Accelerating Deep Neural Network Training with Inconsistent Stochastic Gradient Descent