Training Spiking Neural Networks with Real-Time Propagation Through Time

Yaokun Wang, Zhiying Long

Keywords: Spiking Neural Networks, Online Learning, Real-time Propagation Through Time, Drift of membrane potential distribution

TL;DR: We propose Real-time Propagation Through Time, a memory-efficient online training algorithm for SNNs that stabilizes convergence and mitigates drift of membrane potential distribution.

Abstract: Online learning algorithms for Spiking Neural Networks (SNNs) offer a memory-efficient alternative to Backpropagation Through Time (BPTT), but suffer from two critical issues: training instability and membrane potential distribution drift. To address these challenges, we introduce Real-Time Propagation Through Time (RPTT), a novel online learning framework. RPTT computes gradients using only the spatial component and integrates two synergistic regularization mechanisms: Membrane Potential Distribution Regularization (MPDR), which statistically constrains membrane potentials to counteract distributional drift, and Spatio-Temporal Gradient Regularization (STGR), which smooths weight updates to ensure stable convergence. We theoretically prove that RPTT converges to a stationary point. Extensive experiments on CIFAR-10/100, ImageNet-1k, and DVS-CIFAR10 demonstrate that RPTT achieves state-of-the-art performance while significantly reducing memory consumption. Experimental analysis reveals that RPTT achieves strong performance by effectively alleviating the membrane potential drift. Our work thus provides an effective framework for the online training of SNNs, significantly advancing their application in dynamic and realistic environments.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15399