A much larger picture collection

A learning system needs examples paired with the answers it should learn. For images, that means people must identify what each picture contains.

and her team built ImageNet, a collection organised by labels such as “school bus,” “coffee mug,” and thousands of other categories. Online workers checked pictures and attached the correct labels.

The first ImageNet paper in 2009 described 3.2 million images across 5,247 categories.

Think of it as an enormous set of study cards. Each picture sits on one side and a human-provided answer sits on the other. The labels did not appear by magic. People supplied the supervision that made training possible.