Verbeterde kalibratie in deep learning: theorie, methoden en toepassingen

In recent years, deep learning has transformed numerous fields, from computer vision and natural language processing to autonomous systems and healthcare. The ability of deep neural networks to learn complex patterns from large datasets has enabled significant advances in AI-driven decision-making systems, including autonomous driving, fraud detection, and medical diagnostics. However, as these models are increasingly deployed in critical domains, where decisions based on their outputs can have significant consequences, their reliability becomes just as important as their accuracy. Uncertainty estimation allows practitioners to gauge the confidence of model predictions, an essential requirement for ensuring safety and trustworthiness. In applications like healthcare, where deep learning models assist in medical image analysis and diagnostics, calibrated uncertainty estimates are essential to ensure that clinicians can trust the predictions. Similarly, in autonomous driving, where models must make rapid decisions in dynamic environments, poorly calibrated predictions can lead to overconfident or overly cautious behavior, both of which carry safety risks. Despite their success, deep learning models are known to be poorly calibrated, especially when applied to challenging scenarios such as imbalanced datasets, evolving environments where data distributions shift over time, and tasks requiring calibration of the entire probability distribution. Model calibration has therefore emerged as a key area of focus in deep learning research, ensuring the alignment between predicted probabilities and observed outcomes. Standard calibration techniques, while effective for simpler models and tasks, often fall short when applied to these more complex conditions. This thesis focuses on addressing the challenges of calibration in deep learning, developing new methods that ensure models provide reliable uncertainty estimates, not only in classification tasks but also in complex applications such as object detection and volume estimation from segmentation models. By advancing calibration techniques, we aim to improve the overall reliability of deep learning models across a wide range of high-impact applications. In Part I of this thesis, we introduce a family of calibration error estimators that address three fundamental challenges: canonical calibration, calibration under label shift without access to target labels, and calibration errors induced by proper scoring rules. These estimators are designed to be consistent, differentiable, and computationally efficient, making them suitable for integration into both post-hoc and trainable calibration strategies. Extensive empirical evaluation across diverse datasets and scenarios demonstrates their effectiveness and robustness. The proposed methods significantly improve upon traditional approaches, offering a more comprehensive solution for estimating calibration error in neural networks. Part II extends the notion of calibration beyond standard classification tasks to more complex problems like object detection and image segmentation. Specifically, we propose a novel density-based calibration error estimator designed for object detection, which improves the reliability of uncertainty estimates in tasks involving structured output spaces. Furthermore, we investigate the role of calibrated segmentation models in volume estimation, showing that well-calibrated models lead to more accurate volume estimations of medical structures, such as tumors and lesions. This advancement is especially important for medical diagnosis and treatment planning, where accurate volumetric predictions can directly influence clinical decisions. By combining theoretical foundations with practical applicability, this thesis contributes new methods for enhancing the reliability and trustworthiness of deep learning models across diverse tasks. The work not only advances calibration in classification but also extends the scope to more complex domains where reliable uncertainty estimates are critical. Through the development of these novel methods, this thesis paves the way for more trustworthy deep learning systems, making them better suited for deployment in real-world applications.