When Explanations Lie: Why Many Modified BP Attributions Fail

This repository provides the code for our paper "When Explanations Lie: Why Many Modified BP Attributions Fail" under review for ICML.

To create an enviroment with the correct package versions:

$ conda create --name wel --file ./conda_list_output.txt

Additionally, you need to install the deeplift and innvestigate packages. After you activated the wel conda enviroment:

$ cd <root dir>
$ pip install ./repos/deeplift ./repos/innvestigate

Finally, you have to specify the path of the ImageNet dataset in imagenet_dir.json.

Sources of the images for the two_classes notebook, both are under a free license:

Elefant & Zebra (Free for commercial use, No attribution required): https://pixabay.com/photos/zebra-elephant-africa-safari-3742242/

Name		Name	Last commit message	Last commit date
Latest commit History 12 Commits
development		development
images		images
.gitignore		.gitignore
README.md		README.md
conda_list_output.txt		conda_list_output.txt
convergence_simulation.ipynb		convergence_simulation.ipynb
cosine_similarity_convergence.ipynb		cosine_similarity_convergence.ipynb
create_code_submission.sh		create_code_submission.sh
deeplift_resnet.py		deeplift_resnet.py
imagenet_dir.json		imagenet_dir.json
imagenet_dir.json.template		imagenet_dir.json.template
imagenet_labels.json		imagenet_labels.json
monkey_patch_lrp_resnet.py		monkey_patch_lrp_resnet.py
pattern_attr.ipynb		pattern_attr.ipynb
sanity_checks.ipynb		sanity_checks.ipynb
train_cifar10.ipynb		train_cifar10.ipynb
two_classes.ipynb		two_classes.ipynb
when_explanations_lie.py		when_explanations_lie.py

Provide feedback