Decaying error term in stochastic gradient section

[lesniewski]

Thanks for an excellently written and beautifully illustrated document.

I was confused by the final formula, which may reflect my misunderstanding, or may be an error in the formula. In "a decaying sum of the errors", shouldn't the epsilon term be inside the summation? As written now, it would seem that only the last minibatch's error is included.

See the following commit for the proposed fix.
lesniewski@cdfe014