Weakly Supervised Disentanglement with Triplet Network

Variational Autoencoders have gained considerable attention due to their capacity of encoding high dimensional data into a lower dimensional latent space. In this context, several methods have been proposed with the objective of producing disentangled representations. In this work, we propose a weakly supervised model that explicitly disentangles the factors of variation of a dataset in separate subspaces using a pairwise architecture. We also create a framework that encourages conditional image generation according to the desired factor of variation, by controlling these subspaces. This is achieved by introducing an additional network trained with a triplet loss. Its output approximates representations of images generated from the same factor and push the ones of images generated from different factors apart. Experiments are carried out on widely used datasets, and show that our model is able to disentangle specified factors of variation, and to generate new data while constraining desired properties, even when these factors have small influence on reconstruction loss.