Title:

InfoGAN Disentanlement Framework

Description:

Image recognition and artificial image synthesis are a common field of application where neural network architectures can be utilized to solve the problem at hand. An aggravated extension
of image synthesis deals with learning data representations in a way that the features learned by the network represent human-interpretable attributes of the data. This is called disentangled representation learning and many different approaches have been pursued to solve achieve that. This thesis provides the InfoGAN Disentanglement Framework, an application encapsulating an InfoGAN architecture, which is one of the aforementioned approaches. It utilizes concepts of information theory to separate the desired features into a specified code structure provided by the user of the framework. Furthermore, the framework’s ability to find such disentangled representations is demonstrated on different datasets of different size and complexity. Lastly, a quantitative evaluation is performed, examining the effect of parameter tuning on the quality of the disentanglement.

Supervisor: Boris Lorbeer [1], Tanja Hagemann [2]

Type:  Bachelor Thesis

Duration: 4 months