New Algorithm Boosts Direct Training of Spiking Neural Networks

Researchers propose a novel direct training algorithm for Spiking Neural Networks (SNNs) to close the performance gap with Artificial Neural Networks (ANNs). The algorithm introduces three innovations: a circulate-firing spiking neuron model that enhances information representation by better utilizing membrane potentials; a time-step-wise learnable surrogate gradient function for accurate gradient estimation during backpropagation; and a positive-negative balanced loss function. The work addresses key limitations of conventional spiking neurons and fixed surrogate gradients, aiming to improve SNN energy efficiency and performance. The paper is available on arXiv under reference 2605.27412.