The boundary between learned hidden representations and the output layer that maps them to predictions.