vtsne theoretical basis
bhomass opened this issue · comments
It seems vtsne is just tsne with an additional loss derived from reparameterization, something reminiscent of variational auto encoding. is there any paper explaining the theoretical basis?
I think I get the point. You like to apply a 2d isotropic gaussian as the prior for p_ij. Therefore, in addition to minimizing the KLD between the similarity based pdf of the final and original vector, you like the final tsne distribution to form k gaussian like clusters. Intuitively, that would be nice. Its always easier to focus on clustered points to try to find their latent characteristics. We all know that the original tsne can sometimes give misleading formations. It would be nice to apply this technique to some of the tough cases to verify that it leads to more accurate interpretations.