Thema der Dissertation:
An interpretable omnigenic neural network architecture for the human genome Thema der Disputation:
From predictive performance to scientific knowledge: benchmarking in biomedical machine learning
An interpretable omnigenic neural network architecture for the human genome Thema der Disputation:
From predictive performance to scientific knowledge: benchmarking in biomedical machine learning
Abstract: Machine learning increasingly produces useful predictions before the underlying patterns or mechanisms are fully understood, creating a gap between predictive performance and scientific understanding. Benchmarks are among the primary instruments for bridging this gap and turning predictive success into scientific knowledge. A benchmark is not merely a dataset paired with a performance score. Rather, it is a community-level measurement system that links tasks, metrics, assumptions, and scientific claims. In this talk, I introduce key concepts from the science of benchmarking and examine common failure modes. Benchmarks may reward spurious artifacts, reduce complex constructs to a single score, encourage overfitting to public test sets, or emphasize aggregate accuracy while overlooking local, subgroup-specific, or mechanistic validity. I then apply this framework to biomedical machine learning, with a particular focus on risk prediction in population cohorts. In this setting, phenotype definitions, shifts in ancestry and population composition, calibration, privacy constraints, and the choice of comparators make rigorous benchmarking particularly challenging. In conclusion, better benchmark design in biomedical machine learning can help ensure that gains in predictive performance translate into robust and scientifically meaningful advances.
Zeit & Ort
25.09.2026 | 11:00
Seminarraum 031
(Fachbereich Mathematik und Informatik, Arnimallee 7, 14195 Berlin)