UniForm shares one attention matrix across all attention heads and adds multi-scale depthwise convolutions to value projections, claiming faster edge inference at comparable accuracy.
Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2024 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
UniForm: A Reuse Attention Mechanism Optimized for Efficient Vision Transformers on Edge Devices
UniForm shares one attention matrix across all attention heads and adds multi-scale depthwise convolutions to value projections, claiming faster edge inference at comparable accuracy.