Abstract
Spiking Neural Networks (SNNs) have attracted significant research attention due to their superior temporal information processing capability, low power operation, and biological plausibility. In SNN training, typically, only the synaptic weights are adjusted. At the same time, neuron-specific parameters such as membrane potential time constant ( τ mem ) and threshold voltage (V th ) remain fixed, often uniform across the network and manually tuned to an optimal value. However, studies have demonstrated that the computational performance of SNNs can be substantially enhanced by training neuron parameters alongside synaptic weights during network training [1] . Additionally, developing artificial neurons with programmable τ mem is crucial to effectively processing a broad spectrum of temporal signals. Therefore, there is a need to develop artificial neurons with adjustable τ mem and V th to not only facilitate the hardware implementation of such novel training algorithms but also to improve the performance of SNNs by exploiting the diversity in neuronal dynamics.