Muon optimizer
Algorithms AIMech:Muon DRAFT
Aliases: Muon
A neural-network optimization method that approximately orthogonalizes momentum-based matrix updates using Newton–Schulz iterations.
Identity Notes & Concept Scope
Muon is a matrix-update optimizer distinct from coordinatewise Adam-family scaling. The name here denotes the optimization algorithm, not unrelated uses of the word muon.
Claims & Grounded Evidence
1 Claim
Mechanism
#update-rule
The inspected implementation forms a momentum update, optionally uses a Nesterov combination, applies iterative matrix orthogonalization, and scales the resulting update for the parameter shape while applying decoupled weight decay.
⚠️ Curated Limitations & Negative Results
- The cited implementation is scoped to two-dimensional hidden-layer parameters and directs other parameters to a standard optimizer.
- The implementation provides alternative shape-scaling conventions; no universal exact orthogonality, convergence or performance guarantee is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:TrainingStrategy
TrainingStrategy
|
related_match | This algorithm can implement parameter fitting, but the pinned TrainingStrategy definition uses a preprocessing genus. This is a contextual training relationship, not equivalence or a subclass assertion. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
pytorch-docs
|
PyTorch contributors
|
official documentation
documentation
|
sha256:abaa98c693e0…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T03:25:18Z
Updated: 2026-10-11T03:25:18Z
Source File:
data/curated/Muon.yaml