Momentum Methods: Second-Order ODE
An interactive live film: Nesterov acceleration on the narrow quadratic valley; inertia shoots past the floor while 3/t damping carries the particle rapidly down the flat axis to achieve O(1/t^2) convergence.
Narration
Gradient descent and gradient flow follow the steepest slope. In an elongated valley, they crawl along the floor and waste steps oscillating against steep walls.
Momentum changes the game. Adding inertia turns first-order gradient flow into a second-order differential equation, like a heavy ball rolling in a potential.
The rest of the narration is in the book's Companion.
More from Chapter 20
Interactive1:45Flow Matching: Velocity Fields and the Continuity EquationChapter 20
Preview1:28From a Residual Network to a Neural ODEChapter 20
Preview1:27Gradient Flow: Gradient Descent With Infinitely Small StepsChapter 20
Preview1:26Memory by Orthogonal Projection (HiPPO)Chapter 20
Preview1:34Probability Flow ODE: Deterministic Paths, Identical MarginalsChapter 20
Preview1:40Backward Error Analysis: What Gradient Descent FollowsChapter 20
Preview1:46Rectified Flow: Straightening by RetrainingChapter 20