t-SNE Perplexity Lab (2D)

This 2D companion runs the exact same live t-SNE gradient descent as the 3D original — real per-point Gaussian affinities in 5D solved for a target perplexity, a Student-t distribution in the 2D map, and honest gradient-descent updates every frame — but draws it straight onto the map plane instead of a rotating 3D scene. Two things a camera makes hard to read become directly visible: a neighbor-affinity graph draws a line between any two points whose symmetrized high-dimensional probability clears a threshold, so you can see exactly which pairs the optimiser is trying to pull together, and a live KL-divergence sparkline tracks convergence speed as you change perplexity, cluster separation, cluster count or the learning rate.