Repeatability package for AISoLA 2023

This is the repeatability evaluation (RE) package for the paper The inverse problem for neural networks presented at AISoLA 2023. To cite the work, you can use:

@inproceedings{ForetsS24,
  author       = {Marcelo Forets and
                  Christian Schilling},
  title        = {The inverse problem for neural networks},
  booktitle    = {{(A)ISoLA}},
  series       = {LNCS},
  volume       = {14380},
  publisher    = {Springer},
  year         = {2024},
  url          = {https://doi.org/10.1007/978-3-031-46002-9\_14},
  doi          = {10.1007/978-3-031-46002-9\_14}
}

The code has since been integrated in the package NeuralNetworkReachability.jl.

How to use

First install the Julia compiler following the instructions here. Once you have installed Julia, open a terminal in the examples/ folder and execute

$ julia --project=. run_all.jl

to run all experiments. Alternatively, each experiment can be run individually. Check the file run_all.jl to identify the relevant scripts.

Each experiment creates plot files in the current folder.

Parabola experiment

The neural network in parabola2d/parabola2d.network was created with the script parabola2d/parabola2d_train.jl. This script results in different neural networks every time it is run.