RACE Attention is a strictly linear-time attention mechanism that approximates softmax attention outputs using Gaussian projections and soft LSH to enable training on contexts up to 12 million tokens.
21 Proof.(1) Row-sum control.Since ˆS=S+ ∆and ˆD= diag( ˆS1), we rewrite E:= ˆD−D= diag ( ( ˆS−S)1 ) = diag(∆1)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
RACE Attention: A Strictly Linear-Time Attention Layer for Training on Outrageously Large Contexts
RACE Attention is a strictly linear-time attention mechanism that approximates softmax attention outputs using Gaussian projections and soft LSH to enable training on contexts up to 12 million tokens.