New dimensionality reduction algorithm (UMAP?)

This came up in the discussion around transformpoint and fluid.mds~, but figured it would be good to make a thread about it.

Basically, it would be good to have a non-autoencoder-based alternative to fluid.pca~ for being able to transform a single datapoint on a precomputed fit.

On top of this, it would be nice to have something that takes our PCA or UMAP and turns it into an equal-shaped field, like the Poisson Disc:

http://extremelearning.com.au/an-improved-version-of-bridsons-algorithm-n-for-poisson-disc-sampling/

1 Like

Totally.

Having a way to remap (via equally spacing and/or re-ordering via user-specified associated descriptors (i.e. loudness = y axis, centroid = x axis)) the space after a reduction would be fantastic too.