Naive solution (1 line):

https://gist.github.com/applecrazy/deda2fac6e83c07b93e001731...

Edit: I literally took newlines, converted to \n in a string, and then exec()ed the whole thing. Here's a repl of it working: https://repl.it/@applecrazy/Code-Golfing-a-Neural-Net

Sure! Optimizing for number of lines (instead of number of characters):

* Remove 'import numpy as n' and use __import__('numpy') in place of it everywhere

* Remove s and d functions; inline them where they're called

* Get rid of 'class N', as it's unnecessary. If adding globals is cheating, then you can do '__import__('math').x = x' instead of 'self.x = x' (yes this will work and persist).

* Technically, you don't have to print the result at the end

Where do I go for my beer?

Not readable, doesn't count ;)