Google's AI Watched YouTube And Found… Cats?
In 2012, Google let an AI watch YouTube with no labels, and the first thing it figured out on its own was cats.
The script
Google showed an AI ten million frames from YouTube videos... and never told it what was in them.
No labels. No answers. Just pixels.
It ran on sixteen thousand processor cores for three straight days.
Then researchers looked inside.
One single artificial neuron had started firing for ONE thing.
Cat faces.
Nobody taught it what a cat was. It worked it out alone... because YouTube is full of them.
It learned human faces too... but the cats made the headlines. And it all happened because...
The facts, checked
- 2012, Google Brain (Le, Ng, Dean et al.), "Building high-level features using large scale unsupervised learning".
- Trained on 10 million 200×200 frames from YouTube videos, with no labels.
- 16,000 CPU cores across 1,000 machines, for three days.
- One neuron became a strong cat-face detector; others responded to human faces and bodies.
