Fig. 10

Graphical model of our saVAE. All plain arrows are multi-layered perceptrons, except for the curved one. The curved arrow reflects the similarities among positive samples in the latent space. The dotted arrows represent the process of making the paired samples according to the precalculated similarity table of the observed data. The gray circles are observable; the white ones are not. The variable z is computed using the reparametrization trick [9], that is, \(z = \varvec{\mu } + \varvec{\epsilon } \odot \varvec{\sigma }\), where \(\varvec{\epsilon } \sim N(\varvec{0},I)\). Here, h is a hidden variable, and s is an optional input. See Methods section