How Can One Choose the Best CAM-Based Explainability Method for a CNN Model?

There is no single "best" metric for evaluating AI explanations: The paper challenges the widespread reliance on Intersection over Union (IoU) as the standard metric for assessing explainability methods, suggesting that different metrics may better capture what makes an explanation useful to humans.

Human judgment should guide evaluation of AI explanations: The researchers used crowdsourcing to collect human preferences for explanations and compared these to automated metrics, finding that human perception of "good" explanations may differ from what traditional metrics measure—an important consideration for deploying AI in sensitive applications.

Current explainability methods need better validation: The study reveals a gap between how we mathematically measure explanations and how humans actually perceive them as helpful or trustworthy, indicating that existing CAM-based explainability methods may not be optimized for real-world human understanding.

Explainability evaluation requires multiple approaches: Rather than relying on a single metric like IoU, the paper advocates for using multiple distance metrics to assess whether AI explanations align with human perception, suggesting a more comprehensive evaluation framework is needed.

Explainability matters for high-stakes applications: The paper emphasizes that as deep learning models are increasingly used in sensitive domains affecting human life, having reliable methods to explain and validate those explanations becomes critical for responsible AI deployment.