AlexNet
Era 4 · Deep learning and foundation models (2012–2021) › Deep learning
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
University of Toronto
What happened
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton of the University of Toronto submit AlexNet to the ImageNet competition on September 30, 2012. It wins with a top-5 error of 15.3 percent, more than ten points below the runner-up.
Why it mattered
A deep convolutional network trained on two consumer GPUs clearly beats hand-designed vision methods. It is the starting point of the deep learning boom.
The full story
In 2012 the annual ImageNet competition asked entrants to classify images into a thousand categories, and the best systems combined hand-designed visual features with statistical classifiers. Krizhevsky, one of Hinton's students, had already written cuda-convnet, a program for training convolutional networks on GPUs, and extended it to split a network across two cards.
AlexNet had eight layers, five convolutional and three fully connected, with 60 million parameters and 650,000 neurons. It used the ReLU activation function, which trained better than the usual ones, and dropout, which randomly switches off neurons, to keep it from memorizing the examples. It was trained on 1.2 million images for five or six days on two Nvidia GTX 580 GPUs, in Krizhevsky's bedroom at the family home.
Its error of 15.3 percent left the second-place entry more than ten points behind. The margin convinced much of the computer vision community that deep networks, graphics hardware, and large datasets worked together.
Shortly afterward the three authors founded DNNresearch and sold the company, along with the AlexNet code, to Google. The paper appeared in the NIPS 2012 proceedings and, in 2017, in Communications of the ACM.
Builds on
CUDA (February 15, 2007), ImageNet (June 2009)
Documented direct influence
AlexNet is trained on GPUs with CUDA, on ImageNet
DanNet (2011)
Documented direct influence
The AlexNet paper cites two papers by Cireşan et al., including the one on multi-column networks, and says its architecture is "somewhat similar" to the "columnar" CNN of Cireşan et al. (section 3.2)
Threads that pass through here
- Data, GPUs, and AlexNet (step 4 of 4)
- The neural network winter (step 11 of 11)
Sources
Original source
- ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6), pp. 84-90 (opens in a new tab). Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, 2017
Further reading
How to cite
“AlexNet.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alexnet
This entry was last reviewed on November 13, 2026.