Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: He, Chuan, Deng, Zhanwang, Lu, Zhaosong
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917419003412480
author He, Chuan
Deng, Zhanwang
Lu, Zhaosong
author_facet He, Chuan
Deng, Zhanwang
Lu, Zhaosong
contents Neural network (NN) training is inherently a large-scale matrix optimization problem, yet the matrix structure of NN parameters has long been overlooked. Recently, the optimizer Muon \citep{jordanmuon}, which explicitly exploits this structure, has gained significant attention for its strong performance in foundation model training. A key component contributing to Muon's success is matrix orthogonalization. In this paper, we propose \textit{low-rank orthogonalization}, which performs orthogonalization by leveraging the low-rank nature of gradients during NN training. Building on this, we introduce low-rank matrix-signed gradient descent (MSGD) and a low-rank variant of Muon. Numerical experiments demonstrate the superior performance of low-rank orthogonalization, with low-rank Muon achieving promising results in GPT-2 and LLaMA pretraining -- surpassing the carefully tuned vanilla Muon on tasks with large model sizes. Theoretically, we establish the iteration complexity of low-rank MSGD for finding an approximate stationary solution, and the iteration complexity of low-rank Muon for finding an approximate stochastic stationary solution under heavy-tailed noise. The code to reproduce our numerical experiments is available at https://github.com/dengzhanwang/Low-rank-Muon.
format Preprint
id arxiv_https___arxiv_org_abs_2509_11983
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training
He, Chuan
Deng, Zhanwang
Lu, Zhaosong
Machine Learning
Optimization and Control
49M37, 90C30, 90C90
Neural network (NN) training is inherently a large-scale matrix optimization problem, yet the matrix structure of NN parameters has long been overlooked. Recently, the optimizer Muon \citep{jordanmuon}, which explicitly exploits this structure, has gained significant attention for its strong performance in foundation model training. A key component contributing to Muon's success is matrix orthogonalization. In this paper, we propose \textit{low-rank orthogonalization}, which performs orthogonalization by leveraging the low-rank nature of gradients during NN training. Building on this, we introduce low-rank matrix-signed gradient descent (MSGD) and a low-rank variant of Muon. Numerical experiments demonstrate the superior performance of low-rank orthogonalization, with low-rank Muon achieving promising results in GPT-2 and LLaMA pretraining -- surpassing the carefully tuned vanilla Muon on tasks with large model sizes. Theoretically, we establish the iteration complexity of low-rank MSGD for finding an approximate stationary solution, and the iteration complexity of low-rank Muon for finding an approximate stochastic stationary solution under heavy-tailed noise. The code to reproduce our numerical experiments is available at https://github.com/dengzhanwang/Low-rank-Muon.
title Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training
topic Machine Learning
Optimization and Control
49M37, 90C30, 90C90
url https://arxiv.org/abs/2509.11983