self-correcting-networks

Library for training self-correcting neural networks that automatically revise their outputs to satisfy non-relational safety properties.

This repository implements the method described in this paper (appearing in FoMLAS 2022). If you use this code, please use the following citation:

@INPROCEEDINGS{leino21self-correcting,
  title = {Self-Correcting Neural Networks for Safe Classification},
  author = {Klas Leino and Aymeric Fromherz and Ravi Mangal and Matt Fredrikson and Bryan Parno and Corina Păsăreanu},
  booktitle = {Workshop on Formal Methods for ML-Enabled Autonomous Systems (FoMLAS)},
  year = {2022}
}

Installation

We recommend performing installation within a virtual environment. The library can be installed via the following steps:

1. Clone the repository and change into its root directory.
2. Install from source via

pip install -e .

Code to run the experiments from the paper can be found in the evaluation.py files located in scnet/acas (for ACAS Xu experiments), scnet/collision (for collision avoidance experiments), and scnet/misc (for CIFAR-100 and synthetic experiments). To run the experiments on synthetic data, you must first generate the synthetic datasets using scnet/misc/synthetic_data.py.