Unsupervised Discovery and Manipulation of Continuous Disentangled Factors of Variation

Abstract

Learning a disentangled representation of a distribution in a completely unsupervised way is a challenging task that has drawn attention recently. In particular, much focus has been put in separating factors of variation (i.e., attributes) within the latent code of a Generative Adversarial Network (GAN). Achieving that permits control of the presence or absence of those factors in the generated samples by simply editing a small portion of the latent code. Nevertheless, existing methods that perform very well in a noise-to-image setting often fail when dealing with a real data distribution, i.e., when the discovered attributes need to be applied to real images. However, some methods are able to extract and apply a style to a sample but struggle to maintain its content and identity, while others are not able to locally apply attributes and end up achieving only a global manipulation of the original image. In this article, we propose a completely (i.e., truly ) unsupervised method that is able to extract a disentangled set of attributes from a data distribution and apply them to new samples from the same distribution by preserving their content. This is achieved by using an image-to-image GAN that maps an image and a random set of continuous attributes to a new image that includes those attributes. Indeed, these attributes are initially unknown and they are discovered during training by maximizing the mutual information between the generated samples and the attributes’ vector. Finally, the obtained disentangled set of continuous attributes can be used to freely manipulate the input samples. We prove the effectiveness of our method over a series of datasets and show its application on various tasks, such as attribute editing, data augmentation, and style transfer.