Guided AbsoluteGrad: Magnitude of Gradients Matters to Explanation's Localization and Saliency

TL;DR

This paper introduces Guided AbsoluteGrad, a gradient-based explanation method that leverages both positive and negative gradient magnitudes and a new evaluation metric, RCAP, to improve saliency map explanations across multiple datasets and models.

Contribution

It presents a novel gradient-based XAI method, Guided AbsoluteGrad, and a new evaluation metric RCAP, with theoretical justifications and extensive empirical validation.

Findings

Guided AbsoluteGrad outperforms existing gradient-based methods in saliency map quality.

The RCAP metric effectively evaluates localization and noise in explanations.

Gradient magnitude plays a crucial role in explanation quality.

Abstract

This paper proposes a new gradient-based XAI method called Guided AbsoluteGrad for saliency map explanations. We utilize both positive and negative gradient magnitudes and employ gradient variance to distinguish the important areas for noise deduction. We also introduce a novel evaluation metric named ReCover And Predict (RCAP), which considers the Localization and Visual Noise Level objectives of the explanations. We propose two propositions for these two objectives and prove the necessity of evaluating them. We evaluate Guided AbsoluteGrad with seven gradient-based XAI methods using the RCAP metric and other SOTA metrics in three case studies: (1) ImageNet dataset with ResNet50 model; (2) International Skin Imaging Collaboration (ISIC) dataset with EfficientNet model; (3) the Places365 dataset with DenseNet161 model. Our method surpasses other gradient-based approaches, showcasing the quality of enhanced saliency map explanations through gradient magnitude.