Unsupervised Latent Regression through Information Maximization - Contrastive Regularized GAN

Most of labelled image public datasets are discrete in nature, e.g. cat vs dog, human, car, etc. With the growing complexity of tasks, fine-grained label data is needed. Fine-grained labels are costly because there is a need for experts to label them. Generative Adversarial Network (GAN) has been gaining a lot of attentions due to its ability to not only generate realistic images but also to disentangle the attributes of the images. Unfortunately, GAN’s disentanglement methods usually lack the ability to quantify such attributes. The objective of this work is to quantify the attributes of the target (object/image) based on the disentangled properties without supervision. In order to get the (disentangled) attributes, we leverage GAN with information maximization and contrastive regularizer. Regression is done by adding additional layer to the contrastive networks of the model. The regression quality of the proposed method is quantified by order quality measured using normalized Kendall’s Tau. Furthermore, an application in denoising image is also presented.