Reviews: A Bridging Framework for Model Optimization and Deep Propagation

Paper summary: The paper proposed a learning based hybrid proximal gradient method for composite minimization problems. The iteration is divided into two modules: the learning module does data fidelity minimization with certain network-based priors; consequently the optimization module generates strict convergence propagations by applying proximal gradient feedback on the output of the learning module. The generated iterates were shown to be a Cauchy sequence converging to the critical points of the original objective. The method was applied to image restoration tasks with performance evaluated. Comments: The core idea is to develop a learning based optimization module to incorporate domain knowledge into conventional proximal gradient descent procedure.