To Learn: AlexNet and the Neural Network

Watson could retrieve an answer someone had already written down. The next leap was a machine that could learn to recognize something no one had explicitly told it how to see.

Key Milestones

• AlexNet wins the ImageNet competition, 2012. A deep convolutional neural network, trained on 1.2 million labeled images, outperformed every hand-engineered computer vision method before it by a wide margin.

The Energy, Social, and Economic Impact of Learning

ImageNet was a large public test: 1.2 million photos, each labeled with what is in it, a dog, a bicycle, a coffee cup, and the challenge was simple to state and hard to solve. Build a program that looks at a new photo it has never seen and correctly names what is in it.

For years, researchers did this by hand, writing rules such as "look for round shapes and pointed ears to spot a cat." AlexNet discarded that approach. It was a neural network, simply shown millions of labeled photos and left to work out the visual patterns itself.

The gap was not subtle. The best 2012 entry using hand-engineered techniques posted a top-5 error rate of 26.2%, meaning the correct answer was not even in its top five guesses more than a quarter of the time. AlexNet, using its self-taught approach, posted 15.3%, nearly eleven points better. We believe every serious computer vision system built since descends from what AlexNet proved: a large enough neural network, trained on enough data, with enough GPU power, could outlearn methods engineers had spent a decade hand-tuning.

View chart in the full report

This is where the arbitrage in this series quietly changed shape. Deep Blue and Watson were both, in their own way, told what to look for. AlexNet was not. It was shown data and left to find the pattern itself: energy converted not into an answer, but into the capacity to learn one.

This marked the arrival of the neural network, the architecture nearly every AI system since has been built on.