Brain-inspired Graph Spiking Neural Networks for Commonsense Knowledge Representation and Reasoning

Hongjian Fang, Yi Zeng, Jianbo Tang, Yuwei Wang, Yao Liang, Xin LiuarXiv 2022 · DOI · arXiv

Abstract

This paper introduces the Knowledge Representation and Reasoning - Graph Spiking Neural Network (KRR-GSNN), a novel architecture designed to bridge the gap between symbolic knowledge graphs and biologically plausible neuromorphic computing. By integrating population encoding, Spike-Timing-Dependent Plasticity (STDP), and reward-modulated learning, the model represents commonsense knowledge as dynamic neural assemblies. Experimental results using the ConceptNet knowledge base demonstrate that KRR-GSNN achieves faster convergence than traditional Graph Convolutional Networks (GCNs) and naturally aligns with the 15% inhibitory neuron ratio observed in the human hippocampus, offering a more interpretable and biologically grounded approach to artificial intelligence.

Introduction: The Commonsense Bottleneck

Artificial Intelligence has long struggled with "commonsense"—the intuitive understanding of the world that humans use to navigate daily life. While Large Language Models (LLMs) and traditional Artificial Neural Networks (ANNs) can simulate reasoning, they typically rely on high-dimensional, fixed-length vectors that function as "black boxes." These representations lack the transparent, sparse, and temporal nature of biological neural processing.

In the human brain, knowledge is not a static list of numbers but a dynamic web of firing neurons. To replicate this, researchers have developed the KRR-GSNN, utilizing Spiking Neural Networks (SNNs)—the "third generation" of neural networks—which communicate via discrete spikes over time. The goal is to create a system where logical reasoning, such as transitivity, emerges from biological constraints rather than just statistical correlation.

Figure 1: Conceptual framework linking brain semantic maps, knowledge graphs, and graph spiking neural networks.

Methods: The KRR-GSNN Architecture

The KRR-GSNN model integrates several key biological principles to represent the ConceptNet knowledge base, which contains millions of "Subject-Relation-Object" triples.

Population Encoding

Moving away from single-vector representations, the model adopts a population-based format inspired by the "Grandmother-Cell" hypothesis. Each concept (e.g., "Apple") is represented by a specific population of Leaky Integrate-and-Fire (LIF) neurons. This allows the network to represent symbols through collective neural activity rather than isolated mathematical values.

Synaptic Plasticity and Learning

The model employs Spike-Timing-Dependent Plasticity (STDP) to learn the edges of the knowledge graph. Connections between neuron populations are strengthened or weakened based on the precise timing of their spikes, adhering to the principle that "neurons that fire together, wire together." To facilitate logical reasoning, a reinforcement learning layer (R-STDP) was added. When the network correctly identifies a logical relationship, a simulated "dopamine" signal reinforces the synaptic changes, guiding the model toward correct reasoning paths.

Figure 2: Mapping knowledge triples to neuron population circuits and the resulting weight distributions and spike trains.

Key Insights: The 15% Solution

A pivotal discovery in this research was the optimal ratio of inhibitory to excitatory neurons. By varying this ratio, researchers found that the KRR-GSNN performed most efficiently when exactly 15% of the neurons were inhibitory. This finding is remarkably consistent with the 14-15% inhibitory neuron ratio observed in the CA1 area of the human hippocampus—the region primarily responsible for memory consolidation and concept learning. This biological alignment suggests that the model’s architecture captures fundamental constraints of efficient information processing in the brain.

Results and Performance

The KRR-GSNN demonstrated several advantages over traditional deep learning methods:

  • Faster Convergence: The spiking architecture reached high performance levels significantly faster than standard Graph Convolutional Networks (GCNs) during the training phase.
  • Traceable Reasoning: Unlike the opaque nature of standard ANNs, the reasoning paths in KRR-GSNN are observable. By monitoring the sequence of firing populations, researchers can trace the logic the model uses to reach a conclusion, moving toward "white-box" AI.
  • Self-Inductive Learning: The network showed an emergent ability to generalize. For instance, after being exposed to specific instances of political leadership (e.g., "Biden is the president of America"), it could autonomously induce the abstract relationship: "A person is the president of a country."

Discussion and Future Directions

While KRR-GSNN represents a significant step toward biologically plausible AI, limitations remain. Currently, the model is purely symbolic and lacks "embodied intelligence"—the ability to learn from multimodal sensory inputs like vision or sound. Furthermore, while the model uses population indexing, it does not yet incorporate the 3D spatial mapping (semantic maps) found in the human cortex.

Future work will focus on integrating these spiking networks with spatially accurate models of the brain, potentially leading to AI systems that are as energy-efficient and adaptable as the biological systems that inspired them.

Summary updated

BibTeX
@misc{fang2022brain,
  author = {Hongjian Fang and Yi Zeng and Jianbo Tang and Yuwei Wang and Yao Liang and Xin Liu},
  title = {{Brain-inspired Graph Spiking Neural Networks for Commonsense Knowledge Representation and Reasoning}},
  year = {2022},
  eprint = {2207.05561},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2207.05561},
}
Figure