There is More to Attention: Statistical Filtering Enhances Explanations in Vision Transformers

Attention weights alone aren't enough for understanding Vision Transformers: The paper shows that directly using attention maps from Vision Transformers produces "noisy maps" that don't clearly show what the model is actually focusing on, highlighting a key limitation in how we currently explain these increasingly popular AI models.

Statistical filtering significantly improves explanation clarity: By applying statistical filtering techniques (originally developed for older CNN models) to attention maps, the researchers produced "sharper and more interpretable maps," demonstrating that combining existing techniques can yield better explanations without requiring entirely new methods.

Human judgment should validate AI explanations: The researchers evaluated their method not just through computational metrics but also by comparing explanations to human eye-tracking data, arguing that "human interpretability remains essential for XAI"—a crucial reminder that explanations should align with how humans actually perceive and understand images.

Better explanations don't require sacrificing efficiency: The proposed approach maintains computational efficiency while outperforming state-of-the-art methods, suggesting that improving AI explainability doesn't necessarily mean accepting slower or more resource-intensive systems.

Explanations can be tailored to specific decisions: The paper introduces a "class-specific variant" that produces "discriminative explanations," meaning the method can highlight different features depending on what decision the AI is making, providing more targeted insights into model behavior.