add mini-batch training

[A3Ackerman]

Stochastic gradient descent should be more efficient and less likely to get stuck in local minima