Bayesian Learning of Neural Network Architectures

Disclaimer: This repository contains a reimplementation of the methods presented in this paper. This is not the original code used in the experimental section of the paper.

For a high-level overview of the approach refer to the following blog post.

Project structure

The bal package contains a Pytorch implementation of the Bayesian adaptive size and skip connection layers, a custom implementation of a truncated normal distribution by subclassing torch.Distribution, and some other utilities.

The examples folder contains notebooks and utility functions to visually demonstrate the approach described in the paper and the blog post mentioned above.

Citation

@InProceedings{pmlr-v89-dikov19a,
  title     = 	 {Bayesian Learning of Neural Network Architectures},
  author    = 	 {Dikov, Georgi and Bayer, Justin},
  booktitle = 	 {Proceedings of Machine Learning Research},
  pages     = 	 {730--738},
  year      = 	 {2019},
  volume    = 	 {89},
  month     = 	 {16--18 Apr}
}