A recent trend in computational imaging is to learn an appropriate regularizer from training data consisting of noisy measurements and ground-truth images. At test time, the learned regularizer is used within a familiar iterative scheme, for example proximal-gradient descent. Given noisy measurements as inputs, this scheme (hopefully) converges to a reconstruction of the measured image. In this talk I'll describe an alternative application of this idea known as decision-focused learning. Here the ""image"" is instead a solution to a combinatorial optimization problem, and the ""measurements"" can encompass any observable data that is correlated with the solution.