Recent work has shown that the structure of deep convolutional neural
networks can be used as a structured image prior for solving various inverse
image restoration tasks. Instead of using hand-designed architectures, we
propose to search for neural architectures that capture stronger