Optimization in Machine Learning: Robust or global minimum?

We understand that in convex problems it is much easier to find the global optimum. We appreciate the opportunity to participate in this discussion. KD, MF: No, the convexified problem can have a minimum that is quite different from the original problem. The motivation for our paper comes from the fact that in many problems (like control and reinforcement learning) one is interested in a "robust" minimum (a minimum such that the cost does not increase much when you perturb the parameters). Our method destroys non-robust minima and preserves a single robust minimum of the problem.