Added dropout, replaced sgd with adam optimizer

[ghost]

These changes seem to speed up training, (measured by number of epochs rather than wall clock time), in my experiments.

Please note only tested without :cuda() / cutorch.