In transfer learning, what is typically done with the earlier layers of the pre-trained neural network?