hnaik/vibi

Explaining a black-box using Deep Variational Information Bottleneck Approach

VIBI: Explaining a black-box using Deep Variational Information Bottleneck Approach

Overview

This repo includes pytorch implementation of VIBI. VIBI is a system-agnostic interpretable machine learning method that provides a brief but comprehensive explanation. It adopts the inspring information theoretic principle, information bottleneck principle. Using an information theoretic objective, VIBI selects instance-wise key features that are maximally compressed about an input (briefness), and informative about a decision made by a black-box on that input (comprehensive). The selected key features act as an information bottleneck that serves as a concise explanation for a black-box decision. Please see our recent paper -- arXiv preprint.

Usage

Download and install the environment from Cloud.

conda env create SeojinBang/py36
conda activate py36

See main.py for possible arguments.

To learn a black-box model for MNIST digit recognition:

cd mnist
python original.py --model_name original.ckpt --epoch 5

To learn VIBI to explain the black-box model:

python main.py --dataset mnist --epoch 40 --beta 0.1 --K 4 --explainer_type cnn4 --chunk_size 4 --mode train

Credit

DeepVIB Repo: pytorch implementation of deep variational information bottleneck.

L2X Repo: keras implementation of L2X.

References

Bang et al. 2019. Explaining a black-box using Deep Variational Information Bottleneck Approach. ArXiv Preprint arXiv:1902.06918.

Contact

Please feel free to contact me by e-mail seojinb at cs dot cmu dot edu, if you have any questions.