Supplemental Material A Direct Approach for Designing Gradient-Driven Networks

Here we aim to design a deep network which by construction is the gradient of a certain function. The first property is a necessary condition for a network to be a gradient, and there is no apparent way to directly enforce it. One may consider additional regularization or constraints on the solution. D.1 Proof of Theorem 1 By inequality (14), for any k 0 we have || F (x The selected hyperparameters, used in the experiments, are detailed in Table 2.