Do Explanations Explain? Model Knows Best

TL;DR

This paper introduces an empirical framework that uses neural networks to evaluate the reliability of explanation methods by testing if they conform to specific axioms through controlled experiments.

Contribution

It proposes a novel, model-based evaluation framework for explanation methods, enabling systematic and axiomatic assessment of their trustworthiness.

Findings

Different explanation methods often highlight different features.

The framework can reveal properties and limitations of explanation methods.

It provides a toolset for evaluating current and future explanation techniques.

Abstract

It is a mystery which input features contribute to a neural network's output. Various explanation (feature attribution) methods are proposed in the literature to shed light on the problem. One peculiar observation is that these explanations (attributions) point to different features as being important. The phenomenon raises the question, which explanation to trust? We propose a framework for evaluating the explanations using the neural network model itself. The framework leverages the network to generate input features that impose a particular behavior on the output. Using the generated features, we devise controlled experimental setups to evaluate whether an explanation method conforms to an axiom. Thus we propose an empirical framework for axiomatic evaluation of explanation methods. We evaluate well-known and promising explanation solutions using the proposed framework. The framework provides a toolset to reveal properties and drawbacks within existing and future explanation solutions.