Google Unveils Historic Source Code of AlexNet: The Deep Learning Revolution Enshrined in a Museum Exhibit

In 2012, three enthusiasts who have since become prominent figures in artificial intelligence — Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton — developed a model that fundamentally transformed the field of computer vision and ushered in a new era of Deep Learning. This model, trained on two GPUs in Krizhevsky’s bedroom, was named AlexNet in his honor.

Over the years, the AlexNet architecture gained such significance that the original paper, titled «ImageNet Classification with Deep Convolutional Neural Networks,» has become one of the most cited works in the history of science. Numerous researchers have painstakingly reconstructed the AlexNet code from the paper, yet the original code from the authors remained unpublished—until now.

Today, nearly 13 years later, Google, which currently holds the rights to AlexNet, has finally made the original neural network code available to the public. Moreover, it has been donated to the Computer History Museum as a significant historical artifact.

The Computer History Museum notes that publishing the source code was no small feat; museum staff collaborated with a team of Google engineers for (brace yourself!) five years to accomplish it.

«Back in 2020, we reached out to Alex Krizhevsky with a proposal to transfer AlexNet’s source code to the CHM museum, considering its historical importance. Alex connected me with Geoffrey Hinton, who was then working at Google (as the rights to AlexNet transferred to the company following its acquisition of DNNresearch). Together with the Google team, we spent five years preparing the release, meticulously assembling the original version of the code from 2012 piece by piece.»

Currently, all scripts are publicly available on GitHub here.

If you’re interested in delving deeper into this code and better understanding how modern technologies like ChatGPT, reasoning models, and more function, we invite you to join our Telegram channel Data Secrets.

There, our team of active ML engineers shares applied notes on ML, analyses of key recent papers, and intriguing news like this on a daily basis. So feel free to join us: our community of 55,000 is always happy to welcome new specialists and enthusiasts 🙂

Working with this code truly gives a sense of touching something historically significant. And understandably so: this is the very code that passed through the hands of Nobel laureate Geoffrey Hinton, OpenAI founder Ilya Sutskever, and renowned researcher Alex Krizhevsky.

However, the reason AlexNet is deemed so important that its source code has been entrusted to a museum goes beyond the illustrious names of its creators. Let’s remember that back in 2012, neural networks were still seen as a niche area with limited prospects. It was AlexNet that convincingly demonstrated for the first time that deep learning is the future of artificial intelligence. Notably, this neural network combined two critical factors that propelled deep learning to unprecedented heights.

The first was the utilization of vast datasets, which were beginning to emerge in 2012 thanks to the expansion of the internet. Specifically, the ImageNet dataset, created in 2009 by the renowned Fei-Fei Li and her team, featured millions of images painstakingly labeled by crowdsourcers.

The second factor was training neural networks on GPUs. Previously, most training was done on CPUs, but the launch of NVIDIA’s CUDA platform in the early 2010s marked the beginning of an era of accessible GPU programming.

All these developments converged in 2012, when graduate students Geoffrey Hinton’s team, including Alex Krizhevsky and Ilya Sutskever, decided to merge neural networks, massive datasets, and GPU technology into a single project. Sutskever was convinced that the performance of neural networks would improve with the quantity of data available (a true prophet!), and ImageNet provided the opportunity to validate this hypothesis. He suggested to Krizhevsky, who already had experience optimizing networks for GPU, to tackle this project under Hinton’s guidance, and he agreed.

The work took place in Krizhevsky’s bedroom, where an unceasing training and hyperparameter tuning of the network occurred on two NVIDIA GPUs over the course of a year. The resulting performance was staggering: AlexNet overwhelmingly outperformed competitors in the 2012 ImageNet competition, forever altering the course of artificial intelligence development.

Prominent scientist and one of the pioneers of AI, Yann LeCun, already recognized this milestone as a turning point. He was absolutely right: prior to AlexNet, neural networks were scarcely mentioned in leading scientific literature; afterward, they became the industry standard, laying the groundwork for the neural networks we associate with winning chess championships, chatbots, and computer vision systems today.