This is the official repository for the paper Concept Gradient: Concept-based Interpretation Without Linear Assumption published in ICLR 2023. Concept gradients is a drop-in replacement for Concept Activation Vectors (CAVs) to improve the representation of concept with non-linear functions while following the same line of intuition for constructing attribution. You can find the code to reproduce all the experiment results.
Setup environment
conda env create -n cg python
conda activate cg
pip install -e .
The scripts for training and attributing various models and datasets are in ./scripts.
Running concept gradients on your custom models and datasets is rather simple. src/cg/concept_gradients.py is a standalone file that performs the full functionality of attributing with concept gradients. You can download just this file to your local directory and make sure both torch and captum are installed as prerequisites. Our implementation depends on the Captum library to calculate input gradients and with respect to a certain layer. See the commets in file for more argument details. Example usage of concept gradients can be found here.
Please cite our work if you find this repo useful
@inproceedings{
bai2023concept,
title={Concept Gradient: Concept-based Interpretation Without Linear Assumption},
author={Andrew Bai and Chih-Kuan Yeh and Neil Y.C. Lin and Pradeep Kumar Ravikumar and Cho-Jui Hsieh},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=_01dDd3f78}
}