From 3D data to t-SNE embedding

24 observations · 3 groups · interactive dimensionality reduction
Original 3D data
Group AGroup BGroup CDrag to rotate 3D · Scroll to zoom
Start with the original 3D observations. Select Projection to reveal a randomly initialized t-SNE embedding on the floor (z=0).
The floor points start at random 2D coordinates. t-SNE uses pairwise Gaussian affinities in 3D and Student-t affinities in 2D, optimized by gradient descent with momentum. The learned embedding is not constrained to the original x–y coordinates.