On Using Simulated Annealing in Training Deep Neural Networks