Research Notes

Curated overviews of research topics and completed studies.

F2HF: Feed-Forward Network as a Noisy Head Filter in Vision Transformer Explanations

F2HF: Feed-Forward Network as a Noisy Head Filter in Vision Transformer Explanations

We propose F2HF, a novel interpretability method for Vision Transformers that uses the feed-forward network as an attention filter, achieving a 16.4% improvement in IoU over baseline methods by significantly reducing noise.

vision-transformers interpretability attention-mechanism explainable-ai