Translations:Overfitting and Regularization/23/en
At test time, all neurons are active but their outputs are scaled by $ (1 - p) $ to compensate for the larger number of active units (or equivalently, outputs are scaled by $ 1/(1-p) $ during training — inverted dropout).