A training-free Nyström attention variant with farthest-point sampling preserves ViT performance at lower cost, and masking attention-sink tokens yields modest downstream gains.
Lawrence Zitnick
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Artifacts and Attention Sinks: Structured Approximations for Efficient Vision Transformers
A training-free Nyström attention variant with farthest-point sampling preserves ViT performance at lower cost, and masking attention-sink tokens yields modest downstream gains.