Header

UZH-Logo

Maintenance Infos

Towards better understanding of gradient-based attribution methods for Deep Neural Networks


Ancona, Marco; Ceolini, Enea; Öztireli, Cengiz; Gross, Markus (2018). Towards better understanding of gradient-based attribution methods for Deep Neural Networks. 6th International Conference on Learning Representations (ICLR) 1711.06104, Institute of Neuroinformatics.

Abstract

Understanding the flow of information in Deep Neural Networks (DNNs) is a challenging problem that has gain increasing attention over the last few years. While several methods have been proposed to explain network predictions, there have been only a few attempts to compare them from a theoretical perspective. What is more, no exhaustive empirical comparison has been performed in the past. In this work, we analyze four gradient-based attribution methods and formally prove conditions of equivalence and approximation between them. By reformulating two of these methods, we construct a unified framework which enables a direct comparison, as well as an easier implementation. Finally, we propose a novel evaluation metric, called Sensitivity-n and test the gradient-based attribution methods alongside with a simple perturbation-based attribution method on several datasets in the domains of image and text classification, using various network architectures.

Abstract

Understanding the flow of information in Deep Neural Networks (DNNs) is a challenging problem that has gain increasing attention over the last few years. While several methods have been proposed to explain network predictions, there have been only a few attempts to compare them from a theoretical perspective. What is more, no exhaustive empirical comparison has been performed in the past. In this work, we analyze four gradient-based attribution methods and formally prove conditions of equivalence and approximation between them. By reformulating two of these methods, we construct a unified framework which enables a direct comparison, as well as an easier implementation. Finally, we propose a novel evaluation metric, called Sensitivity-n and test the gradient-based attribution methods alongside with a simple perturbation-based attribution method on several datasets in the domains of image and text classification, using various network architectures.

Statistics

Downloads

3 downloads since deposited on 12 Mar 2019
2 downloads since 12 months
Detailed statistics

Additional indexing

Item Type:Working Paper
Communities & Collections:07 Faculty of Science > Institute of Neuroinformatics
Dewey Decimal Classification:570 Life sciences; biology
Language:English
Date:2018
Deposited On:12 Mar 2019 11:34
Last Modified:25 Sep 2019 00:27
Publisher:Arxiv - Computer Science
Series Name:6th International Conference on Learning Representations (ICLR)
OA Status:Green
Free access at:Official URL. An embargo period may apply.
Official URL:https://arxiv.org/abs/1711.06104

Download

Green Open Access

Download PDF  'Towards better understanding of gradient-based attribution methods for Deep Neural Networks'.
Preview
Content: Published Version
Filetype: PDF
Size: 2MB