The paper derives a Riemannian update rule for the spectral factors of a positive-definite preconditioner, making arbitrary matrix roots fast and numerically stable for low-precision NN training.
Learning rate grafting: Transferability of optimizer tuning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
stat.ML 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Spectral-factorized Positive-definite Curvature Learning for NN Training
The paper derives a Riemannian update rule for the spectral factors of a positive-definite preconditioner, making arbitrary matrix roots fast and numerically stable for low-precision NN training.