Unlabeled Disentangling of GANs with Guided Siamese Networks

Disentangling underlying generative factors of a data distribution is important for interpretability and generalizable representations. In this paper, we introduce two novel disentangling methods. Our first method, Unlabeled Disentangling GAN (UD-GAN, unsupervised), decomposes the latent noise by generating similar/dissimilar image pairs and it learns a distance metric on these pairs with siamese networks and a contrastive loss. This pairwise approach provides consistent representations for similar data points. Our second method (UD-GAN-G, weakly supervised) modifies the UD-GAN with user-defined guidance functions, which restrict the information that goes into the siamese networks. This constraint helps UD-GAN-G to focus on the desired semantic variations in the data. We show that both our methods outperform existing unsupervised approaches in quantitative metrics that measure semantic accuracy of the learned representations. In addition, we illustrate that simple guidance functions we use in UD-GAN-G allow us to directly capture the desired variations in the data.