Incorporating memristive autapse in spatio-temporal attention SNN for neuromorphic speech recognition

Spiking neural networks (SNNs) have gained significant attention for their biological plausibility, event-driven operation, and low power consumption, establishing them as a leading model for processing event stream data. However, current models often oversimplify neuronal dynamics to balance computational cost and performance. To address this limitation and enhance the dynamical behavior of spiking neurons, this paper introduces two key innovations. First, inspired by biological autaptic connections and memristive devices, we propose the memristive autapse (M-Autapse), a self-connection mechanism that enables adaptive modulation of a neuron's membrane potential. Second, recognizing the need for attention mechanisms that match SNNs' spatio-temporal nature, we design a spatio-temporal synergistic attention (STSA) mechanism to bolster simultaneous focus on both temporal and spatial dimensions of input data. Extensive experiments on the neuromorphic speech benchmarks SHD and SSC validate our methods. On SHD, our model demonstrates performance competitive with the state-of-the-art, while also achieving strong results on the SSC dataset.