AlexNet: ImageNet classification with deep CNNs

AlexNet, named after Alex Krizhevsky and built with Ilya Sutskever and Geoffrey Hinton, is a convolutional neural network. That is a kind of model that learns to spot patterns in images by scanning small patches and combining them into larger and larger features, from edges to shapes to whole objects.

The contest was ImageNet, a challenge to correctly label photos across a thousand categories. AlexNet made far fewer mistakes than the methods that came before it. The gap was wide enough that people paid attention immediately.

Two things came together. First, the network was deep, with many layers stacked so it could learn rich features on its own instead of relying on hand-designed rules. Second, the team trained it on GPUs, the chips originally made for video games, which could do the heavy math fast enough to make a big network practical.

There was also plenty of labeled data to learn from. Deep model, fast hardware, and lots of examples turned out to be a powerful recipe, and that combination is still the backbone of modern AI.

AlexNet is widely seen as the spark of the deep learning era. After it, researchers and companies poured effort into neural networks, and progress in vision, speech, and eventually language followed.

For anyone building with AI today, this paper is the origin story of the tools you use. The lesson it teaches, that scale of data and compute can unlock learning that hand-tuned methods cannot, has guided almost everything since.