Inverse UMAP transforms
TIL that the dimensionality reduction algorithm UMAP1 has support for inverse transforms!
Naturally, these are lossy inverses, but being able to generate novel examples of, for example, handwritten digits (yep, classic MNIST once again) is exceedingly cool.
The example images of handwritten digits here are sampled from the compressed planar space, and “inverted” into the original “image space”.
On a related note, the whole explanatory article for UMAP is a beautiful work of exposition, full of rich ideas that make one ponder deeply. I only wish I understood everything in here end-to-end.