Efficient Training of Low-Curvature Neural Networks