<ul><li>A Bayesian framework for training binary and spiking neural networks has been proposed.</li><li>The framework achieves state-of-the-art performance without normalization layers.</li><li>Importance-weighted straight-through (IW-ST) estimators have been introduced as a part of this approach.</li><li>Experiments on CIFAR-10, DVS Gesture, and SHD show that this method matches or exceeds existing approaches without normalization or hand-tuned gradients.</li></ul>

A Principled Bayesian Framework for Training Binary and Spiking Neural Networks

Discover more