Parametric Dual Maximization for Non-Convex Learning Problems