Backpropagation

Also written: backprop, backward pass

The chain rule applied backward through a network's layers, one local derivative at a time, to get the gradient for every parameter. Alex's cold brew network is trained this way.