"In traditional machine learning, engineers often had to define or extract the features that might matter, while deep learning largely automates the construction of internal representations."
LeCun, Bengio, and Hinton explain how traditional pattern-recognition systems depended heavily on manually engineered feature extractors, while deep neural networks learn multiple levels of representation directly from data. Their influential Nature review establishes the technical distinction behind the book's transition from handcrafted features to learned representations.
View Source