Early Stopping-But When?

Abstract: Validation can be used to detect when over#tting starts duringsupervised training of a neural network; training is then stoppedbefore convergence to avoid the over#tting ##early stopping"#. The exactcriterion used for validation-based early stopping, however, is usuallychosen in an ad-hoc fashion or training is stopped interactively. This trickdescribes how to select a stopping criterion in a systematic fashion; itis a trick for either speeding learning procedures or improving...