DeNA-SGN: Deterministic Neighborhood Aggregation for Verified Spiking Graph Inference
Organizations: Oak Ridge National Laboratory, Oak Ridge, TN, USA
Abstract
Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation, supporting control and classification. We propose DeNA-SGN, which combines fixed OR and tie-inclusive Majority aggregation of binary node features with a learned signed-weight threshold classifier. A degree-dependent inhibitory construction implements Majority through shared control neurons, enabling a complete circuit compilation into SuperNeuroMAT. Across seven datasets and 41 mapped classifiers, native simulation matches the exported ANN on every evaluated class decision and output-spike pattern. On the new confirmatory Amazon Photo dataset, DeNA-SGN achieves 83.93+-0.93% current-readout accuracy and 71.61+-2.22% strict output-spike accuracy; Wiki-CS reaches 75.99+-0.54% over 20 public masks. In the matched Cora study, we achieve 74.90+-2.67% on current-readout and 68.57+-2.94% strict spike accuracy. Complete circuit and activity accounting supports analytical energy comparisons and identifies the cost of clocking inactive neurons in large deployments. Current argmax is decoded on the host, and energy is modeled rather than measured on hardware. These findings establish a reproducible route to useful graph classification with decision-matched spiking execution and explicit implementation costs.
Figures & tables
| Graph operator in spikes | Node features | Learning | Time steps | |
|---|---|---|---|---|
| SpikingGCN | no (real conv.) | yes | offline training | (tuned) |
| S-GNN lineage | yes (citation synapses) | no / optional | online STDP, BO | 8, 14, 20 |
| DeNA-SGN | yes (OR and Majority) | yes | offline STE, no STDP | 7 (output at 4) |
| Population | Count | Fires at tick | Role |
|---|---|---|---|
| Query | 0 (input) | receives the single input spike | |
| Relay | 1 | forwards the query to its closed neighborhood | |
| Control | 1 | delivers to Majority neurons if bit of is set | |
| OR feature | 2 | fires if | |
| Majority feature | 2 | fires if | |
| Hidden | 3 | fires if |
| Dataset / source | Nodes | Undirected edges | Node representation before encoding | ||
|---|---|---|---|---|---|
| Cora [ 7 , 9 ] | 2,708 | 5,278 | 1433 | 7 | Binary word presence |
| CiteSeer [ 7 , 9 ] | 3,327 | 4,552 | 3703 | 6 | Binary word presence |
| PubMed [ 8 , 9 ] | 19,717 | 44,324 | 500 | 3 | TF–IDF word attributes |
| KarateClub [ 12 , 28 ] | 34 | 78 | 34 | 4 | Identity features (PyG) |
| Wiki-CS [ 10 ] | 11,701 | 216,123 | 300 | 10 | Article-word embeddings |
| ogbn-arxiv [ 11 , 29 ] | 169,343 | 1,157,799 | 128 | 40 | Title/abstract embeddings |
| Dataset | Protocol | Runs | Train | Validation | Test | Unused |
|---|---|---|---|---|---|---|
| Cora | provided | 3 | 140 | 500 | 1000 | 1068 |
| CiteSeer | provided | 3 | 120 | 500 | 1000 | 1707 |
| PubMed | provided | 3 | 60 | 500 | 1000 | 18157 |
| KarateClub | derived | 3 | 4 | 4 | 26 | 0 |
| Wiki-CS | provided | 20 | 580 | 1769 | 5847 | 3505 |
| ogbn-arxiv | provided | 3 | 90941 | 29799 | 48603 | 0 |
| Dataset | Prot. | Current (%) | Strict (%) | Macro-F1 |
|---|---|---|---|---|
| Cora | P | 74.90 2.67 | 68.57 2.94 | 0.741 |
| CiteSeer | P | 60.93 1.90 | 51.67 5.90 | 0.568 |
| PubMed | P | 72.53 1.12 | 61.97 5.20 | 0.723 |
| KarateClub | D | 62.82 8.88 | 20.51 2.22 | 0.532 |
| Wiki-CS | P | 75.99 0.54 | 67.36 1.66 | 0.721 |
| ogbn-arxiv | P | 58.36 0.69 | 46.20 0.99 | 0.244 |
| Dataset | Runs | Native test queries | Full replays |
|---|---|---|---|
| Cora | 3 | 3,000 | 7 |
| CiteSeer | 3 | 3,000 | 7 |
| PubMed | 3 | 3,000 | 7 |
| KarateClub | 3 | 78 | 5 |
| Wiki-CS | 20 | 116,940 | 7 |
| ogbn-arxiv | 3 | 145,809 | 7 |
| Dataset | Current | Strict | Bin. Argmax | Silent | Multi |
|---|---|---|---|---|---|
| Cora | 74.90 | 68.57 | 73.03 | 10.40 | 6.57 |
| CiteSeer | 60.93 | 51.67 | 57.57 | 17.90 | 11.07 |
| PubMed | 72.53 | 61.97 | 67.80 | 13.73 | 6.57 |
| KarateClub | 62.82 | 20.51 | 57.69 | 76.92 | 0.00 |
| Wiki-CS | 75.99 | 67.36 | 70.44 | 11.80 | 6.07 |
| ogbn-arxiv | 58.36 | 46.20 | 49.89 | 23.46 | 8.96 |
| Model | Current (%) | Strict (%) | Neurons | Synapses |
|---|---|---|---|---|
| OR | 73.27 0.21 | 66.03 5.13 | 6,970 | 209,749 |
| OR+Majority | 74.90 2.67 | 68.57 2.94 | 7,521 | 304,868 |
| GAT | 81.47 0.90 | – | – | – |
| Method | Purely spiking? | Features | Reported evaluation / readout | Cora | CiteSeer | PubMed |
|---|---|---|---|---|---|---|
| Cong et al., ICONS 2022 [ 4 ] | Core only | no | Five random-split runs; Table 2 | 56 | 41 | 54 † |
| Cong et al., ICMLA 2023 (BO) [ 5 ] | Core only | no | Optimized structure-only; Table III | 67 | 51 | 79 |
| Fixed-point Lava [ 6 ] | Core only | no | Cora validation; 140 nodes | 62.86 ‡ | – | – |
| Structure-only SuperNeuroMAT [ 14 ] | Core only | no | Citation example; Section 6.2.2 | 62.7 | 42.4 | 54.7 |
| DeNA-SGN: strict spikes | Yes | yes | Public test; spike decision | 68.57 2.94 | 51.67 5.90 | 61.97 5.20 |
| DeNA-SGN: current readout | Core only | yes | Public test; host argmax | 74.90 2.67 | 60.93 1.90 | 72.53 1.12 |
| Dataset | Neurons | Synapses | Spikes | Events | ANN (nJ) | SNN (nJ) | Ratio |
|---|---|---|---|---|---|---|---|
| Cora | 7,521 | 304,868 | 111.7 | 8,485 | 1321.7 | 12.91 | 102.35 |
| CiteSeer | 11,842 | 715,338 | 141.5 | 12,925 | 3409.6 | 19.94 | 171.03 |
| PubMed | 40,521 | 2,179,908 | 223.7 | 20,232 | 461.6 | 46.60 | 9.91 |
| KarateClub | 214 | 4,461 | 16.7 | 569 | 33.1 | 0.66 | 49.94 |
| Wiki-CS | 24,133 | 4,046,321 | 505.7 | 55,995 | 285.8 | 67.34 | 4.24 |
| ogbn-arxiv | 339,135 | 24,380,220 | 228.0 | 22,546 | 136.9 | 257.71 | 0.53 |
Graph Neural Networks
Graph Neural Networks learn from graph-structured data by combining information across connected nodes, edges, and larger structures.
See all 775 papers