Design learning and statistical methodologies to effectively identify explanatory features (e.g., genetic variations) truly linked to a phenomenon under study (e.g., disease risk) while rigorously controlling the number of false positives among the reported features.
Develop methodologies that provide provably robust predictions in a challenging setting where the train and test distribution differ, e.g., due to adversarial attacks.
Develop statistical tools that can work in combination with any complex machine learning algorithm (e.g., a deep neural network) to provide a reliable assessment of prediction uncertainty. The tools we invent treat regression, classification, and out-of-distribution detection problems.
