Variational autoencoders (VAEs) are widely used deep generative models
capable of learning unsupervised latent representations of data. Such
representations are often difficult to interpret or control. We consider the
problem of unsupervised learning of features correlated to specific