This repository provides the code to reproduce all the results in the paper: "Learning dynamics of linear denoising autoencoders." (ICML 2018)

To reproduce Figures 1-4, all that is required is numpy and matplotlib.

To reproduce Figures 5 and 6, a docker image was created to provide an identical research environment to the one used to run the initial experiments. Below are the instructions to reproduce these plots using this docker image and the notebooks provided.

docker pull arnu/research_env

git clone https://github.com/arnupretorius/lindaedynamics_icml2018.git

Change directory to the cloned repository on your local machine and run the bash script.

research_up.sh

This should create a volume bind mount with the current directory for persistent data storage as well as launch a Jupyter notebook accessible at https://0.0.0.0:8888/. Now, you can simply run the notebook corresponding to the figure in the paper you wish to reproduce.

To stop the docker container from running simply shutdown the notebook by pressing ctrl+c (the container will automatically be removed once stopped).