Bayesian Reasoning with Trained Neural Networks

We use independently trained neural networks to represent abstract concepts and combine them through Bayesian reasoning to approach tasks outside their initial scope. Prior knowledge is provided by deep generative models and classification or regression networks are used to express knowledge on complex features of the system. The task at hand is then formulated as a Bayesian inference problem, which we approximately solve through variational or sampling techniques. We demonstrate how this leads to an alternative way to obtain conditional generative models. By imposing multiple constraints at once, we formulate riddles and solve them through reasoning. We also demonstrate how additional information on features can be combined with conventional noisy measurements to reconstruct high-resolution images of human faces.