STAR-SNN: A spatio-temporal adaptive recurrent spiking neural network with separated propagation surrogate gradient for hardware efficient real-time learning

Abstract

Backpropagation Through Time (BPTT) trains Recurrent Spiking Neural Networks (R-SNNs) effectively but incurs high computational and memory costs, limiting real-time applications. To mitigate resource demands, we adopt truncated BPTT (K=1), reducing memory cost by three orders of magnitude. However, this truncation weakens sequence learning by limiting gradient propagation. To compensate, we introduce the Spatio-temporal Adaptive Recurrent Spiking Neural Network (STAR-SNN), which incorporates adaptive parameters to enhance high-dimensional representations and effectively retain sequence information despite truncation. Additionally, R-SNNs suffer from unstable training due to the entanglement of spike generation and suppression in weight updates. To resolve this, we develop Separated Propagation Surrogate Gradient (SPSG), which decouples these processes by selectively propagating error signals, stabilizing learning and improving convergence. Our approach achieves a 393-fold reduction in MSE loss for chaotic system forecasting and delivers high performance in event-driven DVS-Gesture recognition, establishing a scalable, hardware-efficient framework for real-time neuromorphic computing.