Exploring the loss landscape of regularized neural networks via convex duality