Structure of a variational autoencoder with a 4-3-2-3-4 architecture (similar structure to the autoencoder in Figure 1). The dashed lines indicate that each latent value ﹩{z}_{\mathrm{1,2}}﹩ is drawn from a Gaussian distribution defined by a latent mean μ1,2 and log variance ﹩\mathrm{ln}{\sigma }_{1,2}^{2}﹩.