AWhat changed
Approximated matrix-valued second-moment accumulators from row and column statistics and combined update clipping, decay scheduling and relative-step scaling for low-memory adaptive optimization.
Scientific scope
Approximated matrix-valued second-moment accumulators from row and column statistics and combined update clipping, decay scheduling and relative-step scaling for low-memory adaptive optimization. The O(n+m) auxiliary-state saving refers to a matrix's factored second moment, not all training memory. Translation experiments do not guarantee Adam-equivalent behavior for arbitrary models; no dense nonconvex convergence proof is claimed.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. The O(n+m) auxiliary-state saving refers to a matrix's factored second moment, not all training memory. Translation experiments do not guarantee Adam-equivalent behavior for arbitrary models; no dense nonconvex convergence proof is claimed.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- Adafactor: Adaptive Learning Rates with Sublinear Memory Cost
- Shazeer and Stern: Adafactor factored optimizer state — author/proceedings record
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction