Landmark-Based Spectral Embedding

[joshpoll]

The spectral embedding step of UMAP is really really slow. I experimented with replacing UMAP's typical spectral embedding step with the Variational Nystrom approximation: http://proceedings.mlr.press/v48/vladymyrov16.pdf. My results show that (at least on MNIST) this can speed up the embedding step by 10x while not affecting the final embedding quality very much. I was wondering if it might make sense to upstream this into UMAP.jl? One of the cons is that it adds another hyper-parameter for the number of landmarks.

I'm happy to provide more details and my implementation (the core algorithm is really small), if people are interested in this extension.