Layerwise LQR

Uses

How it works

Evidence quality

Highlighted claims

• LLQR rewrites the exact dense geometry-aware update as a layerwise Linear Quadratic Regulator problem before imposing structure on an inverse preconditioner. — Layerwise LQR for Geometry-Aware Optimization of Deep Networks
• The exact LLQR steepest-descent update is equivalent to a finite-horizon LQR problem under linearized layer dynamics. — Layerwise LQR for Geometry-Aware Optimization of Deep Networks
• The exact Riccati formulation is used mainly as a reference rather than as a practical algorithm for modern networks. — Layerwise LQR for Geometry-Aware Optimization of Deep Networks
• LLQR separates network dynamics from descent geometry in dense curvature matrices. — Layerwise LQR for Geometry-Aware Optimization of Deep Networks
• LLQR is positioned as both a practical optimizer wrapper and a framework for studying scalable approximations to dense geometry-aware updates. — Layerwise LQR for Geometry-Aware Optimization of Deep Networks