ML//Training//early stopping

Regularization by stopping training when validation performance stops improving, rather than when training loss bottoms out.

Regularization by stopping training when validation performance stops improving, rather than when training loss bottoms out.

Training loss keeps falling while validation loss bottoms out early and then drifts up; the gap is overfitting. Stop at the validation minimum.

Requires a held-out fold that model selection never trains on. Under grouped data, the stopping fold must respect the same cross-validation split.

Cheap and effective for LSTM and gradient boosting alike.