Customization of latent space in semi-supervised Variational AutoEncoder

SeungHwan An, Jong-June Jeon

Abstract: Highlights•The latent distribution can be customized by the label-specific conceptual means and variances.•The interpolated images corresponding to the user’s specific purpose can be achieved by customization.•The variability of generated samples depends on a specific subspace, called active latent subspace.•Through our proposed V-nat measure, the active latent subspace can be detected.•The features of images can be analyzed in association with the active latent subspace.