Abstract
We propose a novel unsupervised generative model, Elastic-InfoGAN, thatlearns to disentangle object identity from other low-level aspects inclass-imbalanced datasets. We first investigate the issues surrounding theassumptions about uniformity made by InfoGAN, and demonstrate itsineffectiveness to properly disentangle object identity in imbalanced data. Ourkey idea is to make the discovery of the discrete latent factor of variationinvariant to identity-preserving transformations in real images, and use thatas the signal to learn the latent distribution's parameters. Experiments onboth artificial (MNIST) and real-world (YouTube-Faces) datasets demonstrate theeffectiveness of our approach in imbalanced data by: (i) better disentanglementof object identity as a latent factor of variation; and (ii) betterapproximation of class imbalance in the data, as reflected in the learnedparameters of the latent distribution.