Graph Spiking Attention Network: Sparsity, Efficiency and Robustness.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Existing Graph Attention Networks (GATs) use dense attention, making them sensitive to noisy graph data.
• Current sparse GATs face challenges with high training complexity and difficulties in inductive learning.

Purpose of the Study:

• To develop a novel sparse Graph Attention Network (GAT) that overcomes the limitations of existing methods.
• To enhance robustness against graph noise and simplify inductive learning.

Main Methods:

• Proposed Graph Spiking Attention (GSAT), leveraging the spiking neuron (SN) mechanism.
• Utilized SNs to generate sparse attention coefficients, creating an edge-sparsified graph for GNNs.
• Enabled message passing on selective neighbors for compact and robust processing.

Main Results:

• GSAT demonstrated effectiveness in learning sparse attention coefficients.
• The method naturally performs message passing on selected neighbors, enhancing robustness.
• GSAT proved straightforward for inductive learning tasks.

Conclusions:

• GSAT effectively addresses the noise sensitivity and complexity issues of traditional GATs.
• The spiking neuron mechanism provides a robust and efficient solution for sparse graph attention.
• GSAT shows significant promise for both transductive and inductive graph learning applications.