Rotating gradients into their SVD coordinate system before Adam-style updates can roughly halve the number of steps LLaMA models need to reach a given perplexity.
Practical gauss-newton optimisation for deep learning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Improving Adaptive Moment Optimization via Preconditioner Diagonalization
Rotating gradients into their SVD coordinate system before Adam-style updates can roughly halve the number of steps LLaMA models need to reach a given perplexity.