Mathematische Modellierung und Datenanalyse

Feature importance methods promise to provide a ranking of features according to importance for a given classification task. A wide range of methods exist but their rankings often disagree and they are inherently difficult to evaluate due to a lack of ground truth beyond synthetic datasets. In a recent paper, PTB  together with colleagues from NPL, IMBIH, King´s College, KIT, Fraunhofer HHI and...

The Bayesian approach to solving inverse problems strongly depends on the choice of a prior. Usually, a prior is constructed from expert knowledge or known physical constraints in a probabilistic fashion. A modern alternative to formulate such expert knowledge are generative models, a popular tool in machine learning to generate data whose properties closely resemble a given database by an...

Scanning probe based spectroscopy using broadband infrared radiation emerged as a promising imaging technique at nanometer spatial resolution. However, the pixel-by-pixel data acquisition leads to prohibitive imaging times and enhanced radiation damage. To overcome this issue, a novel hyperspectral imaging scheme was developed in this project using Bayesian compressed sensing (BCS).