Stochastic gradient descent explained