TAS-GNN: Topology-Aware Spiking Graph Neural Networks for Graph Classification

SunJong Park, Hyeyoon Lee, Kanghyun Choi, Dain Kwon, Jongkil Park, Seongsik Park, Jinho Lee

Keywords: Spiking Neural Network, Graph Neural Network, Graph Classification

TL;DR: TAS-GNN, a spiking graph neural network, addresses neuron starvation in graph classification tasks by optimizing neuron utilization, resulting in up to a 27.20% performance improvement over traditional models.

Abstract:

The recent integration of spiking neurons into graph neural networks has been gaining much attraction due to its superior energy efficiency.
Especially because the irregular connection among graph nodes fits the nature of the spiking neural networks, spiking graph neural networks are considered strong alternatives to vanilla graph neural networks.
However, there is still a large performance gap for graph tasks between the spiking neural networks and artificial neural networks.
The gaps are especially large when they are adapted to graph classification tasks, where none of the nodes in the testset graphs are connected to the training set graphs.
We diagnose the problem as the existence of neurons under starvation, caused by the irregular connections among the nodes and the neurons. To alleviate the problem, we propose TAS-GNN.
Based on a set of observations on spiking neurons on graph classification tasks, we devise several techniques to utilize more neurons to deliver meaningful information to the connected neurons.
Experiments on diverse datasets show up to 27.20% improvement, demonstrating the effectiveness of the TAS-GNN.

Primary Area: Neuroscience and cognitive science (neural coding, brain-computer interfaces)

Submission Number: 14022