Bayesian Neural Network Priors Revisited
Abstract
Isotropic Gaussian priors are the de facto standard for modern Bayesian neural network inference. However, it is unclear whether these priors accurately reflect our true beliefs about the weight distributions or give optimal performance. To find better priors, we study summary statistics of neural network weights in networks trained using stochastic gradient descent (SGD). We find that convolutional neural network (CNN) and ResNet weights display strong spatial correlations, while fully connected networks (FCNNs) display heavy-tailed weight distributions. We show that building these observations into priors can lead to improved performance on a variety of image classification datasets. Surprisingly, these priors mitigate the cold posterior effect in FCNNs, but slightly increase the cold posterior effect in ResNets.
Cite
@article{arxiv.2102.06571,
title = {Bayesian Neural Network Priors Revisited},
author = {Vincent Fortuin and Adrià Garriga-Alonso and Sebastian W. Ober and Florian Wenzel and Gunnar Rätsch and Richard E. Turner and Mark van der Wilk and Laurence Aitchison},
journal= {arXiv preprint arXiv:2102.06571},
year = {2022}
}
Comments
Accepted at ICLR 2022