Deeper Understanding of Black-box Predictions via Generalized Influence Functions

This repository is the official implementation of Deeper Understanding of Black-box Predictions via Generalized Influence Functions.

Requirements

This repository is conda-based. Please install Conda first, then precede the installation. To install requirements:

conda env create -f conda_env.yaml
conda activate GIF
cd GIF && source python_path.sh

Training

You can train the model by using the train.py. Please change the preset in the config.py.

python train.py

Results

All figures and tables can be reproduced by the notebook in the scripts

Acknowledgements

I referred to the following repos:

• Models: kuangliu/pytorch-cifar
• Lanczos method:noahgolmant/pytorch-hessian-eigenthings

Reference

@misc{lyu2023deeper,
      title={Deeper Understanding of Black-box Predictions via Generalized Influence Functions}, 
      author={Hyeonsu Lyu and Jonggyu Jang and Sehyun Ryu and Hyun Jong Yang},
      year={2023},
      eprint={2312.05586},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}