Learning Latent Subspaces in Variational Autoencoders

Such representations are often difficult to interpret or control. We consider the problem of unsupervised learning of features correlated to specific labels in a dataset.