Learning-efficient spiking neural networks with multi-compartment spatio-temporal backpropagation.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation inspired by biological neurons.
• Traditional SNNs are limited by simplistic point neuron models, hindering complex learning.
• Simulating soma-dendrite interactions is crucial for advanced neural computation.

Purpose of the Study:

• Introduce a multi-compartment neuron model (MCN) for SNNs.
• Investigate the role of trainable cross-compartment connections in learning dynamics.
• Develop a stable backpropagation algorithm for MCNs.

Main Methods:

• Developed a multi-compartment spiking neuron model (MCN) with trainable cross-compartment connections.
• Provided theoretical proof that these connections act as spatiotemporal momentum.
• Proposed a multi-compartment spatiotemporal backpropagation (MCST-BP) algorithm for enhanced gradient flow.

Main Results:

• MC-SNNs demonstrated superior performance over traditional SNNs on benchmark datasets (S-MNIST, CIFAR-10, SHD, ECG).
• The MCN model significantly improved convergence speed and classification accuracy.
• Trainable cross-compartment connections were shown to guide learning dynamics toward global optima.

Conclusions:

• The MCN model effectively simulates soma-dendrite interactions, enhancing SNN capabilities.
• MCST-BP algorithm ensures stable gradient flow, facilitating effective training.
• This research provides a theoretical and practical foundation for high-performance brain-inspired learning systems.