Spectral maps for slow collective variables

They did something that I always wanted to do which is to put a neural network between the data and the variables from which you construct a (diffusion kernel) transfer operator.  You can then optimize the encoding to produce the largest possible timescale separation.