Training improves a model in rounds; each full pass through the data is an epoch. Score on the training data usually keeps improving, but at some point the model starts memorizing quirks of those examples and gets worse on new ones. Early stopping watches a separate validation set after each round and stops when that score hasn’t improved for a set number of checks, often called the patience.
Say a team fine-tunes a small model on internal help-desk answers. Training loss falls every epoch, but validation loss bottoms out at epoch 4 and rises for the next three. With a patience of three, training stops at epoch 7. Whether the team ends up with the epoch 4 model or the epoch 7 model depends on a separate setting: Keras, for instance, keeps the last weights unless restore_best_weights is turned on. Gradient-boosted trees use the same idea to decide how many trees to add.
Early stopping is a form of regularization, and the patience and the metric to watch are hyperparameters. Because the validation set now helps choose the model, judge the final result on a separate test set. Early stopping won’t fix wrong labels, test data that leaked into training, or the world changing after deployment.
Sources
- Keras: EarlyStopping callback — Stops training when a monitored metric such as validation loss has stopped improving, with patience and min_delta settings and a restore_best_weights option that defaults to False; checked October 24, 2026.
- Goodfellow, Bengio and Courville: Deep Learning, Chapter 7 (Regularization) — Treats early stopping as a regularization strategy: return the parameters from the point of lowest validation error rather than the last training step; checked October 24, 2026.
- scikit-learn: Gradient boosting — Documents early stopping for gradient-boosted trees through validation_fraction, n_iter_no_change and tol, so the model stops adding trees when the validation score stalls; checked October 24, 2026.