Learning Latent Subspaces in Variational Autoencoders