Does Deep Learning Solve the Phase Retrieval Problem?
Tuesday, September 8, 2020 - 1:25pm - 2:25pm
Phase retrieval is a difficult inverse problem, with three types of intrinsic symmetries. These symmetries can fail even the classic methods, e.g., HIO on complex-valued images when the object support is not precisely known --- allowing free translations. Do these symmetries cause learning difficulty if one deploys the deep learning approach? We show that that’s indeed the case, and we present two solutions to the problem: one is active symmetry breaking based on careful pre-processing on the training data, and the other passive symmetry breaking exploiting massive amounts of data and implicit regularization.