Muon optimizer's grokking speed attributed to orthogonalization, study finds

A new research paper published on arXiv details the active ingredient behind the Muon optimizer's faster grokking threshold in modular arithmetic compared to AdamW. The study isolates orthogonalization, specifically the Newton-Schulz iteration, as the key mechanism responsible for this speedup. Reducing the Newton-Schulz iteration count can accelerate grokking but leads to a fragile solution, with five iterations proving to be the most robust choice across various learning rates. AI

IMPACT Identifies a key mechanism for faster model generalization, potentially informing future optimizer development.

RANK_REASON The cluster contains a research paper detailing a novel finding about an optimization algorithm. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AdamW
• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• Muon
• Newton-Schulz iteration
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →