Optimizers such as Adam, Shampoo, and SOAP are unified as structured Fisher approximations, and two new derived optimizers, RACS and Alice, achieve faster LLaMA pre-training than Adam at lower memory.
Effects of last layer One crucial setup difference during evaluation for low-rank methods is whether the last layer is trained by full-rank Adam or not
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank Extension
Optimizers such as Adam, Shampoo, and SOAP are unified as structured Fisher approximations, and two new derived optimizers, RACS and Alice, achieve faster LLaMA pre-training than Adam at lower memory.