What role do input features and desired outputs play in supervised learning?