Explaining image classifiers by removing input features using generative models

TL;DR

This paper introduces a method that uses generative models to remove input features in image classifiers, resulting in more realistic counterfactuals and improved explanation accuracy across multiple datasets and models.

Contribution

The paper integrates generative inpainters into attribution methods, enhancing the realism and accuracy of feature removal explanations for image classifiers.

Findings

More plausible counterfactual samples generated

Improved accuracy in object localization, deletion, and saliency metrics

Greater robustness to hyperparameter variations

Abstract

Perturbation-based explanation methods often measure the contribution of an input feature to an image classifier's outputs by heuristically removing it via e.g. blurring, adding noise, or graying out, which often produce unrealistic, out-of-samples. Instead, we propose to integrate a generative inpainter into three representative attribution methods to remove an input feature. Our proposed change improved all three methods in (1) generating more plausible counterfactual samples under the true data distribution; (2) being more accurate according to three metrics: object localization, deletion, and saliency metrics; and (3) being more robust to hyperparameter changes. Our findings were consistent across both ImageNet and Places365 datasets and two different pairs of classifiers and inpainters.