Henrik Boström

Title of the project
Reliable Machine Learning 

Background and summary of fellowship:
In many application areas, it is not sufficient to present the output of machine learning models to the users without providing any information on what leads to the specific predictions or recommendations and how (un)certain they are. The strongest machine learning models are however often essentially black boxes. In order to enable trust in such models, techniques for explaining the predictions in the form of interpretable approximations are currently being investigated. Another cornerstone for enabling trust is that the uncertainty of the output of the machine learning models is properly quantified, e.g., that the output prediction intervals or probability distributions are well-calibrated. 

Motivated by collaborations with Karolinska Institutet/University hospital on sepsis prediction, Scania on predictive maintenance and the Swedish National Financial Management Authority on gross domestic product (GDP) forecasting, techniques for quantifying uncertainty and explaining predictions will be developed and evaluated. In addition to scientific papers, the output of the project will be Python packages to support reliable machine learning, enabling predictions of state-of-the-art machine learning models to be complemented with explanations and uncertainty quantification.