Google simulates the human brain
with 1000 machines, 16000 cores and a love of cats

Google stitched together 1,000 computers totaling 16,000 cores to form a neural network with over 1 billion connections, and sent it to YouTube looking for cats. Unlike the popular human time-sink, this was all in the name of science: specifically, simulating the human brain. The neural machine was presented with 10 million images taken from random videos, and went about teaching itself what our feline friends look like. Unlike similar experiments, where some manual guidance and supervision is involved, Google's pseudo-brain was given no such assistance.

It wasn't just about cats, of course -- the broader aim was to see whether computers can learn face detection without labeled images. After studying the large set of image-data, the cluster revealed that indeed it could, in addition to being able to develop concepts for human body parts and -- of course -- cats. Overall, there was 15.8 percent accuracy in recognizing 20,000 object categories, which the researchers claim is a 70 percent jump over previous studies. Full details of the hows and whys will be presented at a forthcoming conference in Edinburgh. (source)