UFOExplainer

Code for UFO: A unified method for controlling Understandability and Faithfulness Objectives in concept-based explanations for CNNs

To cite,

@misc{ramaswamy2023ufo,
      title={UFO: A unified method for controlling Understandability and Faithfulness Objectives in concept-based explanations for CNNs}, 
      author={Vikram V. Ramaswamy and Sunnie S. Y. Kim and Ruth Fong and Olga Russakovsky},
      year={2023},
      eprint={2303.15632},
      archivePrefix={arXiv},
}

Steps to run

1. Download Broden images + annotations as in https://github.com/CSAILVision/NetDissect-Lite
2. Get scene scores as done in https://github.com/princetonvisualai/OverlookedFactors
3. Use load_data.py to preprocess data
4. Use attr_rot.py to learn attribute classifiers
5. Use learn_attr_cutoff.py to learn explanations.
6. analsis.ipynb contains the analysis done in the paper.