[ ] Using Gaussian process: having the embedding result with some 'anchor' perplexities, can we use GP to predict the distribution of the function that describe the 'constraint-preserving score' w.r.t prep
[ ] Using constraints to force sigma of each Gaussian at each data point x_i in high dim. to be large enough to "cover" x_j if (i, j) is must-link.
Some initial ideas:
prepsigmaof each Gaussian at each data pointx_iin high dim. to be large enough to "cover"x_jif(i, j)is must-link.