On the Steplength Selection in Stochastic Gradient Methods