This repo contains the I-Repair code for localised repairs of Neural Networks published in https://dl.acm.org/doi/10.1145/3477314.3507059.
I-repair is used to repair a set of misclassified inputs by iteratively rescaling weights connected to individual nodes. In order to determine good repair-node candidates, I-Repair requires a small set of correctly classified inputs which is used to heuristically identify nodes that have a large impact on the repair set and a small impact on the correctly classified set.
All dependencies can be installed via pipenv:
$ cd <your_repair_tool_path>/
$ pipenv install
The experiments reported in the paper can be run with the scripts in the repair_tool/evaluation/benchmark_scripts/ folder; the results are stored in repair_tool/benchmark_results.
Remember to enable the pipenv environment before running the library with:
cd <your_path>/repair_tool/ pipenv shell
Patrick Henriksen: [email protected]
Francesco Leofante: [email protected]
Alessio Lomuscio