Numeric underflow when loss gets small.

[dhruvrajan]

When training MNIST for > 14 epochs, there seems to be an underflow problem, resulting in NaNs when the loss becomes ~ 7E-3.