trends.barnas.net

method · first seen 8 Oct 2026

muP

Also: μP, maximal update parametrization, maximal update parameterization, feature learning in neural networks

A parametrization scheme that allows hyperparameters to transfer across model width so optimal settings remain stable at scale.

Mentions per week

0319 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

hyperparameter transferLaplace approximationelastic pretrainingDP-SGDresidual neural networkdepth-width scaling