BorgwardtLab / Neural-Persistence

Code for the paper 'Neural Persistence: A Complexity Measure for Deep Neural Networks Using Algebraic Topology'

Neural Persistence: A Complexity Measure for Deep Neural Networks Using Algebraic Topology

This repository contains the code for our paper Neural Persistence: A Complexity Measure for Deep Neural Networks Using Algebraic Topology, which was published as an ICLR 2019 conference paper.

This repository is a work in progress. We aim to add more experiments over time.

Deep learning best practices in light of neural persistence

This repository can be used to reproduce the experiment from Section 4.1 of the publication. To ensure ease of use and reproducibility, it relies on Docker.

To install Docker, please follow the official manual. Having set up Docker for your operating system, the subsequent sections guide you through the process.

Build docker container

cd $REPODIR
docker build -t neuralpersistence .

Run experiments and summarize results

docker run -v $PWD/results/:/Neuralpersistence/results/ neuralpersistence python3 -u run_experiments.py
docker run -v $PWD/results/:/Neuralpersistence/results/ neuralpersistence python3 combine_runs.py results/runs/* --output results/combined_runs.csv

Plot the results

docker run -v $PWD/results/:/Neuralpersistence/results/ neuralpersistence python3 create_plots.py results/combined_runs.csv results/combined_runs.pdf

The visualizations of the mean normalized neural persistence, as well as the test accuracy can be found in results/combined_runs.pdf.

Citation

Please use the following citation to refer to this paper:

@inproceedings{Rieck19a,
  title     = {Neural Persistence: {A} Complexity Measure for Deep Neural Networks Using Algebraic Topology},
  author    = {Bastian Rieck and Matteo Togninalli and Christian Bock and Michael Moor and Max Horn and Thomas Gumbsch and Karsten Borgwardt},
  booktitle = {International Conference on Learning Representations~(ICLR)},
  year      = {2019},
  url       = {https://openreview.net/forum?id=ByxkijC5FQ},
}