Research map / Graph learning / Graph self-supervised learning
A Graph is Worth 1-bit Spikes: When Graph Contrastive Learning Meets Spiking Neural Networks
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 SpikeGCL, a graph contrastive learning framework that uses spiking neural networks with blockwise surrogate-gradient training to learn 1-bit node representations, reporting near-full-precision accuracy with about 32x storage compression.
Problem
Full-precision graph contrastive learning requires large hidden dimensions, causing high memory, storage and energy costs unsuitable for resource-constrained settings; binarized graph representation learning under self-supervision remains underexplored.
Approach
Partition node features into T non-overlapping groups fed as sequential inputs to shared-parameter peer GCN encoders followed by a spiking neuron that emits binary spikes; train with a margin ranking contrastive loss under blockwise (per-time-step) local learning with stop-gradient to avoid vanishing gradients in SNN surrogate learning.
Key results
- On six large-scale datasets, SpikeGCL reaches 88.9% on Computers and 70.9% on arXiv, comparable to full-precision GCL methods (roughly 95-99% of their performance per the authors).
- Efficiency: about 7x lower theoretical energy and about 1/60 model size on MAG relative to full-precision GCL baselines; representation storage compressed nearly 32x.
- Theorem gives an approximation error bound of order 1/T between SNN firing rates and an oracle GCN, with the authors showing the bound is exponential in depth L.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpikeGCL instead of BGRL | Graph representation learning · Amazon-Computer | accuracy (%): 90.3$_{\pm0.2}$ -> 88.9$_{\pm0.3}$ | Hurt |
| Use SpikeGCL instead of GA-SNN | Graph representation learning · Amazon-Photo | accuracy (%): 93.5$_{\pm0.6}$ -> 93.0$_{\pm0.1}$ | Hurt |
| Use SpikeGCL instead of SpikeNet | Graph representation learning · Coauthor-CS | accuracy (%): 93.4$_{\pm0.2}$ -> 92.8$_{\pm0.1}$ | Hurt |
| Use SpikeGCL instead of SpikeNet | Graph representation learning · Coauthor-Physics | accuracy (%): 95.8$_{\pm0.7}$ -> 95.2$_{\pm0.6}$ | Hurt |
| Use SpikeGCL instead of BGRL | Graph representation learning · ogbn-arXiv | accuracy (%): 71.6$_{\pm0.1}$ -> 70.9$_{\pm0.1}$ | Hurt |
| Use SpikeGCL instead of SUGRL | Graph representation learning · ogbn-MAG | accuracy (%): 32.4$_{\pm0.1}$ -> 32.0$_{\pm0.3}$ | Hurt |
| Use SpikeGCL instead of full-precision GCL baselines | Graph representation learning · Amazon-Computer, Coauthor-CS, Coauthor-Physics, ogbn-arXiv, ogbn-MAG | energy (mJ) | Helped |
Surprising findings
- Binary 1-bit representations do not necessarily cause accuracy loss when properly trained, per the authors.
- Vanishing gradients in SNN surrogate learning are alleviated by horizontal blockwise training along the time dimension.
Limitations
- Linear evaluation assumes downstream tasks are linearly separable, which may not capture all representation quality.
- Theoretical guarantee requires reset-by-subtraction; reset-to-zero incurs a non-vanishing error term.
- Experiments do not include very large graphs (N around 10^6 or more).
Experimental setup
Datasets: Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics Cora CiteSeer PubMed ogbn-arXiv ogbn-MAG
Benchmarks: Amazon-Computers Amazon-Photo Coauthor-CS Coauthor-Physics ogbn-arXiv ogbn-MAG Cora CiteSeer PubMed
Metrics: classification accuracy parameter size (KB) theoretical energy consumption (mJ) training time GPU memory usage
Baselines: GCN GAT SpikeNet SpikingGCN GC-SNN GA-SNN Bi-GCN BinaryGNN BANE DGI GRACE CCA-SSG
Builds on: GRACE GraphCL GCN PLIF BGRL
Closest papers by meaning
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.89)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.88)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.88)
- TopoGCL: Topological Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
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.