Cascading Generative Adversarial Networks for Targeted Imagination


Abundance of labelled data played a crucial role in the recent developments in computer vision, but that faces problems like scalability and transferability to the wild. One alternative approach is to utilize the data without labels, i.e. unsupervised learning, in learning valuable information and put it in use to tackle vision problems. Generative Adversarial Networks (GANs) have gained momentum for their ability to model image distributions in unsupervised manner. They learn to emulate the training set and that enables sampling from that domain and using the knowledge learned for useful applications. Several methods proposed enhancing GANs, including regularizing the loss with some feature matching. We seek to push GANs beyond the data in the training and try to explore unseen territory in the image manifold. We first propose a new regularizer for GAN based on K-Nearest Neighbor (K-NN) selective feature matching to a target set Y in high-level feature space, during the adversarial training of GAN on the base set X, and we call this novel model K-GAN. We show that minimizing the added term follows from cross-entropy minimization between the distributions of GAN and set Y. Then, we introduce a cascaded framework for GANs that try to address the task of imagining a new distribution that combines the base set X and target set Y by cascading sampling GANs with translation GANs, and we dub the cascade of such GANs as the Imaginative Adversarial Network (IAN). Several cascades are trained on a collected dataset Zoo-Faces and generated innovative samples are shown, including from K-GAN cascade. We conduct an objective and subjective evaluation for different IAN setups in the addressed task of generating innovative samples and we show the effect of regularizing GAN on different scores. We conclude with some useful applications for these IANs, like multi-domain manifold traversing.

[MS Thesis]