End-to-End Calibration of Neural Networks
M.Sc. thesis. Theory and empirical evaluation in a biomedical context. TU Braunschweig, supervised by Prof. Dr. Nicole Mücke.
Master’s thesis at the Institute for Statistics and Foundations of Machine Learning, TU Braunschweig (supervisor: Prof. Dr. Nicole Mücke).
- Investigating end-to-end calibration of neural networks: incorporating calibration objectives directly into model training rather than calibrating post hoc, and evaluating empirically how this affects predictive reliability and uncertainty estimation.
- Studying the relationship between predictive performance, confidence and reliability under distribution shift, using Transformer-based gene regulatory network inference as the application.
- Designing and implementing Python/PyTorch experiments to evaluate neural-network calibration methods on biological regulatory network data.