Reading the Loss Curves

Your best guide is the pair of curves: training loss and validation loss over steps. Read them together. They tell you which failure mode you are in — and when to stop.

Intuition

Pattern What it means What to try
Underfitting Both losses stay high and plateau early Train longer, improve data, or give the model more room to learn
Overfitting Train loss keeps falling while validation loss turns back up Stop earlier, simplify updates (freeze / smaller learning rate), add better or more diverse data
No learning Both curves flat near the starting value Check data pipeline, labels, learning rate, and that gradients are actually flowing
Healthy fine-tune Both losses fall together with a small, stable gap Stop near the validation-loss minimum

How it works

What good learning looks like

Hallmarks of a healthy run:

Tie-back to earlier chapters

What goes wrong

One-line summary

Watch training and validation loss together: fall together with a small gap means healthy learning; stop at the validation minimum.

Key terms