G2T-SNN: Fusing Topological Graph and Gaussian Prior Spiking Neural Networks for Tactile Object Recognition

Inspired by the event-driven nature of biological systems and sparse spiking networks, research on event-driven tactile learning has been stimulated by recent advancements in event-driven tactile sensors and spiking neural networks (SNNs). However, a major challenge in tactile object recognition research lies in addressing the generalization problem caused by the complex spatiotemporal characteristics of high-dimensional event-driven tactile data. Additionally, sparse connections of SNN models are resulted by difficulties in fully utilizing timing information and structural features of tactile nodes, due to the limited representation of tactile information graphs. To address these issues comprehensively, this article proposes a fusing topological graph and Gaussian prior event-driven tactile object recognition SNNs. First, the incorporation of Gaussian prior effectively integrates prior knowledge to alleviate the generalization problem in event-driven tactile object recognition. Furthermore, both topological sorting and R-tree tactile graph construction methods are employed to enhance the connection structure within the tactile graph and resolve the issue of sparse connections. Finally, various approximate leaky integrate-and-fire neuron activation functions are introduced for comprehensive comparison and evaluation regarding their impact on performance in event-driven tactile object recognition. Experimental results using EvTouch-Containers and EvTouch-Objects datasets demonstrate that topological graph and Gaussian prior tactile object recognition (G2T)-SNN achieves superior recognition performance compared to TactileSGNet methods: 7.50% improvement on EvTouch-Containers dataset and 2.23% improvement on EvTouch-Objects dataset.