How Can I Explain This to You? An Empirical Study of Deep Neural Network Explanation Methods [NeurIPS 2020]

This repository is the official implementation and it contains the code, trained models, and data used to generate explanations for each of the image, text, audio, and ECG domains.

Prerequisites

Install the required Python packages.

pip3 install -r requirements.txt

Contents

Data

Here's the Google Drive Link to the preprocessed data: Link

Download each of the folders there and place them in data/

Training Code

Inside the training_code/ folder, there are the Jupyter notebooks for training the models used during the study.

Trained Models

Inside the trained_models/ folder, there are the pretrained models used for the study, named as [domain].hdf5 for each of the domains: image, text, audio, ECG.

Explanations Code

Inside the explanations_code/ folder, there are the Jupyter notebooks for generating explanations using each of the methods used for the study for each domain.

Note: The font file included is required for generating some of the explanations for the text dataset.

Survey Responses

Inside the survey responses/ folder, we have the survey responses collected from the Amazon Turk Study and the code to process them.

Results from the AMT Study

The values indicate the rate by which users selected a particular method when it is an available explanation, with 95% bootstrap confidence intervals

BibTex

If you find this code and results useful in your research, please cite:

@article{jeyakumar2020can,
  title={How Can I Explain This to You? An Empirical Study of Deep Neural Network Explanation Methods},
  author={Jeyakumar, Jeya Vikranth and Noor, Joseph and Cheng, Yu-Hsi and Garcia, Luis and Srivastava, Mani},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}