Tianzhu Ye

Oct 8, 2024 | huggingface.co | Tianzhu Ye |Li Dong |Yuqing Xia |Yutao Sun

Abstract Transformer tends to overallocate attention to irrelevant context. In thiswork, we introduce Diff Transformer, which amplifies attention to the relevantcontext while canceling noise. Specifically, the differential attentionmechanism calculates attention scores as the difference between two separatesoftmax attention maps. The subtraction cancels noise, promoting the emergenceof sparse attention patterns.