AM-SGCN: Tactile Object Recognition for Adaptive Multichannel Spiking Graph Convolutional Neural Networks

In recent years, event-driven tactile data learning and event-driven spiking neural network (SNN) characteristics have provided new methods for tactile event perception. However, it is difficult to effectively extract tactile information from highly irregular taxel structures; additionally, the step property of spikes leads to nondifferentiability, making it difficult to directly apply the error backpropagation method to the training processes of SNNs. In this context, we propose an adaptive multichannel spiking graph convolutional network (GCN) framework that fully utilizes taxel features and topology information to build tactile topology and feature graphs for irregular taxels, and then extracts specific common tactile event data information from the feature, topology information, and their combinations via adaptive multichannel spiking graph convolution. The temporal spike sequence learning-based backpropagation algorithm is used to accurately calculate the spike gradient loss by analyzing the temporal and spatial dependence of neurons on the activity states of spike neurons. Experimental comparisons are conducted on the EvTouch-Objects tactile event dataset, and the experimental results show that the proposed method can effectively extract and fuse tactile information. Compared with the state-of-the-art methods, the proposed method achieves better classification performance, attaining an accuracy of 91.32% and a high precision rate of 92.43% in tests.