How models actually learn, from vanilla gradient descent to Adam
Early stopping uses validation performance to decide when to stop training. If validation loss stops improving for long enough, you keep the best checkpoint and stop. Progress here is counted in epochs: one epoch is one full pass through the training data, and validation is typically checked after each one.
This is both a compute saver and a regularizer. It prevents the model from continuing to fit the training set after validation performance has started to get worse.
A toaster gives the right instinct. Pulling bread out too early leaves it pale. Waiting too long burns it. You watch the toast and stop when it reaches the best color, not when the heating element has run as long as possible. The figure below shows where the "watching" data comes from: hold out part of the training data, or rotate folds as shown, so the model is judged on examples it never fit. Early stopping reads its signal from exactly that held-out slice.