Poster: Stochastic Weight Sharing for Bayesian Neural Networks

This repository contains the poster for the paper "Stochastic Weight Sharing for Bayesian Neural Networks". The work is presented at The 28th International Conference on Artificial Intelligence and Statistics (AISTATS) The main contributions: We introduce a stochastic weight-sharing technique specifically-tailored to BNNs and that employs Wasser stein distance, gradients, and within-class variance in order to improve model efficiency. Fast BNN Training: Train large BNNs (ResNet, ViT) with 2D Gaussians and Wasserstein loss, cutting compute costs. Compact and Accurate: Up to 50× parameter compression and 75% size reduction, with negligible loss in accuracy. SOTA Accuracy/Uncertainty: Matches full BNNs on CIFAR/ImageNet. Article: https://proceedings.mlr.press/v258/lin25a.html Github link: https://github.com/moulelin/2DGBNN