EngineersofAI LogoCoursesBlogResearch LabAI LettersThe LabCode BankInteractive 3DKodrEarnestJobs
Skip to main content
Interactive 3D/t-SNE & UMAP Dimensionality Reduction
Class A
Class B
Class C
Class D
Class E
TSNE
t-SNE / UMAP
Dimensionality reduction to 2D. Colors = classes.
Dataset
Parameters
Perplexity30
5 (tight)50 (loose)
Iterations200
Learning rate200
Actions
HUD Metrics
Iteration
0 / 200
Stress
1.0000
Clusters preserved
0 / 5
How It Works
t-SNE maps high-dimensional points to 2D by preserving local neighborhood structure. Similar points are pulled together; dissimilar ones are pushed apart.

Perplexity controls how many neighbors each point considers - lower = tighter, isolated clusters.

UMAP uses a different graph-based objective and often preserves global structure better.

t-SNE & UMAP Dimensionality Reduction - Interactive Visualization

t-SNE and UMAP are nonlinear dimensionality reduction methods that project high-dimensional data into 2D for visualization. t-SNE minimizes KL divergence between pairwise similarity distributions in high-D and low-D space - it excels at revealing cluster structure but distorts global distances. UMAP is faster, preserves more global structure, and scales to millions of points while producing similarly tight cluster separations.

  • See the 2D projection update as you change t-SNE perplexity - low perplexity reveals fine local structure, high perplexity shows global layout
  • Compare t-SNE vs UMAP on the same dataset - UMAP runs 10-100x faster and preserves inter-cluster distances better
  • Understand why t-SNE cluster sizes and distances are NOT meaningful - only topology (which groups are separate) can be interpreted
  • Adjust UMAP n_neighbors: small values focus on local structure, large values preserve global topology
  • Learn the crowding problem t-SNE solves: using t-distribution (heavy tails) in low-D space to prevent cluster collapse

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.