Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation

Different AI architectures "see" differently than humans: The study found that CNNs focus broadly across images in ways that align with human attention, while Vision Transformers (ViTs) concentrate on tiny, specific details that humans might miss—raising questions about whether models are learning robust features or exploiting dataset quirks in ways humans wouldn't.

Medical AI requires special scrutiny: The Mamba architecture performed well on medical imaging tasks but showed problematic attention patterns on natural images, demonstrating that an AI system's explainability can't be assumed to transfer across different types of tasks—a critical finding for high-stakes applications like healthcare.

Eye-tracking provides objective validation: Rather than relying on subjective visual inspection, researchers used eye-tracking data as ground truth to quantitatively measure whether model attention maps actually match where humans look, providing a more rigorous standard for evaluating if AI explanations are trustworthy.

Current explanation methods have blind spots: The three complementary similarity metrics (Jensen-Shannon, Chi-square, and Pearson correlation) revealed that models can appear well-aligned with human attention in some measures while showing dangerous divergences in others—suggesting single validation approaches may miss important failure modes.

Explainability varies by task complexity: Fine-grained classification, medical imaging, and emotion recognition showed different alignment patterns across architectures, indicating that "explainability" isn't a universal property but depends on the specific application, requiring task-specific validation before deployment.