Deepening our knowledge of stochastic gradient descent