Towards Faster Stochastic Gradient Search