DeNA-SGN: Deterministic Neighborhood Aggregation for Verified Spiking Graph Inference

arXiv:2611.14775 · cs.NE · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "DeNA-SGN: Deterministic Neighborhood Aggregation for Verified Spiking Graph Inference".

Jane: Here's the gist of the paper. Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation, supporting control and classification Page 1 Pruning redundant reasoning tokens DeNA-SGN proposes a spiking graph network that replaces learned neighborhood weighting with fixed logical operators, specifically fixed OR and tie-inclusive Majority aggregation combined with a learned signed-weight threshold classifier Page 1 This design is deterministic because the aggregation is a fixed function of the graph and its features, and spike generation involves no stochastic encoding Page 2 The core components include an OR neuron indicating if any node in the closed neighborhood has a feature, and a Majority neuron indicating whether at least half of them do Page 2 How the training signal is built The training objective for DeNA-SGN involves minimizing a loss function that includes classification error, margin terms to encourage separation around the initial unit firing threshold, and an embedding regularizer Page 4 The embedding regularizer penalizes positive similarity between nodes with the same label and discourages constant hidden activity Page 4 The training process uses hard thresholds in the forward pass, employing a straight-through estimator (STE) for gradients to maintain a logical threshold-circuit construction Page 3 Where the gains come from The primary contributions of DeNA-SGN are centered on deterministic aggregation and verification Page 2 This includes a degree-adaptive Majority threshold implemented by at most floor(log2 ∆) shared control neurons with binary-weighted inhibitory synapses Page 2 The work achieves useful predictive performance, reaching 83.93±0.93% current-readout accuracy on the new Amazon Photo confirmatory dataset Page 8 Furthermore, the matched Cora study shows a mean gain of 1.

Paper summary: Tom: So wrapping up DeNA-SGN, we’ve looked at how they replace learned neighborhood weights with fixed logical operators like fixed OR and Majority aggregation combined with a learned signed-weight threshold classifier on neuromorphic substrates. Jane, what’s the bigger picture here for us listeners who just want to understand what this means?

Jane: It means that graph learning on neuromorphic hardware needs an explicit spiking realization of neighborhood aggregation, control, and classification. DeNA-SGN provides a circuit design that does all three together deterministically. The implication is a move toward verifiable AI on these types of chips where the inference circuit matches the learned network structure exactly.

Lu: The paper’s focus is really on achieving this deterministic nature while still being able to perform useful tasks like classification, which they achieve by using hard thresholds in the forward pass and a straight-through estimator for gradients.

Meng: For production folks, it suggests that if you can fix the aggregation mechanism to be deterministic, you remove a huge layer of uncertainty when mapping the model onto hardware. But we still have that question about whether this fixed logic is flexible enough for all kinds of complex problems outside of what they tested on page two <ref:2611.14775#pg2>.

Tom: I agree with Meng on flexibility. The work shows a solid result with eighty-three point nine three plus or minus zero point nine three percent accuracy on the Amazon Photo confirmatory dataset, and the matched Cora study showed a mean gain of one point five. That’s concrete performance data to back up their claim about usefulness.

Jane: And they also pointed out some limitations themselves, which is important. They note that offline feature encoding and host current decoding remain outside this circuit design, meaning you still have those external steps to worry about when you build the whole system.

Lalam: So the core contribution is taking a complex aggregation step and making it a fixed, deterministic function of the graph features, which simplifies building the physical circuit structure for inference.

Tom: That’s it. The DeNA-SGN paper gives us a path toward explicitly realizing neighborhood aggregation in spiking neural networks in a way that can be verified through simulation.

Conclusion: Tom: So we're wrapping up this look at DeNA-SGN, which is about making graph learning on neuromorphic chips deterministic by using fixed logic for neighborhood aggregation and a learned threshold classifier on page one.

Jane: So what does that actually mean for us when we think about putting these kinds of models onto hardware? Is it just cleaner code, or something bigger?

Lu: It means you can replace those complex, messy learned weighting schemes with fixed operations like a simple OR function and a majority vote on the neighborhood.

Meng: From my side at the startup, that deterministic aggregation is huge because it makes verifying the inference circuit much simpler. If you know exactly how that part works before you even run it, you reduce deployment risk significantly.

Tom: Right, so we're talking about moving away from stochastic elements in the core neighborhood decision process and replacing them with something fixed, which helps with that verification aspect.

Jane: And they’ve actually shown this can still give pretty good performance on a real dataset, hitting eighty-three point nine three percent accuracy on Amazon Photo data. That's a concrete number to anchor all this up.

Lu: Yeah, but the work also has to handle the training side too; they use hard thresholds and a straight-through estimator for gradients so they can still learn something useful while keeping that logical circuit structure intact on page three.

Meng: The caveat is always about scale, though. I wonder how this fixed logic holds up when you throw in really complex, non-local dependencies that aren't captured by those simple OR and majority rules.

Tom: That’s the next big question for DeNA-SGN—does this deterministic approach limit the kind of complex graphs we can actually model?

Jane: It definitely sets a new standard for how we build these spiking networks, pushing us toward systems where the inference path is as clear and verifiable as possible.

Oak Ridge National Laboratory

cs.NE

Submitted: 2026-11-10

Updated: 2026-11-17

Code: https://github.com/ORNL/superneuromat

Project page: https://ornl.github.io/superneuromat

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 83/100

The gist: Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation…

Key concepts

Spiking Graph Network
This is a type of neural network designed for neuromorphic hardware that uses discrete 'spikes' instead of continuous values. It models data as a graph where nodes represent entities and connections represent relationships, allowing the network to process information in a biologically inspired, event-driven manner.
Fixed Logical Operators
Instead of learning complex weights for neighborhood aggregation, DeNA-SGN uses pre-defined logical rules like fixed OR and tie-inclusive Majority. This makes the aggregation process entirely deterministic; given the same graph structure and features, the output is always the same, which is crucial for reliable spiking implementations.
Straight-Through Estimator (STE)
Since training involves hard thresholds in a logical circuit construction, standard backpropagation cannot be directly applied. The STE is a technique used to approximate the gradient of this non-differentiable threshold function, allowing the network to be trained effectively while maintaining the logical structure of the circuit.

Terminology

Summary

Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation, supporting control and classification Page 1

Pruning redundant reasoning tokens

DeNA-SGN proposes a spiking graph network that replaces learned neighborhood weighting with fixed logical operators, specifically fixed OR and tie-inclusive Majority aggregation combined with a learned signed-weight threshold classifier Page 1 This design is deterministic because the aggregation is a fixed function of the graph and its features, and spike generation involves no stochastic encoding Page 2 The core components include an OR neuron indicating if any node in the closed neighborhood has a feature, and a Majority neuron indicating whether at least half of them do Page 2

How the training signal is built

The training objective for DeNA-SGN involves minimizing a loss function that includes classification error, margin terms to encourage separation around the initial unit firing threshold, and an embedding regularizer Page 4 The embedding regularizer penalizes positive similarity between nodes with the same label and discourages constant hidden activity Page 4 The training process uses hard thresholds in the forward pass, employing a straight-through estimator (STE) for gradients to maintain a logical threshold-circuit construction Page 3

Where the gains come from

The primary contributions of DeNA-SGN are centered on deterministic aggregation and verification Page 2 This includes a degree-adaptive Majority threshold implemented by at most floor(log2 ∆) shared control neurons with binary-weighted inhibitory synapses Page 2 The work achieves useful predictive performance, reaching 83.93±0.93% current-readout accuracy on the new Amazon Photo confirmatory dataset Page 8 Furthermore, the matched Cora study shows a mean gain of 1. p. 0

Where the paper says it

On the new confirmatory Amazon Photo dataset, DeNA-SGN achieves 83. 93% current-readout accuracy and 71.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Deterministic Graph Aggregation Implementation: The DeNASGN framework replaces learned neighborhood weighting with fixed logical operators, specifically OR and Majority aggregation, making the aggregation a fixed function of the graph and its features, with no learned edge weights. This allows for a complete circuit compilation into SuperNeuroMAT.

  2. Explicit Circuit Verification: The system enables complete circuit accounting and allows for verification where exported-model decisions can be checked against native simulation, ensuring that every supporting neuron and synapse is counted in the analysis of analytical energy comparisons.

  3. Decision-Matched Execution: The proposed system achieves Decision-matched execution, meaning that across seven datasets and 41 mapped classifiers, the native simulation matches the exported ANN on every evaluated class decision and output spike pattern, providing a reproducible route to useful graph classification with decision-matched spiking execution.

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.

Sources

Related papers