today I bumped up all my *mnist classifiers up a percent in accuracy through a lot of careful tweaking. I think I'm gonna look into more regularization methods and then put it aside for a while 'cause I really don't think I can outdo convolutional networks with my toy fully connected network