🔍 Explainability in Vision Transformers

This repository contains an interpretability pipeline for Vision Transformers (ViTs) using a pretrained DeiT-Tiny model. It implements both Attention Rollout and Gradient Attention Rollout, and introduces a quantitative metric — Foreground–Background Attention Fraction (FAF/BAF) — using Mask R-CNN to measure how well ViTs focus on meaningful object regions.

The project also compares explanation behaviour between CNNs 🧠 and Vision Transformers 🤖, highlighting how transformers rely on global patch interactions rather than localized feature extraction. Experiments show that gradient-based rollout achieves ~25% higher foreground alignment, producing more precise and class-specific attribution maps.

---

✨ Features

• 🔄 Attention Rollout & Gradient Attention Rollout implementations
• 🧩 Patch-level attention visualization on the 14×14 ViT grid
• 🎯 FAF/BAF metric using Mask R-CNN for quantitative semantic focus evaluation
• ⚖️ Comparison of interpretability behaviour between CNNs and ViTs
• 📈 Reproducible experiments with visualizations

---