Train normal mnist using same input options

[rdaly525]

Train mnist using a normal linear regression. But downsample the image to the same amount that I have been in LUTNet. I want to see if I can get at least accuracy parity with a constrained input.

Right now I have been downsampling 28x28 -> 14x14 and using only the top bit. But it would be good to figure out how to do this with other mnist architectures (DNN, CNN) with whatever downsampling occuring. This will be easier with Issue #3