2012 is the clear watershed in the history of this field.

Two conditions had fallen into place beforehand. One was ImageNet, a large-scale image dataset: roughly 1.2 million images labeled across 1,000 categories, with a competition built around classifying them as accurately as possible. The other was the use of GPUs for general-purpose computation.

That year, AlexNet, submitted by Alex Krizhevsky and colleagues from Hinton's lab, beat every previous method by an overwhelming margin. The error rate dropped in one step from about 26% the year before to about 16%. Given that the usual year-over-year improvement was 1–2%, that is an outrageous number.