AlexNet

In 2012, a team from the University of Toronto submitted an entry unlike anything else in the competition.

, , and Geoffrey Hinton had built a deep neural network, eight layers deep with 60 million , and trained it on two consumer graphics cards, the same kind of GPUs from the last chapter. They hadn't hand-engineered any features. They hadn't told the network what to look for. They had simply shown it the training images, let it adjust its weights, and let it work out what mattered on its own.

Their result: 15.3% error. Nearly half the previous year's best.

The judges thought there had been a mistake. No system had ever improved that much in a single year. The gap between first and second place was larger than the entire improvement the field had made in the two years before.

The network was named AlexNet.

There was no mistake.