Momentum-Enhanced Spiking Neuron Models for Improved Spatiotemporal Processing

Spiking neural networks (SNNs) have garnered significant attention in neuromorphic computing due to their biological plausibility and low-power advantages [1], [2]. Among various models, the leaky integrate-and-fire (LIF) neuron has become fundamental owing to its concise formulation and bio-physical realism [3]. However, traditional LIF models exhibit limitations in capturing temporal dependencies and dynamic information transmission, particularly when handling complex spatiotemporal patterns [4], [5]. The simplistic mechanism of LIF neurons struggles to capture hierarchical information in temporal sequences, limiting the representational capacity of SNNs. To address these limitations, this paper introduces a momentum mechanism to optimize neuronal dynamics. By tracking membrane potential variations, we endow neurons with the ability to perceive neural momentum, enabling detection of subtle differences between input patterns and generation of pattern-specific responses. Through theoretical analysis and experimental validation, we demonstrate significant improvements in spatiotemporal pattern recognition tasks, providing new insights for brain-inspired computing.