FAME: Feature Activation Map Explanation on Image Classification and Face Recognition

Deep neural networks lack transparency in their decision-making: The paper confirms that deep learning models used for image classification and face recognition operate as "black boxes," transforming pixel information into predictions through obscured processes that are difficult for humans to understand.

Existing explanation methods have fundamental limitations: The study demonstrates that Class Activation Mapping (CAM), a widely-used technique for explaining image AI systems, operates under a flawed assumption that "feature map locations are influenced only by underlying regions"—an assumption that "does not hold for deeper networks," meaning current explanations may be misleading.

A new hybrid approach improves AI interpretability: FAME combines gradient-based and perturbation-based explanation methods to produce more accurate attribution maps showing which parts of an image influenced a model's decision, achieving "competitive state-of-the-art" performance on both image classification and face recognition tasks.

Better explanations matter for high-stakes applications: The focus on face recognition systems is particularly significant for responsible AI, as these systems are increasingly used in security, law enforcement, and identity verification—contexts where understanding why a system made a decision is critical for fairness and accountability.