Pages Menu
TwitterRssFacebook
Categories Menu

Posted by on Jun 26, 2012 in Hardware, Software |

Google X creates a network of computers that learn to recognize cats without being taught

Google X creates a network of computers that learn to recognize cats without being taught

The secret lab Google X developed a “virtual neural network,” a group of 1,000 computers with 16,000 cores total, which gave millions of images of 200 X 200 pixels taken at random YouTube videos.

What did the system with the pictures? Find cats (as many YouTube users out there). Interestingly, no one told the system how it was a cat, so that the network learned by itself to identify the recurring image of this cat. The researchers presented the results of their research this week at a conference in Scotland.

The machine began to search for matches between all images which enabled him to detect faces, and also “high concepts” like cat faces and human bodies.

No one taught him how the machine was a face, a body or a cat before the start of the analysis of images. Once discovered an object that was repeated several times, the computer developed maps of images that he then used to detect similar objects. In Google maps X called these “neurons” in reference to the theory that our brains are some neurons in the temporal cortex with the specific task to recognize categories of objects such as faces or hands.

Typically, computers are told how an object is defining the edges of shapes, and then marking images that contain these objects. In this project, “we never told in training ‘this is a cat’. Basically invented the concept of the cat, “said Jeff Dean, who worked with the network.

The system is still new and needs much work, but since this first success, Google will be removed from X to continue work on the search team and business. Google hopes to improve the algorithm and use it in its image search service, voice recognition and language translation.

Link: In a big network of computers, Evidence of machine learning (NYTimes)

Tags: , , , ,