Gradient Descent

A model learns by turning error into direction. Follow one point down a computed loss landscape, then change the learning rate until the same rule becomes cautious, fast, unstable, or divergent.

1 exhibit Computed toy landscape Offline, single file

StudioStart hereComputed
Gradient Descent Studio

Switch between a dimensional loss surface and a top-down contour map. Step through the update, inspect the gradient and parameter delta, then compare stable learning with overshoot and divergence.
Single screen · Canvas + MathML→

Gradient Descent

Gradient Descent Studio