Kernel Density Estimate: Bandwidth, Bias-Variance Tradeoff and Exact MISE
Drop data marbles on a board, put a Gaussian bump on each, and watch the sum become a density estimate. Slide the bandwidth between spiky and oversmoothed, compare the Silverman rule, cross-validation and the exact MISE minimiser, and see the closed-form bias and variance curves. A Blender-built board, marbles and ruler, with numerical checks of the statistics built in.
UMAP reduces high-dimensional data to a handful of dimensions by optimizing a low-dimensional layout so that its k-nearest-neighbor graph structure is preserved — using an attractive force along graph edges and an inverse-square repulsive force between every pair of points, exactly the two-force tug-of-war visualized here in real time. Four synthetic clusters start as random noise in 3D and are pulled and pushed into shape epoch by epoch; tune n_neighbors, min_dist, repulsion strength and optimizer speed while live readouts track layout stress and how much of the original neighbor structure survives the embedding.
Drop data marbles on a board, put a Gaussian bump on each, and watch the sum become a density estimate. Slide the bandwidth between spiky and oversmoothed, compare the Silverman rule, cross-validation and the exact MISE minimiser, and see the closed-form bias and variance curves. A Blender-built board, marbles and ruler, with numerical checks of the statistics built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install