NEURAL COMPUTATION, vol.11, no.4, pp.995-1009, 1999 (Journal Indexed in SCI)
We show that with a uniform prior on models having the same training error, early stopping at some fixed training error above the training error minimum results in an increase in the expected generalization error.