← ~/visualizations

dimensionality-reduction

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.

canvasclick to interact
t=0s

practical uses

  • 01.Compressing features before training models (faster, less overfitting)
  • 02.2D/3D visualization of high-dimensional datasets for exploration
  • 03.Manifold learning/representation learning for downstream tasks (clustering, retrieval)

technical notes

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.