acceptodds
Withdrawn from ICLR 2027

OCS-BAHFL: Online Communication Scheduling for Buffered Asynchronous Hierarchical Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Buffered asynchronous hierarchical federated learning (HFL) reduces synchronization overhead, but introduces a fundamental communication-timing problem: aggregating too frequently wastes communication and may rely on weak evidence, whereas waiting too long increases information age and staleness. We propose OCS-BAHFL, an online communication scheduling framework that independently adapts communication timing at the client–edge and edge–cloud levels. OCS-BAHFL controls an edge holding time and a cloud aggregation deadline using only event-local communication statistics, including effective evidence, event frequency, information age, and model disagreement. These observations are converted into self-normalized accumulation and freshness pressures, whose balance defines a strictly convex scalar surrogate with a closed-form waiting-time target. A projected log-domain update tracks this target while preserving positive and bounded communication timers. The scheduler requires no validation accuracy, training-loss feedback, or additional communication messages. We further establish bounded communication-induced delay, characterize controller tracking, and connect effective evidence and bounded hierarchical staleness to convergence under smooth non-convex objectives. Experiments on non-IID MNIST demonstrate that OCS-BAHFL dynamically adjusts communication timing under different local-training budgets and reduces communication volume relative to hierarchical FedAvg and a non-adaptive configuration. In particular, at the largest evaluated local-training budget, OCS-BAHFL reduces communication volume relative to hierarchical FedAvg while maintaining competitive learning performance. These results show that communication timing itself can serve as an effective online control variable in buffered asynchronous HFL.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Learning, Buffered, Asynchronous

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_hKr5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an online communication scheduler for buffered asynchronous hierarchical federated learning. It adjusts edge holding times and cloud aggregation deadlines using effective evidence, information age, event frequency and model disagreement. The two timers track closed-form targets through bounded log-domain updates, without using training-loss or validation feedback.

Critical strengths

The paper makes a useful distinction between waiting for client updates at the edge and waiting for edge updates in the cloud. The scheduler uses information already available at the aggregation nodes, so it does not require training-loss feedback, validation accuracy, or additional model transmissions, and the timer update itself involves only simple scalar calculations.

Critical weaknesses
  1. The rationale for J(x)=A/x+Bx is reasonable, but the paper does not show why this particular rule should improve the actual communication–learning trade-off. Proving that the surrogate is convex does not establish that connection. The relative age term also raises a concern that it equals one when the current and previous ages are the same, regardless of whether the updates are fresh or consistently stale. It is unclear how the controller handles sustained high staleness.

  2. The cloud waiting-time proof does not match the stated aggregation rule. Equation (3) requires a quorum before the adaptive deadline can trigger aggregation; otherwise, the system relies on T(hard). Lemma 4, which instead claims a bound based on the adaptive deadline whenever a dirty edge exists. This needs either a corrected bound or an additional assumption.

  3. The experiments provide limited evidence of the proposed scheduler's benefit. The comparison with Non-Adaptive is useful, but separate edge-only and cloud-only ablations are needed to understand where the gains come from.

Questions for authors

Under the buffer-capacity, quorum and hard-deadline conditions, how often do T(hold) and T(max) actually determine when aggregation occurs? What evidence from trigger statistics and separate edge/cloud ablations shows that the proposed evidence–freshness rule reduces communication or time to a fixed target accuracy, rather than merely changing the timer values?

Reviewer_yDu2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OCS-BAHFL, an online communication scheduling framework for buffered asynchronous hierarchical federated learning. The method independently adapts the edge holding time and cloud aggregation deadline using communication-side information, including effective evidence, event frequency, information age, and model disagreement. These statistics are converted into accumulation and freshness pressures, which define a convex surrogate objective with a closed-form waiting-time target. A projected log-domain controller dynamically adjusts the communication timers while maintaining bounded waiting times. The paper also provides theoretical analysis concerning controller stability, bounded staleness, and non-convex convergence. Experiments on non-IID MNIST demonstrate communication savings compared with fixed scheduling and selected hierarchical FL baselines.

Critical strengths

The paper addresses a meaningful communication scheduling problem in buffered asynchronous hierarchical federated learning. Independently adapting client-edge holding times and edge-cloud aggregation deadlines is a reasonable design for handling heterogeneous communication dynamics.

Critical weaknesses
  1. Inconsistencies in the controller formulation. The main text and Appendix C provide different formulations of the accumulation and freshness pressures. For example, (15) uses a shifted rate ratio, whereas (50) uses a smoothed rate ratio without the same offset. Similar inconsistencies appear in the age normalization terms.
  2. The theoretical analysis does not sufficiently justify the proposed scheduling policy. Although the scalar surrogate is strictly convex, its formulation in J(x)=A/x+Bx is heuristic. The convergence analysis primarily establishes guarantees under bounded delays rather than demonstrating that the proposed timer targets optimize communication efficiency or convergence behavior.
  3. The abstract emphasizes communication reduction at the largest local-training budget, (E=60 in the experiment). However, Table 1 reports communication volume only for E=5, while Figure 2 mainly compares accuracy against training latency. The authors should provide direct communication-volume comparisons for E=20 and E=60 to substantiate this central claim.
  4. The contribution of the two scheduling levels is not sufficiently isolated. The framework combines adaptive timers with staleness-aware aggregation, proximal momentum, and soft hierarchical broadcasting. The paper does not separately evaluate edge-only adaptation and cloud-only adaptation.
  5. Only MNIST on CNN is used, which is too simple. The scale of experiment is too small.
  6. The number of baselines is also too small.
  7. Literature review is missing, and the baseline methods is missing.
  8. There is no way to reproduce this work since the authors did not provide any implementation details and hyperparameter tuning details.
Questions for authors

See weakness

Reviewer_VLPrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes OCS-BAHFL, which adapts edge buffering times and cloud aggregation deadlines in asynchronous hierarchical federated learning. The controller uses effective evidence, information age, event rates, and model disagreement to compute a waiting-time target, followed by a bounded log-domain update. The paper provides timer-tracking and convergence analyses, and evaluates the method on non-IID MNIST against fixed scheduling and hierarchical FL baselines.

Critical strengths
  • Separating edge buffering from cloud aggregation is a useful way to formulate the scheduling problem. The two stages face different arrival patterns and staleness costs, so allowing them to adapt separately is well motivated. The controller also avoids relying on validation accuracy or training-loss feedback.

  • The scheduler requires no validation accuracy, training-loss feedback, or additional communication messages. Once the event statistics are available, each timer update takes constant time and memory. These modest feedback requirements are appealing for asynchronous FL systems, where additional coordination can be costly.

Critical weaknesses
  • The convergence result is not established by the supplied proof. The transition from Eq. (117) to Eq. (118) introduces an extra factor of the learning rate in the heterogeneity and staleness error terms without justifying it. This leaves the claimed convergence rate unsupported. Lemma 4 also overlooks the quorum requirement when deriving the cloud-waiting bound: with too few dirty edges, aggregation may wait until the fixed hard deadline rather than the adaptive deadline. Both issues affect the theoretical guarantee and require substantive revision.

  • Important details are missing for interpreting the existing comparisons. Figure 2(b) includes relevant baselines, but their curves are digitized from DHA-HFL, and the paper does not establish that the timing conditions are comparable. Table 1 provides a useful Non-Adaptive control, but its timer values, tuning procedure, and whether all other components are identical are unclear. These details matter for judging how much of the reported improvement comes from adaptation. No run-to-run uncertainty is reported, and the communication measurements cover only E=5 despite the abstract’s claim about E=60. The differing controller definitions in the main text and Appendix C further obscure which policy produced the results.

Questions for authors
  • What timer values and tuning procedure were used for Non-Adaptive? Were its remaining aggregation and broadcast components identical to OCS-BAHFL?

  • How were the computation and communication timing assumptions matched to the digitized baselines in Figure 2(b)? Are communication results at E=60 and repeated-run statistics available?

  • What cancellation argument or error bound justifies the ηs2\eta_s^2 dependence of the heterogeneity and staleness terms in the convergence analysis?

  • Which controller generated the experiments? Section 4.3 uses previous observations, while Appendix C introduces smoothed baselines and different normalization.

  • Table 1 reports 23.1 s to reach 95% for N=100,M=10,E=5N=100, M=10, E=5, whereas the OCS-BAHFL (E=5E=5) curve in Fig. 2(b) reaches 0.95 only around 100 s. Which is correct, and do they use the same configuration?

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

Related papers

Discussion (0)

Sign in to comment.