Convergence Analysis in Hyperparameter Optimization (2D)
2D companion: a top-down heatmap of the same non-convex loss landscape as the 3D lab, with an optimizer trajectory, a live loss-vs-iteration chart, and a gradient-norm chart showing exactly when the stopping criterion ε is met, across SGD, momentum and Adam.
The 3D original orbits a camera around a purple loss surface while an optimizer marker descends it. This 2D companion drives the exact same landscape function and the exact same SGD / momentum / Adam update rules, but reads them from directly above as a heatmap instead — which makes the basin structure and the shallow local-minima wells easier to compare at a glance — and adds a second chart the 3D lab never had: gradient norm plotted against iteration with the ε stopping line drawn where it is actually evaluated, since ε is a threshold on the gradient norm, not on the loss. A "Steps to converge" readout turns that into a concrete number you can compare across optimizers and noise levels.
Same non-convex loss surface as the 3D lab (one quadratic basin plus six Gaussian bumps forming shallow local minima), same SGD / SGD+Momentum / Adam update equations. The heatmap is precomputed once into an offscreen canvas since the landscape is static, then redrawn scaled every frame under the live trajectory. Convergence is declared once the true (noise-free) gradient norm stays below 10^ε for 8 consecutive optimizer steps, and the step count at which that happens is reported directly.
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install