1: \begin{abstract}

2: 

3: Phase retrieval (PR) is an important component in modern computational imaging systems. Many algorithms have been developed over the past half century.

4: Recent advances in deep learning have opened up a new possibility for robust and fast PR. An emerging technique, called deep unfolding, provides a systematic connection between

5: conventional model-based iterative algorithms and modern data-based deep learning.

6: Unfolded algorithms, powered by data learning, have shown remarkable performance and convergence speed improvement over the original algorithms. Despite their potential, most existing unfolded algorithms are strictly

7: confined to a fixed number of iterations when employing layer-dependent parameters.

8: In this study, we develop a novel framework for deep unfolding to overcome the existing limitations.

9: Even if our framework can be widely applied to general inverse problems, we take PR as an example in the paper.

10: Our development is based on an unfolded generalized

11: expectation consistent signal recovery (GEC-SR) algorithm, wherein damping factors are left for data-driven learning.

12: In particular, we introduce a hypernetwork to generate the damping

13: factors for GEC-SR. Instead of directly learning a set of optimal damping factors, the hypernetwork learns how to generate the optimal damping factors according to the clinical settings,

14: thus ensuring its adaptivity to different scenarios.

15: To make the hypernetwork work adapt to

16: varying layer numbers, we use a recurrent architecture to develop a dynamic hypernetwork, which generates a damping factor that can vary online across layers.

17: We also exploit a self-attention mechanism to enhance the robustness of the hypernetwork.

18: Extensive experiments show that the proposed algorithm outperforms existing ones in convergence

19: speed and accuracy, and still works well under very harsh settings, that many classical PR algorithms unstable or even fail.

20: 

21: \end{abstract}

22: