Quit While You're Ahead

During training, the training loss keeps decreasing every epoch while the validation loss starts rising after epoch 20. A practitioner wants to combat this specific symptom by halting training at the right moment, rather than by synthetically expanding the training set. Which technique are they using?

Batch normalization, which halts training once running statistics stabilize

Early stopping, which monitors validation loss and halts (or restores the best checkpoint) once it stops improving for a set patience window

Data augmentation, which halts training once synthetic transformations of the input stop improving validation loss

Dropout, which halts training once neuron activations converge to zero