This repository provides the code to reproduce the results in the paper: "On the expected behaviour of noise regularised neural networks as Gaussian processes."
The code was written by Arnu Pretorius. Large portions of the code was originally adapted from code that was made available by Lee et al. (2018) at https://github.com/brain-research/nngp.
To reproduce Figures 1, 3 and 4 all that is required is numpy
, pandas
, seaborn
and matplotlib
. Each figure corresponds to a notebook in the repository.
To regenerate the results for Figures 1 and 3, follow the instructions given below.
Step 1. Install Docker and nvidia-docker.
docker pull ufoym/deepo:tensorflow-py27-cu90
git clone https://github.com/arnupretorius/noisyNNGPs_2019.git
Change directory to the cloned repository on your local machine and run the bash script env_up.sh
to start the docker container with the correct environment.
Next change directory to noisy_nngps
, and run the following script run_exp.sh
.
- Pretorius, A., Kamper, H., & Kroon, S. On the expected behaviour of noise regularised neural networks as Gaussian processes. Under review: NeurIPS, 2019.
- Lee, J., Bahri, Y., Novak, R., Schoenholz, S.S., Pennington, J. and Sohl-Dickstein, J. Deep neural networks as gaussian processes. ICLR, 2018.