The algorithm used to train neural networks by calculating gradients of the loss function with respect to each parameter and propagating errors backward through the network layers.