DoctorKatharina Dost

My research focuses on data-centric machine learning for reliability, with particular emphasis on active learning, bias detection and mitigation, applicability domain, and adversarial robustness. A central premise of my work is that, despite its transformative potential, machine learning is only as reliable as the data it learns from, and many model failures can be traced back to gaps, biases, or mismatches in the underlying data distribution. My work therefore aims to identify and address such limitations to improve the reliability, generalisability, and trustworthiness of machine learning systems. Within this, active learning provides a principled framework in which models can identify where they are uncertain and request the most informative additional data, enabling reliable performance under data scarcity while often substantially reducing data collection and annotation effort. My work spans both methodological development and interdisciplinary applications in materials science, microbiology, agriculture, and healthcare, with a particular focus on scientific discovery settings where data is costly, limited, or biased.