Recurrent spiking neural networks with bimodal neuronal time scales for learning performance enhancement.

Recurrent Spiking Neural Networks (RSNNs) represent a crucial paradigm in neuromorphic computing, with their performance heavily dependent on the design of neuronal time scales. Neuroscientific research has established that biological cortical circuits require the coordinated operation of both long- and short-timescale neurons to achieve efficient information processing. However, existing research on RSNNs predominantly adopt unimodal time scales covering short time constants, which formally simulates the heterogeneity characteristic of neuronal time constants. Furthermore, the direct integration of long- and short-timescale neurons fails to maintain the performance advantages of unimodal models, presenting a critical bottleneck for exploring multi-timescale collaboration. To address this challenge, we propose a Long- and Short-timescale RSNN (LSRSNN) model based on bimodal neuronal time scales. The core innovation lies in explicitly delineating long- and short-timescale neuronal populations, determining the time constants of these two neuron types through independent distributions, and developing a specialized learning algorithm tailored for LSRSNN. In experiments on working memory related tasks and speech recognition, LSRSNN demonstrates significant improvements in feature extraction and memory capabilities compared to traditional unimodal RSNNs. These results not only validate the effectiveness of bimodal time scales but also, for the first time, achieve effective collaborative training of long- and short-timescale neurons in RSNNs, providing a more efficient network architecture paradigm for neuromorphic applications such as speech processing and complex decision-making.