Quadratic integrate-and-fire neurons exhibit less fragmented loss landscapes and outperform leaky integrate-and-fire neurons in spike-based gradient descent
Researchers have demonstrated that quadratic integrate-and-fire (QIF) neurons offer a significant advantage over leaky integrate-and-fire (LIF) neurons for training spiking neural networks. Through a comparative study on the Spiking Heidelberg Digits dataset, QIF neurons showed superior performance and more stable training dynamics. The study visualized loss and gradient landscapes, revealing that LIF neurons exhibit fragmented and erratic gradients due to discontinuities, whereas QIF neurons provide a smoother training experience. AI
IMPACT Suggests QIF neurons could enable more stable and effective training for neuromorphic computing applications.