sourced story
30 September 2012Primary source · 2 sourcesWell documented

AlexNet wins ImageNet and starts the deep learning boom

Two gamer GPUs and a network eight layers deep crush the field by more than double digits

On the timeline · around 30 September 2012 · The Statistical and Machine Learning TurnThe Statistical and Machine Learning TurnTransformers and the Generative AI WaveAlexNet wins ImageNet and starts the deep learning boom200820102012201420162018

Quick facts

Researchers
Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton
Institution
University of Toronto
Result
15.3% top-5 error vs 26.2% for runner-up

What happened

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, at the University of Toronto, entered a convolutional neural network in the 2012 ImageNet Large Scale Visual Recognition Challenge. The network, later called AlexNet, had 60 million parameters across five convolutional layers and three fully connected layers, and was trained on two consumer-grade Nvidia GeForce GPUs rather than specialized research hardware. It used ReLU activations to train faster than the standard sigmoid function of the time, and a regularization technique called dropout, which randomly disables neurons during training, to reduce overfitting. AlexNet achieved a top-5 error rate of 15.3 percent, beating the second-place entry's 26.2 percent by nearly 11 percentage points, a margin far larger than the field's usual year-over-year gains.

Why it matters

AlexNet's margin of victory convinced much of the computer vision field almost overnight that deep neural networks, trained on large labeled datasets with GPU hardware, outperformed the hand-engineered feature methods that had dominated the field for years, and the paper's approach became the template for the deep learning decade that followed.

How we know

The original NeurIPS paper documents the network architecture and error rates directly; the Computer History Museum's 2025 release of AlexNet's original source code, reported in its own detailed blog post, independently confirms the architecture described in the paper.

Sources

See something wrong? . Corrections with a source get fixed fastest.

Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →