Regularization networks for inverse problems: A state-space approach