Single-loop methods for bilevel parameter learning in inverse imaging

Abstract: Bilevel optimisation is used in inverse problems for hyperparameter learning and experimental design. For instance, it can be used to find optimal regularisation parameters and forward operators, based on a set of training pairs. However, computationally, the process is costly. To reduce this cost, recently in bilevel optimisation research, especially as applied to machine learning, so-called single-loop approaches have been introduced. On each step of an outer optimisation method, such methods only take a single gradient descent step towards the solution of the inner problem. In this paper, we flexibilise the inner algorithm, to allow for methods more applicable to difficult inverse problems with nonsmooth regularisation, including primal-dual proximal splitting (PDPS). Moreover, as we have recently shown, significant performance improvements can be obtained in PDE-constrained optimisation by interweaving the steps of conventional iterative solvers (Jacobi, Gauss-Seidel, conjugate gradients) for both the PDE and its adjoint, with the steps of the optimisation method. In this paper we demonstrate how the adjoint equation in bilevel problems can also benefit from such interweaving with conventional linear system solvers. We demonstrate the performance of our proposed methods on learning the deconvolution kernel for image deblurring, and the subsampling operator for magnetic resonance imaging (MRI).