Self-Attention

This paper presents an interpretability method for self-attention based models, and specifically for Transformer encoders. The method incorporates LRP and gradients, and achieves SOTA results for ViT, BERT, and DeiT.