Concept-based Explanations for Out-Of-Distribution Detectors

This repository is the official implementation of the ICML 2023 paper: Concept-based Explanations for Out-Of-Distribution Detectors.

Requirements

• It is tested under Ubuntu Linux 16.04.1 and Python 3.6 environment, and requires some packages to be installed: see environment.yml for the list of dependencies of conda environment.

Running Experiments

• To learn concepts given a classifier (i.e., baseline):

python concept_learn.py --name {NAME_OF_EXPERIMENT} --num_concept {NUM_CONCEPTS} --gpu {GPU}
                        --coeff_concept {COEFF_EXPL}

• To learn concepts given a classifier and a detector (i.e., ours):

python concept_learn.py --name {NAME_OF_EXPERIMENT} --num_concept {NUM_CONCEPTS} --gpu {GPU} 
                        --coeff_concept {COEFF_EXPL}
                        --ood 
                        --score {TYPE_OF_DETECTOR} --coeff_score {COEFF_MSE}
                        --feat_l2 --coeff_feat {COEFF_NORM} 
                        --separability --coeff_separa {COEFF_SEP}

• To evaluate the learned concepts (e.g., detection completeness and concept separability):

python concept_eval.py --name {NAME_OF_EXPERIMENT} --result_dir {PATH_TO_CONCEPTS}
                       --out_data {OOD_DATASET} --gpu {GPU} 
                       --score {TYPE_OF_DETECTOR}
                       --separate
                       --visualize

Acknowledgements

We build upon the baseline code by Yeh et al., NeurIPS'20.

Citation

Please cite our work if you use this codebase:

@inproceedings{
choi2023concept-ood,
title={Concept-based Explanations for Out-of-Distribution Detectors},
author={Jihye Choi and Jayaram Raghuram and Ryan Feng and Jiefeng Chen and Somesh Jha and Atul Prakash},
booktitle={International Conference on Machine Learning},
year={2023}
}

License

Please refer to the LICENSE.