A multisynaptic spiking neuron for simultaneously encoding spatiotemporal dynamics

Area of Science:

• Neuromorphic computing
• Computational neuroscience
• Artificial intelligence

Background:

• Spiking neural networks (SNNs) offer biological plausibility and computational power due to temporal dynamics.
• Standard SNN neurons face challenges in simultaneously encoding complex spatiotemporal input dynamics.

Purpose of the Study:

• Introduce the Multi-Synaptic Firing (MSF) neuron, inspired by biological multisynaptic connections.
• Enable SNNs to jointly encode spatial intensity and temporal dynamics for enhanced performance.

Main Methods:

• Propose the MSF neuron model with multiple synapses and varying thresholds on a postsynaptic neuron.
• Derive optimal threshold selection and parameter optimization for surrogate gradients.
• Implement and evaluate deep MSF-based SNNs on various benchmarks.

Main Results:

• MSF neurons generalize Leaky Integrate-and-Fire (LIF) and ReLU neurons.
• Achieve superior accuracy compared to LIF neurons while maintaining low power and latency.
• Outperform ReLU neurons in event-driven tasks, demonstrating high execution efficiency.

Conclusions:

• MSF neurons significantly advance neuromorphic computing capabilities.
• Enable scalable deep SNNs for real-world spatiotemporal applications without performance loss.