STRA-SNN: Spatio-Temporal Residual Attention Mechanism in Spiking Neural Network for Efficient Classification of Event Stream Data

Spiking Neural Networks (SNNs), as a significant algorithmic model in neuromorphic computing, have garnered widespread attention due to their advantages in temporal information processing, low power consumption, and high biological plausibility. The potential of SNNs in extracting spatiotemporal features makes them particularly suitable for processing event stream data. However, existing SNN attention modules mainly focus on a single dimension of the data, making it difficult to fully extract both temporal and spatial features. To address this, this study draws inspiration from biological synapses and residual blocks, proposing a spatiotemporal residual attention mechanism to enhance the spatiotemporal receptive fields of spiking neurons, thereby establishing stronger spatiotemporal dependencies between network layers. In addition, we integrate memristors with neuron circuits to create memristor-coupled neurons, which further enhances the pulse firing mechanism of the neurons. Experimental results indicate that both methods significantly improve the performance of SNNs in event stream classification tasks. Our method was tested on the neuromorphic datasets SHD and SSC, achieving accuracies of 94.83% and 77.56%, respectively, indicating that its performance is comparable to current state-of-the-art methods.