A continual inference formulation of Nyström attention that updates landmark-based low-rank attention incrementally, cutting per-step FLOPs by up to three orders of magnitude.
Trans- formers in the real world: A survey on NLP applications
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
Continual Low-Rank Scaled Dot-product Attention
A continual inference formulation of Nyström attention that updates landmark-based low-rank attention incrementally, cutting per-step FLOPs by up to three orders of magnitude.