SSEL: spike-based structural entropic learning for spiking graph neural networks.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Graph Neural Networks
• Information Theory

Background:

• Spiking Neural Networks (SNNs) represent an energy-efficient, event-driven AI paradigm.
• Spiking Graph Neural Networks (SGNNs) extend SNNs to graph data but are vulnerable to adversarial topology perturbations.
• Existing SGNNs lack robustness against structural manipulation, limiting their practical application.

Purpose of the Study:

• To develop a robust SGNN framework resilient to adversarial topological attacks.
• To enhance the energy efficiency of SGNNs through principled information-theoretic approaches.
• To introduce structural entropy minimization as a core learning objective for SGNNs.

Main Methods:

• Introduced the Spike-based Structural Entropy Learning (SSEL) framework.
• Integrated structural entropy theory into SGNN learning objectives to guide topology refinement.
• Developed an entropy-driven topological gating mechanism to restrict message propagation along optimized edges.
• Leveraged dual sparsity from topology and event-driven computation for robustness and efficiency.

Main Results:

• Achieved exceptional robustness, with accuracy increasing from 30.14% to 64.58% under 0.1 salt-and-pepper noise.
• Demonstrated significant energy reduction of 97.28% compared to conventional Graph Neural Networks (GNNs).
• Maintained state-of-the-art accuracy (85.31% on Cora) while enhancing robustness and efficiency.

Conclusions:

• Principled minimization of structural entropy is a powerful strategy for enhancing SGNN robustness.
• The SSEL framework effectively mitigates adversarial topological perturbations.
• Combining information-theoretic graph principles with neuromorphic computing unlocks significant potential for robust and efficient AI.