[asib]

Looks like this implementation is based (in part) on the Stanford ML course, which teaches nnets using sigmoid activation.

Given that it's intended to introduce to beginners how nnets work, the choice of activation is an aside anyway - the real meat is back/forwardprop.

[ironislands]

Absolutely, though I'll put in softmax, tanh and ReLU activation functions in near time. It isn't that difficult. Putting in more docstrings is also one of my todo.