DeNA-SGN swaps learned graph attention for fixed logic: its spiking circuit matched the exported model on every one of 293,577 test queries
Synopsis
Spiking graph learning has mostly relied on learned edge weights or stochastic spike encoding, which makes a complete inference circuit hard to compile and check query by query. DeNA-SGN instead uses two fixed binary neighborhood summaries, OR and a tie-inclusive Majority, feeding a learned signed-weight threshold classifier, with degree-dependent inhibitory control neurons implementing Majority as a shared control circuit compiled into the SuperNeuroMAT simulator. Across seven datasets and 41 mapped classifiers, native simulation matched the exported artificial network on every evaluated class decision and output-spike pattern; on the new confirmatory Amazon Photo set it reaches 83.93±0.93% current-readout accuracy and 71.61±2.22% strict output-spike accuracy.
Interpretation
Spiking graph classification can now run through an inference path built entirely from threshold neurons and fixed logical operators, with no learned edge weights and no stochastic spike encoding in the aggregation. Earlier spiking graph methods either computed a real-valued graph convolution before spike encoding, or treated citation relations as synapses and propagated labels through online spike-timing-dependent plasticity, leaving no fixed circuit to check for the aggregation step itself. Across seven datasets and 41 mapped classifiers, native simulation matched the exported float64 artificial network on every evaluated class decision and output-spike pattern, over 293,577 benchmark test-query executions, with 3,000 additional OR-control queries passing the same decision checks.
Each binary feature keeps only two kinds of neighborhood evidence: whether the feature appears anywhere in the closed neighborhood, and whether at least half the neighbors support it, and both feed a bias-free two-layer threshold classifier. The Majority threshold varies with query-node degree, and is realized with ordinary fixed-threshold neurons plus signed inhibitory currents, so the number of control neurons grows logarithmically with maximum closed-neighborhood size and is independent of feature dimension. The complete circuit is compiled and executed in SuperNeuroMAT 3.5.0 in sparse CPU mode, with query, relay, control, feature, hidden, output and observer neurons all counted, totaling 7,521 neurons and 304,868 synapses on Cora.
On the same Cora public split, seeds and training budget, adding the Majority channel gives a mean gain of 1.63 percentage points in current-readout accuracy over OR alone. OR alone reaches 73.27±0.21% current-readout accuracy and OR+Majority reaches 74.90±2.67%, at the cost of 551 additional deployed neurons, 7.91% more than OR. The comparison is a mean over three initialization seeds, and the authors explicitly do not infer statistical significance; the GAT artificial-network reference reaches 81.47±0.90% on the same split.
Perspective
The demonstrated scope is transductive node classification in SuperNeuroMAT 3.5.0 sparse CPU mode, covering seven datasets (Cora, CiteSeer, PubMed, KarateClub, Wiki-CS, ogbn-arxiv and Amazon Photo), with resource counts including all graph nodes and supporting neurons. Current argmax is decoded on the host, and energy is analytically modeled rather than measured on hardware.
The authors state that three-run standard deviations are descriptive, and that the 20 Wiki-CS masks share test nodes; binarization discards continuous magnitudes, the arxiv graph is symmetrized, and changes to graph or features require recompilation. The energy ratio depends on graph size, activity and comparator, dropping to 0.53 on ogbn-arxiv because clocking the much larger deployed neuron population dominates the modeled cost.
