Improving Confidence Calibration of Deep Learning Systems Under Label Noise and Domain Shifts

A recent thesis published on arXiv (2608.12100) tackles the issue of confidence calibration in deep learning models, especially in high-stakes environments where accurate confidence estimates are crucial alongside predictions. This research emphasizes situations where clean validation data is lacking due to label noise and domain shifts, frequently encountered in practical applications. It introduces a framework that utilizes an estimated noise model to derive noise-free confidence estimates by analyzing the connection between noisy and clean label distributions. Furthermore, this method is adapted to Conformal Prediction (CP), which offers set-valued predictions with assured coverage. The noise-aware CP technique seeks to uphold calibration guarantees despite unreliable labels. This work is vital for the secure implementation of AI in essential fields like healthcare, finance, and autonomous driving, where miscalibrated confidence can result in perilous decisions. The thesis is classified as a cross-type announcement on arXiv and was released on August 12, 2026.