Generalized Adversarially Learned Inference

Allowing effective inference of latent vectors while training GANs can greatly increase their applicability in various downstream tasks. Recent approaches, such as ALI and BiGAN frameworks, develop methods variables by adversarially an image generator along with encoder to match two joint distributions vector pairs. We generalize these approaches incorporate multiple layers feedback on reconstructions, self-supervision, other forms supervision based prior or learned knowledge about the desired solutions. achieve this modifying discriminator's objective correctly identify more than tuples arbitrary number random consisting images, vectors, generated through auxiliary tasks, reconstruction inpainting outputs suitable pre-trained models. design a non-saturating maximization for generator-encoder pair prove that resulting adversarial game corresponds global optimum simultaneously matches all distributions. Within our proposed framework, we introduce novel set techniques providing self-supervised model properties, patch-level correspondence cycle consistency reconstructions. Through comprehensive experiments, demonstrate efficacy, scalability, flexibility approach variety The appendix paper be found at following link: https://drive.google.com/file/d/1i99e682CqYWMEDXlnqkqrctGLVA9viiz/view?usp=sharing