Email
Password
Remember meForgot password?
    Log in with Twitter

article imageFacebook's latest AI can 'reason' and process video in real time

By James Walker     Nov 8, 2016 in Technology
Facebook has detailed how it's using and developing artificial intelligence. It now supports three different platforms that can be used by engineers to create new apps and Facebook features, including a real time video filtering service for smartphones.
Facebook said it views artificial intelligence as a way to open the world to everyone. It's approaching it as part of a long-term plan that will see AI integrated into every area of Facebook over the next ten years. To make the concept reality, it's conducting extensive research into multiple areas of artificial intelligence. It has also created three infrastructures that support the company's work in AI, machine learning and neutral networks.
The "backbone" of Facebook's AI-based product development is FBLearner Flow, an AI platform that everyone at Facebook can use. FBLearner Flow has dramatically increased AI usage at the company. Twice as many experiments using the technology are now run each month compared with six months ago. 70 percent of the developers using the platform are not experts in AI.
FBLearner Flow is joined by AutoML, a system that uses AI to optimise AI. AutoML can test over 300,000 machine learning models every month, automatically applying the results of previous tests to make new systems perform better. Finally, there's Lumos, a self-serve platform that helps product teams integrate computer vision into their features. Lumos is already at work, improving Facebook's ability to identify content that violates its community standards.
The systems have enabled Facebook to experiment with AI in new ways. Over the past three months, it has developed a new deep learning platform that is capable of running on iOS and Android mobile devices. Caffe2Go can capture, analyse and process the pixels of video streams in real time.
This has allowed Facebook to bring a technique called "style transfer" to mobile devices. It lets you transform your photos and videos into images that resemble works by historical artists, all in real time. The neural network behind the technology is 100x smaller than Facebook's existing image processing model, making it lightweight enough to run on mobile devices. Facebook said that real time AI could have important applications in the future, helping people stay connected and learn from others.
"As we move forward, you can imagine how having on-device AI running in real time could help make the world more open and connected for people in areas like accessibility, education, or others," the company said. "The smart devices in our hands are going to continue disrupting the way we think about intelligence. With a swift, lightweight machine learning system like Caffe2go, we are committed to bringing you more awesome AI and AR experiences like getting access to Van Gogh's brush when you shoot a video."
Facebook has also used artificial intelligence to improve virtual reality experiences. Its new computer vision-powered video stabilisation technology can level 360-degree content faster than any other technology. The company's other work includes an advanced speech recognition system that allows digital avatars' lips to move in sync with speaking voices.
Facebook said its next challenge is to develop systems that are more contextually aware. It highlighted how a computer can identify a pizza in an image but cannot determine whether it's vegetarian or not. To complete this task, it has to find pieces of sausage on the pizza, realise "sausage is meat" and then link that to "vegetarian means that there is no meat." AI's current lack of this kind of advanced context prevents this understanding being achieved.
Facebook is already progressing towards contextually aware AI. It recently demonstrated a system that could solve 20 synthetic tasks designed to test simple reasoning. However, it recognised it's still a long way off developing a truly "intelligent" system. The company said it will continue its work in the field because computers with common sense will be more useful machines.
"When computers have common sense they can interact with us in better, more natural ways, from surfacing the most relevant information for us and assisting us with tasks to enabling whole new ways for people to connect," said Facebook Chief Technology Officer Mike Schroepfer. "We’re off to a good start, and I can’t wait to see what tomorrow brings."
More about Facebook, Ai, Artificial intelligence, machine learning, neural networks
 
Latest News
Top News