How we can generalise Backprop to any neural network - see how TensorFlow and Keras compute their partial derivatives! We'll use the LSTM cell as an example