Backpropagation

Also: backprop

Computing how much every parameter contributed to the loss by applying the chain rule backwards through the network, reusing intermediate results so the whole thing costs about as much as one forward pass.

A definition is not an understanding. This is taught properly in How training actually works, part of Backpropagation.