Backpropagation implementation. This is hidden in all modern deep learning libraries behind one single method call: backward(), where the magic happens. I want to un-magic that magic.
Originally I implemented all vector and matrix arithmetics from scratch, but then realised it was simply too slow when dealing with "larger" datasets such as mnist. So bye ~60 commits and hi import numpy as np.