Inhalt / Abstract:

This lecture elaborates on both theoretical and practical aspects of machine learning and deep learning, with a focus on uncertainty, respectively. Firstly, the most important aspects of statistical learning theory are introduced. This includes the so-called probably approximately correct learnability of predictors under specific assumptions as well as the limitations of learnability. Thereafter we study neural networks including algorithmic aspects. This is followed by notions of uncertainty for machine learning in general but also specifically for deep learning. In particular we deduce why neural networks oftentimes tend to be overconfident although they are completely wrong. We study different methods of uncertainty quantification for machine learning and deep learning and their applications. This is complemented with illustrative nutshell examples but also examples from cutting edge applications such as automated driving. If time permits, we conclude this lecture with a brief excursion into the ethical problems and consequences when using deep learning in safety relevant applications. The corresponding exercise course contains both theoretical and computational tasks.