t-SNE Explained: the Math Behind Preserving Local Neighborhoods in High-Dimensional Visualization

burny_tech · x · 2026-09-19

A refresher on t-SNE's core math: conditional similarities in high-dimensional space are computed with Gaussian kernels, while low-dimensional similarities use a heavy-tailed Student-t distribution. The embedding is found by minimizing KL divergence via gradient descent — the heavy tails help separate moderately distant points while keeping neighbors together, making it ideal for 2D/3D visualization.

Original post →

More from Research

Research channel →