To understand why Twitter just bought an artificial intelligence company called Madbits, it helps to watch a video where a modern day computer learns to play a 35-year-old video game.
Captured at a conference in Paris this spring, the video (see above) shows a machine coming to grips with a game called Breakout, something so many kids spent so many hours playing on the Atari game console in the early '80s. Breakout is kinda like Pong, where a tiny digital ball bounces around the screen and players use a tiny digital racket to knock it against various colored bricks, and at first, the machine does about as well as those kids in the early '80s, missing the ball on many occasions. But then the video shows that if the machine spends about two hours practicing, it becomes better at the game than any human could ever be. And after four hours, it not only hits the ball every time, but also figures out a wonderfully clever way of knocking down more bricks, more quickly.
The machine draws on an artificial intelligence technique known as a convolutional neural network. With this technique---a rough mimic of the networks of neurons in the human brain---a computer can learn to better handle certain tasks by doing them over and over again. The machine in the video uses convolutional neural nets to learn Breakout, Pong, and other Atari games, but the technology is also very well suited to teaching machines how to recognize what's pictured in digital photos. And judging from research published by the founders of Madbits, it seems this type of artificial intelligence lies at the heart of the image recognition technology built by the tiny New York company.
Twitter and Madbits decline to discuss the acquisition, but in a brief message posted to the Madbits website, the company's founders---Clément Farabet and Louis-Alexandre Etezad-Heydari---do say that the company has built a "visual intelligence technology that automatically understands, organizes and extracts relevant information from raw media" and that this technology is based on "deep learning," a form of AI that includes convolutional neural nets. In any event, the video above---which shows off the work of another deep learning startup called DeepMind---goes a long way toward showing what this technology is all about. Deep learning is essentially a way for machines to very rapidly teach them themselves how to do stuff.
"By the end of the video, you can see how well the machine learned," says Adam Gibson, founder of a third deep learning startup called Skymind. "Unlike human players, it takes really short jumps, never higher than it had to, which makes it faster."
>Deep learning is essentially a way for machines to very rapidly teach them themselves how to do stuff.
Deep learning is so effective, most of the biggest names in tech are now applying it to their own internet services. Before Twitter acquired Madbits, Google bought both DeepMind and DNNresearch, a startup founded by the academic at the heart of the deep learning movement, Geoff Hinton. Microsoft used deep learning to built its new Skype Translation tool. And Facebook hired Yann LeCunn, another big-name researcher in the field.
Farabet and Etezad-Heydari, the founders of Madbits, were students of LeCun's at New York University. Information about the technology their company has built is scant, but Farabet published several papers related to convolutional neural nets while at NYU and his resume says that the Madbits technology is based on his previous research. Like other deep learning techniques, convolutional neural nets are basically multi-layered algorithms that run across a large number of computers, analyzing large amounts of data in an effort to learn the task at hand.
What the company does say is that its technology is a way of carefully examining images. "Over this past year, we've built visual intelligence technology that automatically understands, organizes and extracts relevant information from raw media," reads the company's webpage. "Understanding the content of an image, whether or not there are tags associated with that image, is a complex challenge."
It is indeed. But researchers like Hinton, LeCun, and Farabet have already made some significant progress in this area. The trick with deep learning is that, by examining more and more images as time goes on, machines can get better and better at recognizing what's in them, and clearly, this is what Twitter is hoping to draw on. Google is already using convolutional neutral nets to automatically added textual tags to images posted to its Google+ social network, and this is only begins to show what deep learning is capable of. Like Facebook, Google, and others, Twitter could use such technology to power an image search engine, letting you more easily locate images posted to its social network, and it could better analyze the stuff you're posting to its service and use this information to tailor your experience accordingly, which could include carefully targeted ads.
Deep learning allows machines to process information more like humans do. But at the same time, as the Deepmind video shows, it allows machines to move beyond what humans are capable of. That is the goal not only for Twitter, but Microsoft, Facebook, Google, and so many others.
