4 ms·

I could very well imagine that the random batching in stochastic gradient descent is great for regularizing and preventing local optima.