Necessity and optimality of unrolled optimization-inspired architectures in computational imaging

Determine whether unrolled optimization-inspired neural network architectures are necessary or optimal for neural network design in computational imaging tasks such as Magnetic Resonance Imaging (MRI), computed tomography, and astronomical imaging, in contrast to empirically designed end-to-end architectures that do not strictly adhere to optimization-derived operations.

Background

Computational imaging problems are often addressed with unrolled optimization-inspired architectures that explicitly incorporate the measurement operator and observed data. This approach has become standard in domains such as MRI and computed tomography due to its strong performance and principled conditioning on acquisition physics.

However, empirical successes of end-to-end architectures (e.g., UNet variants and transformer-based models) in simpler tasks like denoising suggest that strict adherence to optimization-derived operations may not be necessary. The paper explicitly notes that it remains unclear whether unrolling is required or optimal for neural network design, motivating their proposed non-iterative RAM architecture that conditions on measurement operators without unrolling.