Implicit Modeling - A Generalization of Discriminative and Generative Approaches

We propose a new modeling approach that is a generalization of gen-erative and discriminative models. The core idea is to use an implicitparameterization of a joint probability distribution by specifying only theconditional distributions. The proposed scheme combines the advantagesof both worlds – it can use powerful complex discriminative models asits parts, having at the same time better generalization capabilities. Wethoroughly evaluate the proposed method for a simple classification taskwith artificial data and illustrate its advantages for real-word scenarios ona semantic image segmentation problem.