1 paper · 1 filter
Florian Hübler, Thomas Pethick, Suvrit Sra
Non-Euclidean optimisation methods with matrix-valued updates, such as Muon and Scion, have recently shown strong empirical performance for training Transformer models, yet their t…