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

Better explanations require measuring both accuracy and clarity: The paper introduces RCAP, a new evaluation metric that assesses not just whether explanations point to the right areas (localization) but also whether they avoid highlighting irrelevant noise. This dual focus addresses a gap in how explanation quality is currently measured in AI systems.

Gradient magnitude matters more than previously recognized: Guided AbsoluteGrad outperformed seven other gradient-based explanation methods by paying closer attention to the strength of gradients (both positive and negative), suggesting that current XAI approaches may be overlooking important signal in how strongly different parts of an image influence a model's decision.

Medical AI explanations need rigorous evaluation: Testing on the ISIC skin imaging dataset demonstrates that explanation methods perform differently across domains—what works well for general image recognition (ImageNet) may not work equally well for medical applications, highlighting the need for domain-specific validation of AI explanations in high-stakes fields.

Visual noise in explanations is a real problem worth solving: The paper formally proves that evaluating "visual noise level" (irrelevant highlighted areas) is necessary alongside localization accuracy, suggesting that some current explanation methods may mislead users by highlighting distracting or irrelevant image regions alongside correct ones.