A Benchmark for Interpretability Methods in Deep Neural Networks
Abstract
We propose an empirical measure of the approximate accuracy of feature importance estimates in deep neural networks. Our results across several large-scale image classification datasets show that many popular interpretability methods produce estimates of feature importance that are not better than a random designation of feature importance. Only certain ensemble based approaches---VarGrad and SmoothGrad-Squared---outperform such a random assignment of importance. The manner of ensembling remains critical, we show that some approaches do no better then the underlying method but carry a far higher computational burden.
Cite
Text
Hooker et al. "A Benchmark for Interpretability Methods in Deep Neural Networks." Neural Information Processing Systems, 2019.Markdown
[Hooker et al. "A Benchmark for Interpretability Methods in Deep Neural Networks." Neural Information Processing Systems, 2019.](https://mlanthology.org/neurips/2019/hooker2019neurips-benchmark/)BibTeX
@inproceedings{hooker2019neurips-benchmark,
title = {{A Benchmark for Interpretability Methods in Deep Neural Networks}},
author = {Hooker, Sara and Erhan, Dumitru and Kindermans, Pieter-Jan and Kim, Been},
booktitle = {Neural Information Processing Systems},
year = {2019},
pages = {9737-9748},
url = {https://mlanthology.org/neurips/2019/hooker2019neurips-benchmark/}
}