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

TL;DR

This paper provides a formal comparison and unified framework for four gradient-based attribution methods in DNNs, introduces a new evaluation metric, and empirically assesses their effectiveness across multiple datasets and architectures.

Contribution

It offers a formal analysis of attribution methods, a unified framework for comparison, and introduces a novel evaluation metric for attribution quality.

Findings

Formal conditions of equivalence and approximation between methods

A new unified framework for gradient-based attribution methods

Empirical evaluation using Sensitivity-n metric across datasets

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.