Even with the right model structure, "how we train it" makes or breaks performance. Today covers loss functions, optimizers, learning rate — the training engine — and transfer learning/fine-tuning that achieves great results with limited data. Plus framework (TensorFlow/PyTorch) and SageMaker integration that tests often ask. MLS-C01 frequently asks "When learning fails/is slow/overfits, what do we change?" — this domain.
Quantifies difference between prediction and truth. Choose per problem type.