LearningBackprop

Context: MachineLearning, NeuralNetworks

This is wild. Back-prop (back-propagation) is the way that neural networks feed information about the error between what they are currently generating and what the trainers would like them to be generating. It's a crucial part of their learning.

Here are some people who claim to have an algorithm that can actually learn back-prop itself (presumably by some lower-level "learning method") without having some of the overhead of sending the information back.

Not sure how this can really work, but I'm making a page for it while I try to understand it.