🗺️ Dataset

Blobs (5D)

Two moons (HD)

Concentric

Random

Hyperparameters

Perplexity: 30 Learning rate: 120 Points: 200

Stats

Iteration0

KL divergence—

Perplexity30

n points200

Phaseexaggeration

t-SNE converts high-D distances into Gaussian conditional probabilities p_j|i (σ tuned so each point's perplexity matches the slider), symmetrises them to p_ij, and matches them in 2D with a heavy-tailed Student-t kernel q_ij ∝ 1/(1+‖y_i−y_j‖²). Gradient descent with momentum minimises KL(P‖Q), pulling similar points together and pushing dissimilar ones apart. Early "exaggeration" inflates P to form tight clusters first.