Generative Entity Networks: Disentangling Entitites and Attributes in Visual Scenes using Partial Natural Language Descriptions

Generative image models have made significant progress in the last few years, and are now able to generate low-resolution images which sometimes look realistic. However the state-of-the-art models utilize fully entangled latent representations where small changes to a single neuron can effect every output pixel in relatively arbitrary ways, and different neurons have possibly arbitrary relationships with each other. This limits the ability of such models to generalize to new combinations or orientations of objects as well as their ability to connect with more structured representations such as natural language, without explicit strong supervision. In this work explore the synergistic effect of using partial natural language scene descriptions to help disentangle the latent entities visible an image. We present a novel neural network architecture called Generative Entity Networks, which jointly generates both the natural language descriptions and the images from a set of latent entities. Our model is based on the variational autoencoder framework and makes use of visual attention to identify and characterise the visual attributes of each entity. Using the Shapeworld dataset, we show that our representation both enables a better generative model of images, leading to higher quality image samples, as well as creating more semantically useful representations that improve performance over purely dicriminative models on a simple natural language yes/no question answering task.