Shows a set of points in a simulated high-dimensional space (left) being mapped by f: R^D → R^d into a low-dimensional embedding (right). The animation cycles through different preservation criteria (variance, pairwise distances, neighborhood structure) and highlights how choosing a linear vs nonlinear mapping changes what structure can be preserved under the constraint d << D.
Self-contained canvas 2D rendering with a deterministic synthetic manifold dataset. The mapping f blends between a linear projection and a nonlinear ‘unfolding’ warp. The visualization cycles criteria every ~3.6s, highlights a probe point and neighbors, and displays a simple criterion-specific score. Uses grid-snapped drawing for a blocky green-on-black aesthetic and time-based easing for smooth motion.