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

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer efficient temporal processing and low power consumption.
• Existing spiking neuron models often require extensive tuning and add complexity, limiting generalization and edge device applicability.

Purpose of the Study:

• To propose a novel self-adaptive spiking neuron model, the Multi-Parametric Leaky Integrate-and-Fire (MPLIF) neuron.
• To enhance the adaptability and generalization of SNNs while maintaining brain-inspired computational characteristics.

Main Methods:

• Introduced the MPLIF neuron model with novel self-adaptive mechanisms.
• Conducted extensive experiments on six benchmark datasets (CIFAR-10, CIFAR-100, Caltech101, DVS128-Gesture, CIFAR10-DVS, N-Caltech101).
• Utilized Spiking ResNet-18 and Spiking VGG-11 architectures for evaluation.

Main Results:

• MPLIF-based SNNs consistently outperformed standard Leaky Integrate-and-Fire (LIF) and advanced variants under identical training settings.
• Achieved accuracy improvements of 2.1%-4.3% on static datasets and 1.6%-3.8% on neuromorphic datasets.
• Demonstrated enhanced adaptability and generalization capabilities of the proposed neuron model.

Conclusions:

• The MPLIF neuron model offers a significant advancement in SNNs by improving self-adaptive capabilities.
• MPLIF provides a more generalized and stable solution compared to existing spiking neuron variants.
• The proposed model shows promise for energy-efficient and memory-constrained edge AI applications.