Introduction
Once the sanity checks pass (topic 4.1), you train for real. The most important tool you have while training is the loss curve: the training loss and the validation loss plotted over time, together with a validation metric like accuracy.
A doctor looking at a patient's temperature chart over a week can often tell what's going on without any other test: a steady fever, a sudden spike, a slow recovery. Loss curves work the same way. Their shape tells you whether the model is learning, whether it's memorizing, whether training is unstable, and what to try next.
This topic is a field guide to the common shapes.