Transfer learning is when a machine learning system that has been trained on one kind of data, can also be modified or easily retrained to operate on other kinds of data. For example, a deep neural network trained on English text may learn deep grammatical patrerns as well as the precise words, and so be able to be modified, with less re-training, to operate on other Germanic languages. Similarly, game playing systems may be abe to quickly pick up similar games as they have learnt aspects of deep strategy.
Used on Chap. 19: pages 465, 468