Skip to content

Latest commit

 

History

History
5 lines (3 loc) · 357 Bytes

200709 Fast Transformers with Clustered Attention.md

File metadata and controls

5 lines (3 loc) · 357 Bytes

https://arxiv.org/abs/2007.04825

Fast Transformers with Clustered Attention (Apoorv Vyas, Angelos Katharopoulos, François Fleuret)

attention 쿼리들을 k-means로 클러스터링한 다음 같은 클러스터에 있는 쿼리들은 같은 attention weight를 사용하게 해서 연산량을 감소시킴. lsh 다음은 k-means! #attention #transformer