Figure 3.

A schematic illustration of a single iteration of the learned proximal gradient method, algorithm 2, for a CT reconstruction problem. The choice of E _y is described in section 5.1.1. Knowledge of the forward model is incorporated into the reconstruction through ∇E _y, which is not an equivariant operator in general. Motivated by proposition 2, we know that ${\widehat{\mathrm{prox}}}_i$ is naturally modelled as an equivariant operator.