Exact Gauss-Newton Optimization for Training Deep Neural Networks