Early Stopping
Halting training once validation performance stops improving for a set number of checks, then restoring the weights from whichever checkpoint scored best.
Watch validation loss every epoch, and stop once it hasn't improved for a set number of checks — the patience. Restoring the weights from whichever checkpoint scored best, rather than the final one, is what actually delivers the improvement; stopping without that restore just freezes training at a worse point than necessary.
It's regularization by training duration instead of by penalty term, which is cheap and reliable, but it spends part of the validation set's information doing the choosing — the same set can't then be reused to make other tuning decisions without leaking.