Isomap

Sampling needed (not so for SVD).
Which scales better?
But, how can we find a *propositional* representation?
 person: Tom
 glasses: T
 direction: left
LLE: can be more robust
isomap: natural dimension discovery