What is gradient descent, and how does it work in the training of machine learning models? How do you decide the learning rate?