Diploma Thesis

With the emergence of Generative Adversarial Networks, Variational Autoencoders and Adversarial Autoencoders as neural networks, we have very useful tools available to encode complex distributions of high dimensional data to a very meaningful and compressed level. Autoencoders usually just encode variations between single data points. With some modifications it is possible to extend the model, to disentangle the encoding of variations between subsets (groups) of the data and the variation of single data points inside each group. This thesis describes a specific training algorithm applied on an Adversarial Autoencoder to find representations of data and separate them into a representation of style and a representation of content. A still open question is how to measure the effectiveness of this disentanglement, especially when the true generative factors are unknown. We will conduct experiments that use the performance of a data classifier to approximate the quantity of the imposed disentanglement and observe how the reconstruction quality is influenced by manipulations of the disentangled representations.