Research map / Graph learning / Knowledge and temporal graphs
Dynamic Reactive Spiking Graph Neural Network
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 a Dynamic Reactive Spiking Graph Neural Network whose LIF neurons each have an optimizable threshold, plus learnable random-walk positional spikes, for energy-efficient node classification on graphs.
Problem
Spiking GNNs use same-reactive spiking neurons (shared thresholds per layer) after few-step feature propagation, which limits expressiveness and fails to capture long-range neighbor information for node classification.
Approach
Replace layer-level or fixed thresholds with neuron-level unique optimizable thresholds (dynamic reactive neurons) and integrate learnable graph positional spikes (Laplacian or random-walk encodings) into the spiking inputs or outputs, so long-range structure is captured while staying spike-based.
Key results
- Ours achieves 82.50% on Cora and 92.80% on Amazon Photos in Table 1, outperforming SpikingGCN (77.72% and 87.01%) and most ANN GNNs.
- Replacing neuron-level reactive mechanism with layer-level one ('Ours r/w LLRM') drops Cora accuracy to 77.79%, below the full model's 82.50%.
- Energy consumption is claimed to be in the same order as SpikingGCN and at least 10x lower than traditional GNNs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable positional spike integration (ablation) | DRSGNN · Cora | accuracy: 81.93±0.54 -> 82.50±0.51 | Helped |
| add learnable positional spike integration (ablation) | DRSGNN · Amazon Photos | accuracy: 92.02±0.44 -> 92.80±0.55 | Helped |
| add learnable positional spike integration (ablation) | DRSGNN · ogbn-arxiv | accuracy: 64.06±0.35 -> 66.26±0.39 | Helped |
| add dynamic reactive mechanism (ablation) | DRSGNN · Cora | accuracy: 80.62±0.37 -> 82.50±0.51 | Helped |
| add dynamic reactive mechanism (ablation) | DRSGNN · Amazon Photos | accuracy: 91.82±0.36 -> 92.80±0.55 | Helped |
| add dynamic reactive mechanism (ablation) | DRSGNN · ogbn-arxiv | accuracy: 64.33±0.29 -> 66.26±0.39 | Helped |
| use neuron-level rather than layer-level reactive thresholds (ablation) | DRSGNN · Cora | accuracy: 77.79±0.49 -> 82.50±0.51 | Helped |
| use neuron-level rather than layer-level reactive thresholds (ablation) | DRSGNN · ACM | accuracy: 91.13±0.43 -> 92.73±0.39 | Helped |
| use neuron-level rather than layer-level reactive thresholds (ablation) | DRSGNN · ogbn-arxiv | accuracy: 63.98±0.25 -> 66.26±0.39 | Helped |
| use learnable rather than unlearnable Laplacian positional signals (ablation) | DRSGNN · Cora | accuracy: 69.62±0.43 -> 82.62±0.32 | Helped |
| use learnable rather than unlearnable random-walk positional signals (ablation) | DRSGNN · Cora | accuracy: 70.06±0.37 -> 82.50±0.51 | Helped |
| initialize positional features with Laplacian eigenvectors rather than random walk, using integration way 1 (ablation) | DRSGNN · Cora | accuracy: 82.62±0.32 vs 82.50±0.51 | No clear effect |
Surprising findings
- On Coauthor CS, GEM reports 82.55% while the proposed model reports 92.85%, but the layer-level variant and some ANN baselines are competitive on several datasets, and the method is slightly below GAT/GEM on ogbn-arxiv-like settings in some cases.
Limitations
- Energy claims are estimated rather than measured on neuromorphic hardware.
- Performance is sometimes below ANN GNNs (e.g., ogbn-arxiv, Coauthor CS) and the proposed model is not always best.
- Gradient uses a pseudo-derivative surrogate for the non-differentiable spike function.
Experimental setup
Datasets: Cora Citeseer Pubmed ogbn-arxiv Amazon Photos Amazon Computers ACM DBLP Coauthor CS Coauthor Physics Flickr BlogCatalog
Benchmarks: Cora Citeseer Pubmed ogbn-arxiv Amazon Photos Amazon Computers ACM DBLP Coauthor CS Coauthor Physics Flickr BlogCatalog
Metrics: classification accuracy area under learning curve energy consumption
Baselines: GCN GAT SGC FastGCN DAGNN SDGNN GEM SpikingGCN
Builds on: SGC LIF neuron model SpikingGCN DAGNN RWPE
Closest papers by meaning
- Dynamic Spiking Graph Neural Networks (AAAI 2024 · similarity 0.91)
- Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph (ICML 2025 · similarity 0.90)
- Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning (ICML 2024 · similarity 0.88)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.87)
- Self-Supervised Discovery of Neural Circuits in Spatially Patterned Neural Responses with Graph Neural Networks (NeurIPS 2025 · similarity 0.87)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.85)
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.