Recent progress in generative AI has not only achieved generation of highly realistic synthetic data, but also strives to allow for manipulation of high-level properties of the data, with significant potential for applications. Building useful representations is a core ingredient to be able to perform such meaningful manipulations: this goes beyond the mere representation of statistical information, and also includes aspects of causality. The seminar will explore how these representations can be learned, how they are leveraged to achieve various purposes, and what are their properties and limitations. For that we will cover a selection of recent algorithmic, empirical and theoretical studies.
Students are enabled to independently familiarize themselves with a topic, prepare it and present it. They also become aware of the effect of their own presentation on other students. In addition, important key skills are acquired: for example, students train and improve their presentation techniques and rhetorical skills.
1 Examination: presentation. The grade depends on active participation in the seminar and the quality of the presentation and any accompanying paper.
TBA