- 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
-
The paper conducts ablation studies to evaluate several design choices, including temporal prioritisation, quantisation, and downlink error feedback.
-
The experimental settings consider multiple Dirichlet concentration parameters.
Critical weaknesses
-
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.
-
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
-
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?
-
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?
-
Why should anti-aligned diagnostic signals contribute positively to block importance? What happens if the absolute value in Eq. 5 is removed?
-
How much additional computation and memory does the diagnostic replay require, and how does this overhead affect the end-to-end efficiency?
-
Can the authors provide comparisons with identical communication budgets, client participation, quantisation, and error-feedback mechanisms?
-
Are the improvements over UL-Shuffle and Plain-RMS statistically significant across independent runs?