why stochastic gradient descent works