The effectiveness of feature attribution methods and its correlation with automatic evaluation scores

Official Implementation for the paper The effectiveness of feature attribution methods and its correlation with automatic evaluation scores (NeurIPS2021) by Giang Nguyen, Daeyoung Kim and Anh Nguyen.

_tldr: Explainable AI (XAI) human evaluation of feature attribution methods (a.k.a. heatmaps).

NeurIPS: Video presentation of the paper is here | Reviews (6,6,7,7)

Dataset

• Images used in Instructions, Training phase and Validation phase of all tested method can be found in Stimuli_Natural andStimuli_Dog in Human_experiments.tar.gz.
• Datasets (Natural/Adversarial ImageNet + Dog) described in Sec. 2.4.3 here and their visualizations using the tested methods can be found in Dataset and Visualization in Human_experiments.tar.gz.
• The data collected from online users and expert users via Gorilla.sc can be found in here.
  • The folders Data and Data-v2 were used to generate the statistics reported in Fig. 1b, Fig. 4, Table. A1, A2, A3, A4, A5, A8, A9 in here. Run ImageNet_analyze_spreadsheet.ipynb and Dogs_analyze_spreadsheet.ipynb to see the statistics.
  • The folder Expert_Data was used to generate the statistics reported in Table. 1 in here. Run Expert_ImageNet_analyze_spreadsheet.ipynb to see the statistics.

User interface

• This video walks you through one entire Gorilla.sc experiment run on a single user. We hope sharing all screens we carefully designed for this study could help future research.

If you wanna try out the User Interface from your device OR get the materials to replicate the exerpiment pipeline, drop me an email at [email protected] so that I could grant you the access.

Citation

If you find our work useful in your research, please consider citing:

@article{nguyen2021effectiveness,
  title={The effectiveness of feature attribution methods and its correlation with automatic evaluation scores},
  author={Nguyen, Giang and Kim, Daeyoung and Nguyen, Anh},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

License

MIT