Labeled data pairs each example with a target outcome. In supervised learning, these pairs let an algorithm associate inputs with the answers it is meant to predict. Deep learning often relies on many such examples, although it can also use other learning approaches. section 1.6 section 3.4
Cats and dogs
For a system that classifies cat and dog photographs, the images are inputs and the animal categories are targets. The image pixels can be represented as numbers for a neural network; those numbers are not themselves the labels. section 3.4
The practical obstacle is supply. Many businesses lack enough data, particularly labeled data, to train a deep network. ImageNet is an unusually large resource, with more than 14 million human-annotated images available for object recognition testing. Having labels does not remove every risk: a trained network may underperform on data unlike its training examples, and changing a project’s objective can require expensive retraining. section 3.4.2
