Disentangling latent factors of variation for visual data

Abstract

Disentangling higher level generative factors as disjoint latent dimensions offer several benefits such as ease of deriving invariant representations, targeted data augmentation with style transfer, better interpretability of the data, etc. In this work, we focus on disentangling factors of variation with weak-supervision (in the form of pair-wise similarity labels) using a nonadversarial approach. We show compelling results for both the quality of disentangled representations and image generation for MNIST and CMU MultiPIE datasets and UTK-face and CelebA datasets for cross-dataset evaluation. We further demonstrate few-shot learning of new previously-unseen classes as a consequence of e ective disentangling of the latent subspace (into style and class).