updated 03:15 pm EDT, Tue June 26, 2012
1000 machine neural network scanned 10M YouTube images
Google's X Lab has built a one-billion-connection "neural network" that can identify cats on YouTube. The project, aimed at simulating object recognition by humans, was able to more than double the accuracy of item identification from a list of 20,000, according to the New York Times, Using 16,000 cores in 1000 connected machines, the system learned to identify objects without human supervision. The technology represents a departure from current vision-learning methods.
The system was fed 10 million images from YouTube thumbnails, at a resolution of 200x200. After three days of learning, it was able to recognize not only the human face and the human body, but also cats, a subject seen frequently in online video clips. The improved 15.8 percent accuracy rate is said to be a leap of 70 percent over the previous state-of-the-art system.
A presentation of the researchers findings will be presented at a conference in Scotland, and it is likely that Google will capitalize on the research, having previously worked with face-recognition technology.