Data generalization for new classes with a single instance via automatic style labeling and transfer

Abstract

In order to synthesize new images from a specific class, most generative models like Generative Adversarial Nets (GANs) require a large amount of data from this class. In other words, modern generative models often lack the ability to create new samples belonging to an unseen class from which they have observed only one instance. In this thesis, we propose a model that can generalize a single instance from an unseen class and create a whole data distribution of the class by learning how data from other classes vary within their own distributions, and transferring this information to the new class. We show that the new samples generated by our model not only preserve the essential visual features for them to be recognized as in the same class that the source instance is from, but also exhibit variety. Experiments on the MNIST dataset show that after hiding away one class of digits and training only on the data of the remaining nine classes, our model can successfully generate new images of the hidden class with controllable features, given just a single image from that class.