Core Concepts
Symmetry breaking is crucial for improving end-to-end learning performance in phase retrieval problems.
Abstract
The content discusses the challenges faced in end-to-end learning for phase retrieval, focusing on the importance of symmetry breaking. It explains how intrinsic symmetries can lead to learning difficulties and proposes a novel technique to preprocess training sets before learning. The paper provides insights into the mathematical principles behind symmetry breaking and its application in far-field phase retrieval (FFPR). Experimental results demonstrate the significant improvement in performance after symmetry breaking, validating its effectiveness.
Introduction
Symmetries in imaging science's forward model.
Challenges with data-driven deep learning approaches.
Methods
Symmetry breaking for learning square root.
Symmetry breaking for phase retrieval.
Experiments
Evaluation dataset: simulated Bragg CDI crystal dataset.
Experiment setup using UNet and SiSPRNet models.
Results
Uniform improvement due to symmetry breaking.
Performance boost on training and test sets after symmetry breaking.
Related Work
Recent research efforts on solving inverse problems using DL.
Author Biography
Brief biographies of Wenjie Zhang, Yuxiang Wan, Zhong Zhuang, and Ju Sun.
Stats
Given Y = |F(X)|^2 ∈RM1×M2 (R+ means nonnegative reals).
To ensure recoverability, M1 ≥ 2N1 − 1 and M2 ≥ 2N2 − 1 are necessary.
Quotes
"Symmetry breaking leads to a much smoother target function, making it easier for DNNs to learn."
"In terms of SA-MSE loss, training on unprocessed sets often performs worse than non-data-driven methods."