Efficient event-based visual recognition with resonate-and-fire spiking neurons

Spiking Neural Networks (SNNs) offer a highly promising framework for processing event-based data, owing to their inherent temporal dynamics and spike-driven computation. However, existing SNN-based event models suffer from two critical limitations. First, they lack the ability to separate frequency-dependent information, making it impossible to distinguish background from motion cues. Second, the reset mechanism of Leaky Integrate-and-Fire neurons severely restricts the networks’ capacity to maintain long-term temporal information. To solve these problems, we propose an event visual model with Resonate-and-Fire (RF) neurons. Specifically, we propose a frequency divider module that achieves separation of high-frequency dynamic information and low-frequency background information through RF neurons. Additionally, we propose a parallel training and serial inference method that significantly enhances long-term memory capability while preserving asynchronous characteristics. Comprehensive experiments on multiple DVS datasets demonstrate that our proposed method achieves state-of-the-art performance, providing crucial technical support for developing efficient and effective neuromorphic vision models.