Second-Order Methods

How models actually learn, from vanilla gradient descent to Adam

First-order methods use gradients. Second-order methods also use curvature, usually through the Hessian. Curvature tells the optimizer how the gradient itself changes as parameters move.

Newton's method uses that curvature to choose a step that can jump directly to the minimum of a quadratic. The price is that Hessians are huge in modern neural networks.

A crane operator uses a load chart because direction is not enough. The load also bends the boom, and that bend changes what move is safe. Second-order optimization reads the bend, not just the pull, before deciding how far to move. In the figure you play the operator: slide the two curvatures and watch the surface become a bowl, a dome, or a saddle. The Hessian's eigenvalues are exactly those two knobs.

Where this lives in MLLarge neural networks usually rely on first-order optimizers because gradients are cheap through backpropagation, while full Hessians are not. Second-order ideas still influence preconditioning, K-FAC, Shampoo, L-BFGS, and optimizer research.
▶ Second-Order Methods
← Regularization as GeometryThe Loss Landscape →