Research map / Graph learning / Knowledge and temporal graphs
Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GRSNN, a spiking neural network that encodes relations via learnable synaptic delays and temporal coding to perform knowledge graph and homogeneous graph link prediction as a neural generalization of path formulations.
Problem
Whether spiking neural networks can perform human-like graph reasoning (link prediction in knowledge graphs and graphs) by exploiting spiking time and synaptic delays, which prior SNN work has not systematically done at the network level.
Approach
Use learnable synaptic delays (conditioned on edge and query relations) as an extra temporal dimension to encode relation/edge properties, with spike-time decoding, so the SNN propagates all paths from a source node simultaneously as a neural generalization of path formulations and Dijkstra's algorithm.
Key results
- On FB15k-237, the synaptic delay variant reaches MRR 0.368 and Hits@10 0.551, versus 0.311 MRR for the synaptic-weight variant and 0.204 MRR for no relation encoding.
- Theoretical estimate indicates roughly 20x energy reduction versus the non-spiking counterpart on FB15k-237, with spike rate about 0.258.
- GRSNN achieves competitive results, surpassing most machine learning methods across metrics, though NBFNet performs better due to GNN tricks omitted from GRSNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add synaptic-delay relation encoding instead of no relation encoding | SNN · FB15k-237 | MR: 396 -> 139 | Helped |
| Add synaptic-delay relation encoding instead of no relation encoding | SNN · FB15k-237 | MRR: 0.204 -> 0.368 | Helped |
| Add synaptic-delay relation encoding instead of no relation encoding | SNN · FB15k-237 | HITS@1: 0.119 -> 0.275 | Helped |
| Add synaptic-delay relation encoding instead of no relation encoding | SNN · FB15k-237 | HITS@3: 0.226 -> 0.407 | Helped |
| Add synaptic-delay relation encoding instead of no relation encoding | SNN · FB15k-237 | HITS@10: 0.380 -> 0.551 | Helped |
| Use synaptic delays instead of synaptic weights to encode relations | SNN · FB15k-237 | MR: 197 -> 139 | Helped |
| Use synaptic delays instead of synaptic weights to encode relations | SNN · FB15k-237 | MRR: 0.311 -> 0.368 | Helped |
| Use synaptic delays instead of synaptic weights to encode relations | SNN · FB15k-237 | HITS@1: 0.220 -> 0.275 | Helped |
| Use synaptic delays instead of synaptic weights to encode relations | SNN · FB15k-237 | HITS@3: 0.343 -> 0.407 | Helped |
| Use synaptic delays instead of synaptic weights to encode relations | SNN · FB15k-237 | HITS@10: 0.491 -> 0.551 | Helped |
Surprising findings
- NBFNet outperforms GRSNN because GRSNN omits many GNN tricks deliberately to preserve SNN properties.
Limitations
- Simulation limited to T=10 discrete time steps due to BPTT cost, possibly losing precise temporal information
- Omits advanced GNN components, creating a performance gap versus NBFNet
- Energy results are theoretical estimates and exclude memory access and delay-implementation costs
Experimental setup
Datasets: FB15k-237 WN18RR FB15k-237 inductive splits v1-v4 WN18RR inductive splits v1-v4 Cora CiteSeer PubMed
Benchmarks: FB15k-237 WN18RR FB15k-237 inductive v1-v4 WN18RR inductive v1-v4 Cora CiteSeer PubMed
Metrics: MR MRR Hits@1 Hits@3 Hits@10 AUROC AP spike rate energy estimate
Baselines: NBFNet DistMult RotatE CompGCN R-GCN GraIL Synaptic weight SNN variant No-relation SNN variant
Builds on: NBFNet Dijkstra's algorithm surrogate gradient training of SNNs DistMult
Closest papers by meaning
- Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph (ICML 2025 · similarity 0.90)
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.88)
- Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.88)
- Dynamic Reactive Spiking Graph Neural Network (AAAI 2024 · similarity 0.88)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.88)
- Efficient and Effective Time-Series Forecasting with Spiking Neural Networks (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.