The network maps two inputs (x₁, x₂) to a single output — its prediction is now a surface over a 2D input plane. Drag to rotate. Hit Train to watch a flat surface morph to fit the noisy sin(x₁)·cos(x₂) target.