- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a general framework for integrating spiking neural networks (SNNs) into message-passing neural networks (MPNNs) and graph Transformers. For MPNNs, the authors reposition linear transformations and introduce spiking neurons to enable sparse accumulation operations. They also provide a theoretical analysis suggesting that spike thresholding can mitigate over-smoothing. For graph Transformers, the paper introduces spike-based attention using binary queries and keys, replaces softmax with layer normalization, and supports different attention structures.
Experiments across nine graph architectures and six datasets demonstrate competitive predictive performance and substantially lower estimated computational energy, with spiking models consuming approximately 6–19% of their non-spiking counterparts' estimated energy.
Critical strengths
Broad architectural applicability. The framework is evaluated on six MPNNs and three graph Transformers, demonstrating that the proposed spiking integration strategy can accommodate diverse graph learning architectures.
Encouraging empirical performance. Table 1 reports competitive results across six datasets, with improvements in most spiking versus non-spiking comparisons. The attention ablation and over-smoothing visualizations provide additional evidence supporting the proposed design.
Critical weaknesses
Energy efficiency is not validated through actual hardware measurements. The reported energy reductions rely on theoretical MAC/AC operation counts (Section 4.2 and Figure 4), without fully accounting for memory access, spike communication, normalization, and temporal execution overhead. Therefore, the claimed practical energy-efficiency gains remain uncertain.
The theoretical analysis does not establish a general guarantee against over-smoothing. Theorem 1 characterizes the local sensitivity of thresholded spike outputs under distributional assumptions, rather than the propagation of node representations across network depth. Furthermore, the proofs in Appendix C establish asymptotic upper bounds, which alone do not guarantee the claimed separation between spiking and continuous activations. Thus, the theoretical result provides useful intuition but does not rigorously establish that the proposed architecture prevents over-smoothing.
Questions for authors
Can the authors clarify how memory access, spike communication, layer normalization, and temporal execution costs affect the estimated energy savings? Are measured latency or energy results available?
How does the local Dirichlet-energy analysis in Theorem 1 extend to the evolution of node representations across multiple graph layers? Can the authors clarify the assumptions needed for the claimed over-smoothing mitigation?
Could the authors provide an ablation separating the effects of spiking neurons, linear-layer repositioning, and layer normalization, to better identify their individual contributions?