AI 101

Early stopping

Also known as: early stop

Training

Early stopping is the simplest anti-overfitting trick: watch the validation loss, and stop training the moment it stops improving (or starts to rise).

  • Train loss keeps falling (memorizing).
  • Validation loss plateaus/rises (overfitting).
  • Stop, keep the best checkpoint.

The discipline: always monitor a *held-out* set. The moment it stops improving, further training is usually just memorization.