Convergence of Stochastic Gradient Descent in Deep Neural Network