A neural network loss L depends on a weight w through several composed functions. How is dL/dw computed?