Incorporating prior knowledge into generative adversarial networks

In the last few years, we have seen several success stories regarding the application of machine learning to different tasks, e.g. deep learning for speech recognition, music processing or image classification. The accuracy of machines in such problems has improved dramatically. However, when the task at hand involves a deeper understanding, machine learning techniques are far behind humans.
The main concept of this thesis is to develop novel machine learning systems that by encoding expert knowledge are able to provide a deeper understanding of heterogeneous media and generalise to new contexts and application domains and to communicate the understanding of the signals to other systems, including humans.
This will mainly be applied to Generative Adversarial Networks which is a model that takes a training set, consisting of samples drawn from an underlying probability distribution, and learns to represent an estimate of that distribution somehow. The generative nature of these networks makes them ideal in order to encode prior knowledge, in a similar manner as one would in probabilistic graphical models.