MPLIF: Multi-parametric leaky integrate-and-fire neuron for spiking neural networks.

Abstract

Spiking Neural Networks (SNNs) have garnered significant attention due to their ability to process temporal information efficiently with low power consumption and high biological plausibility. In prevailing SNNs, spiking neuron models play a crucial role and have led to extensive research on variants of neuron models. Despite significant performance and stability improvements achieved by these spiking neuron variants, they typically require task-specific hyperparameter tuning and additional architectural complexity. This reliance often leads to reduced generalization, unstable convergence, and limited applicability in edge devices constrained by energy and memory budgets. To enhance the self-adaptive capability of spiking neurons while preserving the brain-inspired computational characteristics, this study proposes a spiking neuron model, termed Multi-Parametric Leaky Integrate-and-Fire (MPLIF) neuron. The proposed spiking neuron improves the adaptability and generalization of SNNs by utilizing novel self-adaptive mechanisms across all neurodynamic computation processes. Extensive experiments are conducted on six benchmark datasets, including CIFAR-10, CIFAR-100, Caltech101, DVS128-Gesture, CIFAR10-DVS, and N-Caltech101, using Spiking ResNet-18 and Spiking VGG-11 backbones. Under identical training settings, MPLIF-based SNNs consistently outperform LIF and advanced variants, achieving up to 2.1%-4.3% accuracy improvement on static datasets and 1.6%-3.8% improvement on neuromorphic datasets. The source code is available at: https://github.com/JeffRody/MPLIF.