acceptodds
Withdrawn from ICLR 2027

Temporal Evidence Beyond Magnitude for Communication-Efficient Federated Spiking Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning with spiking neural networks (SNNs) offers a promising route to energy-efficient edge intelligence, but communication remains a primary bottleneck. Existing compressors largely inherit an artificial neural network (ANN)-oriented view, ranking updates for transmission by their accumulated magnitude. In SNNs, however, each synaptic update accumulates learning contributions across timesteps under stateful leaky integrate-and-fire (LIF) dynamics, so similar update magnitudes can correspond to markedly different temporal contribution patterns. Magnitude-only ranking cannot distinguish these patterns. We introduce LIF temporal evidence, derived from the alignment between timestep-wise learning signals obtained from a single diagnostic replay and the accumulated local update. Building on this evidence, we propose FedBLADE (Block-wise LIF-Aware Dynamic Encoding), which aggregates temporal evidence within architecture- and magnitude-matched blocks, combines it with block update magnitude to prioritize blocks, and admits the highest-ranked blocks under an explicit bit budget, followed by within-block Top- sparsification, low-bit quantization, and compact relative-position coding. Experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 under Dirichlet non-IID partitions and uplink budgets of 1–10% of dense transmission show that FedBLADE outperforms representative compression baselines across the tested settings; it also outperforms FedSNN on CIFAR-10 and Fashion-MNIST and achieves competitive performance against several representative federated SNN methods.

Primary area:
optimization
Keywords:
Federated Learning, Spiking Neural Networks, Communication Compression

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_brN2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduce LIF temporal evidence, derived from the alignment between timestep-wise learning signals obtained from a single diagnostic replay and the accumulated local update. Building on this evidence, the authors propose FedBLADE, which aggregates temporal evidence within architecture- and magnitude-matched blocks, combines it with block update magnitude to prioritize blocks, and admits the highest-ranked blocks under an explicit bit budget, followed by within-block Top-k sparsification, low-bit quantization, and compact relative-position coding. Experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 under Dirichlet non-IID partitions and uplink budgets of 1-10% of dense transmission show that FedBLADE outperforms representative compression baselines across the tested settings; it also outperforms FedSNN on CIFAR-10 and Fashion-MNIST and achieves competitive performance against several representative federated SNN methods.

Critical strengths
  1. The paper conducts ablation studies to evaluate several design choices, including temporal prioritisation, quantisation, and downlink error feedback.

  2. The experimental settings consider multiple Dirichlet concentration parameters.

Critical weaknesses
  1. The main contribution is the use of temporal evidence to prioritise blocks for communication. However, the ablation results suggest that the performance benefit specifically attributable to this mechanism is relatively modest.

  2. The diagnostic is computed before local optimisation, while the communicated update is produced after five local epochs. Under non-IID data, local training can substantially change gradient directions, parameter magnitudes, and the relative importance of different blocks. This therefore does not convincingly demonstrate that a one-shot, pre-training diagnostic remains predictive of block importance after local optimisation.

Questions for authors
  1. What is the correlation between the pre-training diagnostic score and the actual importance of blocks in the final local update? Does this correlation persist after five local epochs and under strongly non-IID data?

  2. Why does removing downlink error feedback cause a much larger accuracy drop than removing temporal prioritisation? What evidence shows that temporal evidence provides a meaningful benefit beyond the existing compression pipeline?

  3. Why should anti-aligned diagnostic signals contribute positively to block importance? What happens if the absolute value in Eq. 5 is removed?

  4. How much additional computation and memory does the diagnostic replay require, and how does this overhead affect the end-to-end efficiency?

  5. Can the authors provide comparisons with identical communication budgets, client participation, quantisation, and error-feedback mechanisms?

  6. Are the improvements over UL-Shuffle and Plain-RMS statistically significant across independent runs?

Reviewer_77SnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedBLADE for communication compression in federated spiking neural networks. It combines update magnitude with timestep-wise learning signals to prioritize parameter blocks, followed by within-block sparsification, quantization, and support encoding under a coded-bit budget, together with downlink error feedback. The method is evaluated across multiple datasets, levels of data heterogeneity, and communication budgets.

Critical strengths
  1. The method is motivated by SNN temporal dynamics. Using timestep-wise learning signals to supplement update magnitude provides a relevant approach to update selection under communication constraints.
  2. The communication accounting and experimental coverage are useful strengths. The budget includes values, support information, quantization scales, and other message fields. Experiments across communication budgets, data heterogeneity, and multiple seeds help assess compression effectiveness.
Critical weaknesses

The independent contribution of temporal evidence remains unclear. Plain-RMS removes both temporal modulation and layer normalization, so its performance gap from the full method cannot be attributed entirely to temporal evidence. A more targeted ablation would strengthen this claim.

Questions for authors

Could the authors disable only temporal modulation while retaining layer normalization and all other components, and compare with the full method to clarify the benefit of temporal evidence itself?

Reviewer_vSASOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedBLADE, which combines temporal block prioritization, sparse encoding, and downlink error feedback for communication-efficient federated SNN training.Experiments on three image-classification datasets evaluate its performance under non-IID data and limited communication budgets.

Critical strengths
  1. The temporal-evidence score can distinguish blocks with identical final updates and identical time-summed diagnostic gradients by using their timestep-wise contributions. This gives the prioritization rule access to information beyond conventional update-magnitude statistics.

  2. The diagnostic replay is separated from local optimization: it takes no optimizer step, restores normalization buffers and supported random states, and uses the resulting signals only for transmission ordering. This preserves the local training objective and provides a modular way to incorporate SNN-derived information into an update compressor.

Critical weaknesses
  1. Equation (6) ranks blocks without considering their encoding costs, although Appendix B shows that these costs depend on block size, retained support, and whether an additional layer scale must be transmitted. Consequently, a higher score does not necessarily imply a larger benefit per transmitted bit, and satisfying the budget constraint alone does not establish efficient use of that budget. A comparison against cost-aware ordering or reconstruction-error-based selection using the same codec would clarify whether the proposed allocation rule provides an advantage over simpler alternatives.

  2. The evaluation uses three static-image datasets with four or six timesteps, but the timestep count changes together with the dataset and architecture. There is no controlled experiment varying the timestep count or LIF temporal decay while keeping the task and architecture fixed. It therefore remains unclear when temporal evidence becomes useful, whether its advantage increases with temporal dependence, and when magnitude-based ranking is sufficient. Such experiments would provide a more direct test of the central motivation than comparisons across different dataset–architecture pairs.

  3. Table 2 evaluates each dataset at only one communication budget, while Figure 3 varies the budget only for FedBLADE. These results do not show whether the relative advantage persists across budgets or how much cumulative communication each method needs to reach the same accuracy. Comparing accuracy–budget curves and cumulative communicated bits to a shared accuracy target would better establish the claimed communication efficiency.

Questions for authors

Please check the weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.