Finding Critical and Gradient-Flat Points of Deep Neural Network Loss Functions