Burkholz

Deep learning has achieved major breakthroughs in multiple areas ranging from computer vision and natural language processing to protein folding and continues to transform the standards of modeling across disciplines. Yet, it comes at a considerable computational cost, requires large amounts of data, is sensitive to hyperparameter tuning, and is vulnerable to adversarial attacks. The main goal of the Relational Machine Learning Group is to address these shortcomings by developing models and algorithms that are robust to noise, adapt to a changing environment, and integrate information that can be available in form of small amounts of data as well as various forms of domain knowledge. This makes our approach particularly well suited for the biomedical domain. From a methodological point of view, we like to work at the intersection of machine learning and complex network science. By interpreting deep neural networks as cascade processes, we gain theoretical insights into the computational principles underlying deep learning and translate them into practice by improving contemporary deep learning algorithms and more general probabilistic models.