About this Event
6548 Forest Park Pkwy, St. Louis, MO 63112, USA
Robust Deep Learning Models for Imaging Inverse Problems
Modern imaging problems increasingly rely on deep generative models, including score-based and flow-matching models, as well as neural networks used as implicit priors. Although these methods can produce high-quality reconstructions, their reliability may deteriorate when deployment conditions differ from those encountered during training. Such changes may affect the image distribution, the measurement process, or other characteristics of the imaging problem. This work investigates the impact of these mismatches on reconstruction algorithms and their performance. It develops theoretical tools for understanding their consequences, algorithms for detecting distribution shifts, and strategies for correcting or mitigating their effects. It also introduces frameworks for designing imaging methods that remain reliable and robust under changing conditions. Together, these contributions support the development of trustworthy learning-based imaging systems for real-world applications.
Event Details
See Who Is Interested
About this Event
6548 Forest Park Pkwy, St. Louis, MO 63112, USA
Robust Deep Learning Models for Imaging Inverse Problems
Modern imaging problems increasingly rely on deep generative models, including score-based and flow-matching models, as well as neural networks used as implicit priors. Although these methods can produce high-quality reconstructions, their reliability may deteriorate when deployment conditions differ from those encountered during training. Such changes may affect the image distribution, the measurement process, or other characteristics of the imaging problem. This work investigates the impact of these mismatches on reconstruction algorithms and their performance. It develops theoretical tools for understanding their consequences, algorithms for detecting distribution shifts, and strategies for correcting or mitigating their effects. It also introduces frameworks for designing imaging methods that remain reliable and robust under changing conditions. Together, these contributions support the development of trustworthy learning-based imaging systems for real-world applications.
