6, Tables 21–22). 18 runs, seeds 12–14, protocol fixed before the first run:

  • A same-revision rerun of the original configuration collapsed at one seed of three, not one that collapsed in the archive.
  • At client lr 0.25 and with forecast-norm multiplier 1, no run collapsed. With 3 of 6 original runs collapsing, three seeds cannot show stabilisation, and we do not claim it.
  • Without a collapse, OPTICAST is level with reuse: −0.19 [−1.38, +1.00] pp at lr 0.25.

W5, cost (§5.5, Table 23).

  • Last-update reuse in the same code, which fits nothing, takes 162.4 min vs OPTICAST's 176.1: learning the gains adds 13.7 min (8%).
  • The remaining 53.7 min over FedAU (108.7) are the overhead of the shared code path, which in these runs keeps int8 histories on the CPU and 1.9 GiB of logging state.
  • U-FedVARP, with a client state of similar size, takes 112.9 min.

Do these answers address your concerns? If anything remains unclear, please leave a further comment; we will respond during the discussion period.

AuthorsOpenReview ↗

Follow up on our response

Thank you again for your feedback. We wanted to follow up on our response, especially the revisions clarifying the paper’s contribution relative to prior work and the matched comparison between learned forecasting and last-update reuse. If you have had a chance to look at these changes, we would appreciate hearing whether they address your main concerns. We would be happy to explain any part of the response in more detail.

Thank you for your time.

Thank the authors for the detailed responses. I have reviewed the revised paper carefully. The additional experiments and implementation audits are helpful, particularly the cold-start and aggregation-scale analyses. However, the new results further weaken the necessity of learned forecasting, while the theoretical and stability concerns remain. I appreciate the authors' efforts during the rebuttal, but I will maintain my rating at 2.

AuthorsOpenReview ↗

Response to Reviewer 3gUg (round 2)

Thank you for reading the revision carefully and for noting that the additional experiments and implementation audits are helpful.

Your review credited the paper for carefully reporting negative results and acknowledging limitations; two of the remaining points, necessity and stability, are such reported limits, not errors.

Necessity of learned forecasting. The paper asks when stale-update forecasting helps, and reporting that the learned component's added value is not established is part of its answer, stated in the abstract and §6. OPTICAST as configured still compares well with the published methods:

  • It has the highest mean accuracy against FedAU, MIFA, U-FedVARP, FedStale and FedSteer in all four main settings (Table 3). This is a descriptive ordering of configured implementations, not tuned per method, and on FEMNIST the large forecasting gains are mostly aggregate scale (§5.3).
  • In the prospectively registered test on moderate CIFAR-10, it beat FedAU at every one of five fresh seeds (+1.76 pp, one-sided p = 0.017, Table 5).

Theory. The analysis gives results that are not in the cited work and that bear directly on how such methods are evaluated, together with checks of them:

  • Proposition 2: rescaling the zero-forecast update restores the full-participation mean for every set of client updates if and only if p/q is constant on clients with positive weight. In the two-client example, a server learning-rate change corrects speed but not the displaced stationary point.
  • Corollary 1: a closed-form, aggregation-aware damping criterion. In a numerical check at a frozen CIFAR-10 checkpoint it agrees with a fine-grid search to within 3.0×10⁻⁷, and the optimum is interior.
  • App. A.7: a bias identity for forecasts refitted to the current report. Applied to FedSteer's released code, it gives a generally nonzero bias term even at known propensities (assuming mean independence, which we did not verify for that code), which FedSteer's Algorithm 1 does not have; exact enumeration on synthetic fixtures matches the identity to 7×10⁻¹⁶. This concerns the conditional mean, not its size in training, and is not a claim that FedSteer's theorem is wrong.
  • The exact moments match measurement: at a frozen CIFAR-10 checkpoint, with local-update noise held at zero, predicted-to-measured variance ratios are 0.992 to 1.01 over 10,000 sampled masks.

We do not claim that the convergence bound guarantees OPTICAST's stability (§3); its role is diagnostic.

Stability. §6 states that the instability on Shakespeare prevents an unconditional recommendation of the learned configuration, and its cause is unresolved. At its original setting OPTICAST collapsed in 3 of 6 runs; with client learning rate 0.25 or forecast-norm multiplier 1, none of its six runs collapsed. With three seeds per change we do not claim a fix.

Thank you again for engaging with the revision.

Reviewer_y2AhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how stale-update forecasting interacts with participation weighting and aggregate scale in federated learning. It derives conditional bias and variance expressions, separates aggregate mass from relative client weighting, and gives an oracle criterion for damping stale updates. Experiments compare OPTICAST with existing methods and simpler controls, finding benefits from historical reuse but limited evidence for learned forecasting.

Critical strengths
  1. The analysis separates aggregate scale from relative client weighting. It shows why correcting the total update weight need not restore the intended population objective.
  2. The conditional moment analysis connects forecast residuals with imperfect participation weights. It explains why forecasting can change both bias and variance, providing a stronger basis for evaluation than endpoint accuracy alone.
Critical weaknesses
  1. Practical robustness remains insufficiently established. Participation is simulated using independent Bernoulli draws with fixed, index-assigned probabilities. The evaluation therefore does not test changing or temporally correlated availability, where stale updates may behave differently. OPTICAST also collapses on two of three Shakespeare seeds, with the cause unresolved. Together with the lack of an established advantage over simple reuse in the prospective CIFAR-10 comparison, these results leave its practical benefit uncertain.
  2. The comparisons do not fully isolate the effects under study. The evaluated implementations differ in storage precision, forecast clipping, and, for OPTICAST, cold-start behavior. Appendix F.2 also documents historical code revisions and an unexplained discrepancy between nominally matched configurations. These differences limit attribution of the observed gains to forecasting or participation weighting. The controlled-input tests do not establish equivalence of the historical training pipelines.
Questions for authors
  1. Do the main forecasting gains persist under one fixed implementation with matched precision, clipping, cold-start rules, and update order?
  2. What causes the Shakespeare collapses, and do they persist after reasonable learning-rate and damping adjustments?
  3. Does the FEMNIST scale intervention reproduce across seeds and remain informative against a zero-forecast baseline with a separately tuned server learning rate?
5 comments
AuthorsOpenReview ↗

Response to Reviewer y2Ah (part 1)

We thank the reviewer for a precise and constructive review. (Reply 1 of 2.) All changes are in the uploaded revision and are printed in blue.

PointWhat we didWhere
W1, correlated availabilityMarkov-availability experiment; theory remark§3 after Prop. 1; App. G.10, Tables 38–39; §6
W1, changing availabilityRegime-switch experiment: reuse keeps a clear gainApp. G.12, Tables 41–42; §6
W1 and Q2, Shakespeare18-run panel: same-revision rerun, client lr 0.25, forecast-norm multiplier 1App. D.6, Tables 21–22; §5.5
W1, no advantage over reuseAgreed; now in the abstract and §6Abstract; §6
W2 and Q1, one implementationCold-start-matched control; revision auditsApp. G.9, Table 37; §5.2; App. G.8
Q3, FEMNIST scaleThree seeds; tuned server-lr control: a larger step alone closes the gap; reading revisedAbstract; §5.3; App. G.3, G.11; Tables 15, 40

W1, availability model. Proposition 1 needs, given the history, reports independent of current updates and of each other, not independent rounds, so it holds with p read as the history-conditional reporting probability. Under temporally correlated availability the count estimate converges to the stationary rate instead, so the bias term persists (new remark, §3). We tested this on sparse CIFAR-10: two-state Markov availability, same stationary probabilities, lag-one autocorrelation 0.5, five arms, seeds 12–14. Mean participation and aggregate mass match the Bernoulli runs (Table 39).

Contrast, pp [95% CI]MarkovBernoulli
Last update − zero forecast+2.00 [−6.72, +10.72]+3.88 [+0.16, +7.61]
OPTICAST − last update−1.40 [−5.28, +2.48]−0.48 [−1.39, +0.42]

Every arm's mean accuracy drops. Contrary to our predeclared prediction, the forecasting arms lose more on average than the zero forecast and FedAU (−3.49 and −4.41 vs −1.61 and −1.92 pp). This partly confirms your concern: under correlated absence the benefit of reuse is smaller and uncertain; §6 says so.

W1, changing availability. A second predeclared panel reverses the probability assignment at round 100, so frequent clients become rare and rare clients frequent (same five arms, seeds 12–14, endpoint after the switch; App. G.12). Up to round 100 every run reproduces its archived Bernoulli run exactly. The count estimate lags the new probabilities: after the switch the weighted propensity mismatch is 0.55 against 0.23, and aggregate mass is 1.22 against 0.91, partly because the switch draw has 7% more reports (Table 42).

Contrast, pp [95% CI]SwitchBernoulli
Last update − zero forecast+2.61 [+1.63, +3.60]+3.88 [+0.16, +7.61]
OPTICAST − last update+0.30 [−0.31, +0.91]−0.48 [−1.39, +0.42]

By the predeclared rule, the conclusion survives this change: reuse keeps a gain whose interval excludes zero, and OPTICAST adds nothing detectable. The gain is smaller on average (paired change −1.27 [−4.82, +2.28] pp). Per arm, the zero forecast changes by +0.94 [−2.62, +4.50] and last update by −0.33 [−0.80, +0.14] pp (lower at every seed); on point estimates this fits both attributions the protocol names, and neither interval excludes zero, so we cannot say which side moved.

W1 and Q2, Shakespeare. 18 runs at seeds 12–14, with the protocol fixed before the first run:

  • Same-revision rerun of the original OPTICAST configuration: it collapsed at seed 14 only, whereas the archive collapsed at seeds 12 and 13. Which seeds collapse is not reproducible.
  • Client lr 0.25 (OPTICAST, last update, zero forecast, FedAU) and multiplier 1 (OPTICAST): no run collapsed. With 3 of 6 original runs collapsing, three clean seeds per change cannot show that it removes the failure (even pooling both changes, 0 of 6 vs 3 of 6 gives post hoc Fisher p = 0.09), so we do not claim stabilisation.
  • Without a collapse, OPTICAST is level with reuse: −0.19 [−1.38, +1.00] pp at lr 0.25.
  • Damping, from the archive (§5.5, Table 20): the trust weight, an adaptive damping of the forecast, did not collapse at the three archived seeds, but its median weight in the second half of training was 0.0014, effectively switching forecasting off; FedStale's damped reuse (α = 0.5) collapsed at two of three seeds.

The cause remains unresolved, and the revision says so.

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer y2Ah (part 2)

(Continued from the previous comment.)

W1, no advantage over reuse. We agree; this is the paper's conclusion. The abstract and §6 now state that "we do not recommend the learned configuration over last-update reuse".

W2 and Q1, one implementation.

  • Precision and clipping differ only between the reference-code baselines and our shared code. The zero / last-update / learned contrasts all run in the shared code with the same precision, clipping, estimator and update order.
  • Cold start, the one remaining difference: with the learned rule's cold start given to reuse, OPTICAST − reuse on FEMNIST falls from +1.75 [+0.17, +3.33] to +0.10 [−0.14, +0.34] pp, and is +0.14 [−0.98, +1.26] on sparse CIFAR-10 (Table 37). As predeclared, the FEMNIST gain at the archived setting is attributed to the cold-start rule.
  • Historical revisions (App. G.8): all 122 central comparisons are implementation-equivalent for their claims; the F.2 pair first diverges at a commit that changed random-number consumption (we do not claim this explains the full gap).
  • Direct answer to Q1: the new intervals and floors (App. G.4–G.5) ran at one revision with all arms. Reuse-over-zero gains persist (e.g. +3.90 and +5.00 pp at floors 0.02 and 0.10, Table 32). No learned gain over reuse appears on CIFAR-10; on FEMNIST the learned configuration still leads at floors 0.10 and 0.15 (Table 34), without a matched cold start.

Q3, FEMNIST scale. It reproduces across seeds: recovery is 86.7%, 87.2% and 89.7% at seeds 12–14 (Table 15). The second part of your question changed our reading. The oracle factor equals a server learning rate larger by about 1.66 (weight decay correspondingly smaller), so we ran the zero forecast at server learning rates 1.0, 1.5, 2.0 and 2.5, seeds 12–14, all at one revision, under a protocol committed before the first run (new App. G.11, Table 40):

Server learning rate1.01.52.02.5
Zero forecast, accuracy (%)40.1049.5854.7858.45
  • The best rate, 2.5, is the edge of the grid, so 58.45% is a lower bound for a tuned zero forecast.
  • It exceeds the participating mean (55.68%) by +2.77 [+2.46, +3.07] pp and the oracle-scaled zero forecast by +4.66 [+4.04, +5.28] pp.
  • Against it, last update and OPTICAST at the shared rate give −7.64 [−8.72, −6.55] and −5.89 [−6.89, −4.89] pp, instead of +10.72 and +12.47 against the zero forecast at the same rate.
  • The rerun at 1.0 reproduces the archived zero-forecast endpoints exactly.

So a larger server step alone closes the gap, and the large HT forecast gains on FEMNIST are mostly scale compensation. The rate that passes the participating mean exceeds the one that restores expected mass, so the untuned forecasts and participating mean may also gain from a larger step. The oracle-free Hájek estimator pointed the same way: it puts the zero forecast at 55.65–55.69% at every floor, level with the participating mean (App. G.5). The 10–13% that the oracle factor leaves is therefore not evidence of the relative-weight distortion of Proposition 2, which these runs do not detect (the proposition itself is unchanged); §5.3 now says so, and the abstract reports the result. Only the zero forecast was tuned, so this is not a ranking of tuned methods.

Do these answers address your concerns? If anything remains unclear, please leave a further comment; we will respond during the discussion period.

AuthorsOpenReview ↗

Follow up on our response

Thank you again for your detailed review. We wanted to follow up on our response, particularly the experiments with correlated client availability, the matched cold-start comparison, and the FEMNIST control with a tuned server learning rate. If you have had a chance to look at the revision, we would appreciate hearing whether these additions address your main concerns. We would be happy to clarify any point that remains open.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.