MALA adds a scale-sensitive offset to linear attention so attention sharpens with query magnitude, improving accuracy across several tasks.
Davit: Dual attention vision transformers
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
-
Rectifying Magnitude Neglect in Linear Attention
MALA adds a scale-sensitive offset to linear attention so attention sharpens with query magnitude, improving accuracy across several tasks.