New SNN Algorithm Closes Gap with ANNs Using Circulate-Firing Neurons

Spiking Neural Networks (SNNs) promise ultra-low energy consumption but have historically lagged behind Artificial Neural Networks (ANNs) in accuracy. A new paper from researchers Feifan Zhou, Xiang Wei, Yang Liu, and Qiang Yu tackles this gap with three key innovations. First, they introduce a circulate-firing neuron model that better leverages membrane potential dynamics to encode more information per spike. Second, they replace fixed surrogate gradient functions with time-step-wise learnable gradients, enabling more accurate gradient flow during backpropagation. Third, a positive-negative balanced loss function optimizes both excitatory and inhibitory membrane potentials, further stabilizing training.

The results are striking: the proposed algorithm achieves competitive performance across multiple benchmark datasets and, crucially, generalizes seamlessly to advanced architectures like Transformers, consistently outperforming prior SNN methods. This suggests that with proper training techniques, SNNs can rival ANN accuracy without sacrificing their energy efficiency advantage. The work opens a new avenue for high-performance spiking architectures, potentially enabling real-time, low-power AI on edge devices.