Distributed Training · 2025
Modular Manifolds
Constrains weight matrices to submanifolds (e.g. the Stiefel manifold of unit-singular-value matrices) and co-designs a 'manifold Muon' optimizer that steps in the tangent space, giving principled per-layer learning-rate budgeting.
Editorial record
Plain-language summary
Constrains weight matrices to submanifolds (e.g. the Stiefel manifold of unit-singular-value matrices) and co-designs a 'manifold Muon' optimizer that steps in the tangent space, giving principled per-layer learning-rate budgeting.
Knowledge graph
Relationships
Descendants
Conceptual ancestorEvidence: Strongly supported
Adam: A Method for Stochastic Optimization
Extends the optimizer lineage with manifold-constrained training
TML
Source record
Provenance
- Record ID
- P-656
- Record created
- 2026-07-13
- Last reviewed
- 2026-07-14
- Record version
- 2
Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.