Experiments / live

Gradient descent playground

Left: a 1D loss curve — click anywhere to drop a starting point, then hit step to apply θ ← θ − η·L′(θ) and watch the iterates march toward the minimum. Nudge η up and the steps overshoot; push it past the threshold and they oscillate and blow up. Right: the same descent on a 2D surface, the gold trajectory rolling from a fixed start toward the white minimum. Swap the surface between a friendly bowl, a nasty ravine, and a saddle to see where rolling downhill gets hard. The full story — the update rule, why the learning rate makes or breaks it, and convex vs non-convex landscapes — is in the essay.

iterate 0 — click the left plot to set θ₀