Original Data (High-D Space)
UMAP Embedding (2D)
ψ(x) = 1 / (1 + a·x²ᵇ)
Smooth approximation to fuzzy membership
Cross-Entropy Loss
k-NN Graph
vⱼ|ᵢ = exp(-(d(xᵢ,xⱼ) - ρᵢ) / σᵢ)
Local connectivity with adaptive σᵢ
UMAP Parameters
15
n_neighbors
0.1
min_dist
n_neighbors: local vs global structure
min_dist: how tightly points cluster
Algorithm
0
Epoch
Learning Rate
Statistics
Points
0
Loss
--
Clusters
0
Edges
0
Quick Examples
How it works: UMAP constructs a fuzzy topological representation using k-NN graphs, then optimizes a low-D embedding to match this structure. Unlike t-SNE, it better preserves global structure and is generally faster.