Allowing negative step sizes in a Wolfe line search turns ascent directions into usable descent steps, giving SR1 quasi-Newton methods competitive neural-network training performance.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Don't Be So Positive: Negative Step Sizes in Second-Order Methods
Allowing negative step sizes in a Wolfe line search turns ascent directions into usable descent steps, giving SR1 quasi-Newton methods competitive neural-network training performance.