UFLDL sparse auto-encoder exercise.

The hardest part is to write the backpropagation, with that done, everything is fine.

Not sure why the final training always hit maximum-iteration 400 .
but the final visualization is fine.