Reproduce “ViT-CX: Causal Explanation of Vision Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, Nevin L.Zhang
ViT-CX is a mask-based method for causally explaining Vision Transformer predictions, generating more faithful and interpretable saliency maps by addressing ViT-specific challenges. It achieves superior faithfulness (e.g., 10% lower Deletion AUC on ViT-B) and interpretability (e.g., 86.42% Pointing Game accuracy on ViT-B) compared to existing methods, while maintaining computational efficiency.
Key insights
- Attention weights in Vision Transformers are often insufficient for faithful causal explanations, as they primarily indicate importance to the `[CLS]` token rather than semantic content.
- Vision Transformers are highly sensitive to the *shape* of masked regions, necessitating a debiased causal impact score to prevent artifactual explanations.
- Causal overdetermination is prevalent in ViTs, making pixel coverage bias correction essential for accurately attributing importance and generating meaningful saliency maps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.