LaRVAE: Label Replacement VAE for Semi-Supervised Disentanglement Learning

Learning interpretable and disentangled representations is a crucial yet challenging task in representation learning. In this work, we develop LaRVAE, a novel semi-supervised variational auto-encoder (VAE) for learning disentangled representations by more effectively exploiting the limited labeled data. Specifically, during training, we replace the inferred representation associated with a data point with its ground-truth representation whenever it is available. LaRVAE is theoretically inspired by our proposed general framework of semi-supervised disentanglement learning in the context of VAEs. It naturally motivates both the novel representation replacement strategy in LaRVAE and a supervised regularization term commonly (albeit in an ad-hoc way) used in existing semi-supervised VAEs. Extensive experiments on synthetic and real datasets demonstrate both quantitatively and qualitatively the ability of LaRVAE to significantly and consistently improve disentanglement with very limited supervision.