Hold out is a technique used during the verification of machine learning. A proporation (often 10%) of the training data is held back as test data. The model is then trained using the remainder of the data and the hold out data used to test the quality of the model. Typically the model performs better on the data it has seen during training compared wit the text data. The hold out test data is effectively being used to check the generalisation of the training.
Used in Chap. 9: page 121