Convergence rates for momentum stochastic gradient descent with noise of machine learning type