Core ML Concepts · intermediate · concept 35 of 311 · Updated Nov 4, 2026
Dropout
Dropout randomly switches off a share of a network's units during training, so no unit can rely on specific others. This acts like training many smaller networks at once and reduces overfitting at almost no cost. It is turned off for prediction. It is simple, widely used, and often one of the first things to try when a network memorizes its training data.
Key terms
Dropout rateCo-adaptationEnsemble effectTraining versus inference modeOverfitting
Learn these first
Where you meet it in the real world
Regularizing vision, language and tabular networks, including transformer layers
Guides and articles
Dropout ↗
Aston Zhang and others
Dropout: A Simple Way to Prevent Neural Networks from Overfitting ↗
Srivastava and others