Evolutionary Stochastic Gradient Descent for Optimization of Deep Neural Networks