Spatiotemporal Decoupled Learning for Spiking Neural Networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) show promise for energy-efficient AI.
• Training SNNs effectively remains a challenge, with trade-offs between accuracy (Backpropagation Through Time) and memory efficiency (local learning methods).

Purpose of the Study:

• To introduce Spatiotemporal Decoupled Learning (STDL), a novel training framework for SNNs.
• To achieve both high accuracy and training efficiency in SNNs by decoupling spatial and temporal dependencies.

Main Methods:

• STDL partitions networks into subnetworks for independent training using auxiliary networks.
• Auxiliary networks are constructed under memory constraints to maintain subnetwork synergy.
• Temporal dependencies are decoupled for efficient online learning.

Main Results:

• STDL consistently outperforms local learning methods across seven vision datasets.
• STDL achieves accuracy comparable to Backpropagation Through Time (BPTT).
• STDL significantly reduces GPU memory costs, achieving a 4x reduction on ImageNet compared to BPTT.

Conclusions:

• STDL presents a promising approach for memory-efficient SNN training.
• The framework successfully balances accuracy and computational efficiency.
• This method paves the way for more practical SNN applications.