Sweep over MLP learning rates

[newmanne]

I'm very suspicious that feature normalization means we're now using too low LRs