MMD-critic

Work-in-progress implementation of

Examples are not Enough, Learn to Criticize! Criticism for Interpretability. Been Kim, Rajiv Khanna, Sanmi Koyejo. NeurIPS 2016

Setup

git clone https://github.com/maxidl/MMD-critic.git
cd MMD-critic

Installing Dependencies

• pytorch (for now everything is run on cpu)
• matplotlib
• scikit-learn (for digits only)
• torchvision (for imagenet only)
• tqdm (for imagenet only)

Example dependency install using conda with cpu only pytorch:

conda create -n mmd-critic python=3.8 scikit-learn matplotlib
conda activate mmd-critic
conda install -c pytorch pytorch cpuonly 

Downloading data for the digits example

mkdir ./data
wget https://raw.githubusercontent.com/BeenKim/MMD-critic/master/data/usps -P ./data
wget https://raw.githubusercontent.com/BeenKim/MMD-critic/master/data/usps.t -P ./data

Digits example

python digits.py

Imagenet example

python imagenet.py

This example requires the ImageNet dataset. Make sure to set imagenet_root inside of imagenet.py.

Some example plots for class_name = 'Blenheim spaniel'

Using image embeddings (2084 dim) from resnet50

Using raw images (flattened):

Reference

• Paper: https://proceedings.neurips.cc/paper/2016/hash/5680522b8e2bb01943234bce7bf84534-Abstract.html
• If you find this code useful for your research, please cite the following:

  @inproceedings{kim2016MMD,
  title={Examples are not Enough, Learn to Criticize! Criticism for Interpretability},
  author={Been Kim, Rajiv Khanna and Sanmi Koyejo },
  booktitle={Advances in Neural Information Processing Systems},
  year={2016}
  }
  

• This code is based on the authors original Implementation, consider leaving a ⭐ on their repository to support authors releasing their code.
• Interpretable ML Book: Chapter on Prototypes and Criticisms

To-Do

• add 1NN classifier
• add more examples