Concept-based XAI Library

CXAI is an open-source library for research on concept-based Explainable AI (XAI).

CXAI supports a variety of different models, datasets, and evaluation metrics, associated with concept-based approaches:

High-level Specs:

Methods:

• Now You See Me (CME): Concept-based Model Extraction.
• Concept Bottleneck Models
• Weakly-Supervised Disentanglement Without Compromises
• Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations
• Concept Whitening for Interpretable Image Recognition
• On Completeness-aware Concept-Based Explanations in Deep Neural Networks
• Towards Robust Interpretability with Self-Explaining Neural Networks

Datasets:

• dSprites
• Shapes3D
• Cars3D
• SmallNorb

to get the datasets run script datasets/download_datasets.sh

Requirements

• Python 3.7 - 3.8
• See 'requirements.txt' for the rest of required packages

Installation

If installing from the source, please proceed by running the following command:

python setup.py install

This will install the concepts-xai package together with all its dependencies.

To test that the package has been successfully installed, you may run:

import concepts_xai
help("concepts_xai")

to display all the subpackages included from this installation.

Subpackages

• datasets: datasets to use, including task functions.
• evaluation: different evaluation metrics to use for evaluating our methods.
• experiments: experimental setups (To-be-added soon)
• methods: defines the concept-based methods. Note: SSCC defines wrappers around these methods, that turn then into semi-supervised concept labelling methods.
• utils: contains utility functions for model creation as well as data management.

Citing

If you find this code useful in your research, please consider citing:

@article{kazhdan2021disentanglement,
  title={Is Disentanglement all you need? Comparing Concept-based \& Disentanglement Approaches},
  author={Kazhdan, Dmitry and Dimanov, Botty and Terre, Helena Andres and Jamnik, Mateja and Li{\`o}, Pietro and Weller, Adrian},
  journal={arXiv preprint arXiv:2104.06917},
  year={2021}
}

This work has been presented at the RAI, WeaSuL, and RobustML workshops, at The Ninth International Conference on Learning Representations (ICLR 2021).