Room 4 · The Awakening · 2012

The Network That Found Cats

The Network That Found Cats (Andrew Ng, Jeff Dean and team · Google Brain, 2012): Learning without anyone adding labels.

A Google team trained a network with 16,000 processors for about three days, using ten million images taken from YouTube videos and no labels. One of its neurons ended up responding to cat faces, without anyone having taught it what a cat is.

In June 2012 a team from Google and Stanford presented a paper at the International Conference on Machine Learning with a modest title: "Building high-level features using large scale unsupervised learning." The authors included Quoc Le, Marc'Aurelio Ranzato, Greg Corrado, Jeff Dean and Andrew Ng. The press summarized it in three words: Google found cats.

The experiment tested a question that had been around for years. Can a neural network learn to recognize concepts without anyone telling it what is in each image? Most systems of the time learned from labeled examples: thousands of photos marked "cat" or "not cat." This one received no labels at all.

The team took ten million frames from YouTube videos and trained a very large network, with about a billion connections, spread across 1,000 machines with 16,000 processor cores, for about three days. The network's only goal was to compress the images and rebuild them as well as it could. To do that, it had to discover which patterns repeat across many images.

Afterward the researchers probed individual neurons. One of them responded strongly to cat faces, although the network had never been told what a cat was. Others responded to human faces and human bodies. On the ImageNet benchmark the learned features also improved accuracy considerably over the previous best result for that version of the task.

The cat made headlines, but the lasting impact was organizational. The project was one of the first at Google Brain, the deep learning team founded in 2011, and it showed that a large company's computing infrastructure could be turned into a tool for training very big networks.

Ask your AI: what is the difference between supervised and unsupervised learning? Ask for one example of each with photos.

Did you know? The network also developed neurons that responded to human faces and bodies.

Source: arxiv.org/abs/1112.6209

See it in the museum →

← AlexNet Wins ImageNet · word2vec: Meaning as Geometry →

All of Room 4: The Awakening

Cite this: Nazareno, I. (2026). The Network That Found Cats (2012). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/the-network-that-found-cats-2012/