Generative Adversarial Structured Networks

We propose a technique that combines generative adversarial networks with probabilistic graphical models to explicitly model dependencies in structured distributions. Generative adversarial structured networks (GASNs) produce samples by passing random inputs through a neural network to construct the potentials of a graphical model; maximum a-posteriori inference in this graphical model then yields a sample. To train a GASN, one must differentiate a bi-level optimization, which is non-trivial. We present a solution based on “smoothing” the generator, and propose two methods for obtaining the smoothed gradient. We show preliminary experimental results, demonstrating that training GASNs is feasible.