Schatting en optimalisatie onder onzekerheid

Modern deep neural networks have achieved remarkable performance across diverse domains, yet they face critical challenges in computational efficiency, reliable uncertainty quantification, and trustworthy probability estimates. These challenges become particularly pronounced when deploying models in resource-constrained environments, optimizing complex objective functions under noise, or ensuring safe decision-making in critical applications. Although substantial progress has been made on each front, many approaches treat these issues in isolation and lack frameworks that integrate theoretical rigor with practical applicability. This project addresses these gaps by developing methods that (i) formulate compression problem from the perspective of discrete optimization, (ii) handle noise-aware optimization in Bayesian settings, and (iii) provide statistically grounded estimators and training objectives for selective prediction and calibration.