TL;DR
DeNA-SGN is a new spiking graph network that replaces complex, learned edge weights with simple, deterministic logical operators like OR and Majority. This approach allows for a complete, verifiable circuit-level implementation on neuromorphic hardware, achieving high accuracy on standard benchmarks while providing a precise accounting of energy and resource costs.
What Problem Does This Paper Address?
Graph Neural Networks (GNNs) have become the standard for learning from structured data like social networks or citation graphs. However, GNNs typically rely on continuous, real-valued computations and complex attention mechanisms. This makes them a poor fit for neuromorphic computing, which uses discrete, event-driven "spikes" to process information efficiently.
While previous research has attempted to bridge this gap, existing "spiking GNNs" often rely on stochastic (random) spike encoding or lack a clear, one-to-one mapping between the mathematical model and the physical circuit. This makes it difficult to verify if the hardware is actually doing what the model intended and even harder to accurately measure the true energy cost of the implementation.
Methods
The authors introduce DeNA-SGN (Deterministic Neighborhood Aggregation for Spiking Graph Inference). The core idea is to simplify the way a node gathers information from its neighbors.
Instead of learning how much to "attend" to each neighbor, DeNA-SGN uses two fixed logical channels:
- OR: Fires if any node in the neighborhood has a specific binary feature.
- Majority: Fires if at least half of the nodes in the neighborhood have that feature.
To implement the "Majority" rule in a spiking circuit, the authors developed a clever degree-adaptive inhibitory construction. Since different nodes have different numbers of neighbors, the threshold for "half" changes. They use shared "control neurons" that send inhibitory signals to adjust the firing threshold of the Majority neurons based on the node's degree.
The resulting binary signals are fed into a compact, learned threshold classifier. The entire system is trained offline using a "straight-through estimator" to handle the non-differentiable nature of thresholds and then compiled into an explicit spiking circuit for the SuperNeuroMAT simulator.
Datasets
The model was evaluated on several standard citation and social network benchmarks:
- Cora, CiteSeer, PubMed: Classic citation networks.
- Wiki-CS: A Wikipedia-based graph.
- ogbn-arxiv: A large-scale publication dataset.
- Zachary’s Karate Club: A small, well-known social network.
- Amazon Photo: A new confirmatory dataset introduced in this paper to verify the model's performance on unseen data after the architecture was finalized.
Key Insights and Ideas
- Determinism is Key for Verification: By using fixed logical operators and avoiding stochastic encoding, the authors achieved a 100% match between their mathematical model and the simulated spiking circuit across nearly 300,000 test queries.
- Majority Rules: Adding the "Majority" channel provides a significant boost over just using "OR" because it helps the network distinguish between a feature that is rare in a neighborhood versus one that is widespread.
- The Energy "Crossover" Point: The paper provides a detailed energy analysis. While spiking networks are often touted for their efficiency, the authors found a "crossover" point. For very large graphs (like ogbn-arxiv), the background cost of simply "clocking" millions of inactive neurons can actually make the SNN less efficient than a traditional dense ANN. This highlights a critical deployment challenge for large-scale neuromorphic systems.
Key Results
DeNA-SGN demonstrates competitive predictive performance while maintaining strict circuit-level fidelity:
- Amazon Photo: Achieved 83.93% accuracy using current-readout and 71.61% "strict" spike accuracy (where only the correct class neuron is allowed to fire).
- Wiki-CS: Reached 75.99% accuracy across 20 different data splits.
- Efficiency: On the CiteSeer dataset, the modeled SNN was estimated to be 171x more energy-efficient than a comparable dense ANN, provided the hardware can effectively manage inactive neurons.
Limitations and Future Work
- Binarization Loss: The model requires binary features. Continuous data must be binarized offline, which inevitably discards some information.
- Static Compilation: The current design requires the circuit to be recompiled if the graph structure or features change significantly.
- Analytical vs. Physical Energy: The energy results are based on analytical models and simulations, not direct measurements on physical neuromorphic chips like Loihi.
- Hardware Constraints: Future work needs to address physical hardware limitations, such as maximum fan-in (how many connections a single neuron can receive) and complex routing across large chips.
Summary updated