On the Use of Machine Learning for Computational Imaging

Explore the intersection of machine learning and computational imaging in this 44-minute lecture by George Barbastathis from the Massachusetts Institute of Technology. Delve into the complexities of using supervised machine learning for computational inverse problems, examining the role of regularization in overcoming ill-conditionedness and ill-posedness. Investigate key questions such as whether to explicitly include physics captured in the forward operator into learning architectures or build all-encompassing black boxes. Address challenges like partially known forward operators and the high cost of obtaining sufficient experimental training pairs. Gain insights from implementations of machine learning-aided inverses in three classical inverse problems: electromagnetic field phase retrieval from intensity, 3D dielectric structure retrieval from limited-angle intensity projections, and quantitative analysis of highly scattering surfaces. Consider future directions for joint optimization of forward operators and machine learning inverses to enhance robustness against noise and uncertainties.