Train faster, generalize better: Stability of stochastic gradient descent

  Abstract