Google DeepMind is a new project from the Internet and technology giant aimed at teaching computers how to read and learning how to learn from their own experiences.
An artificial intelligence (AI) system could be developed from a set of basic learning algorithms, researchers at DeepMind contend. The group intends to develop a system capable of examining vast amounts of raw data to create predictions or determine a course of action.
Artificial intelligence may be best trained by looking at examples of real-world actions, with written records of consequences. Media like the Daily Mail in England and CNN are being used as treasure troves of such annotated antidotes.
Each of these websites features notes written above the article, summarizing the story. The AI uses the article itself as the raw data and the notes as annotation. A total of 218,000 Daily Mail articles and 100,000 from CNN were used in developing the new AI system.
The system has also been tested as it played classic Atari video games, including Space Invaders. The AI is only provided with the location of pixels on the screen and the score. Only by associating the two does the system learn how to maximize its score, which is the main goal the computer is designed to achieve.
"Everything is learned completely from scratch — there's absolutely zero pre-programmed knowledge so we don't tell the agent these are Space Invaders or this is how you shoot. It's really learned from the raw pixel inputs," Mustafa Suleyman, cofounder of DeepMind, told a conference of technology experts held in London in June 2015.
When the system first encounters Space Invaders, the AI fires randomly and mostly hides behind the barriers set partway up the screen. This results in quick and frequent deaths for the virtual player. Over time, the artificial mind learns how to fire accurately at the invading spacecraft and hide behind the barriers only when necessary. Toward the end of each round, as the enemy craft come down at an ever-increasing rate, the AI is forced to calculate faster, becoming more skilled.
Traditional AI relies on information and algorithms programmed into the machine and is unable to learn from experience. Such a quality is, for the time being, exclusive to animals, including human beings.
DeepMind was acquired by Google for $633 million in January 2014.
"We combine the best techniques from machine learning and systems neuroscience to build powerful general-purpose learning algorithms," DeepMind states on their website.
Machine reading and learning could soon become common, potentially leading to new advances in robotics.
Photo: Taro Bosatsu | Flickr