Shining light on Facebooks AI strategy

thumb image

In a speech today at Web Summit, Facebook CTO Mike Schroepfer laid out a vision for the role artificial intelligence and machine learning will play in the companys ambitionsto improve global connectivity, technology accessibility, and human computer interaction.

People want to stay connected and close to other people, so whatever is the best current technology to deploy that is the business we want to be in, saidSchroepfer.

Largecompanies like Facebook play an incredibly important role in the artificial intelligence and machine learning ecosystem. Their sheer size and ability to corner the market on talent makes almost every strategic decision they makean industry-wide declaration.

Connecting everything to everything

Despite setbacks, like the explosion of Facebooks satellite aboard a SpaceX Falcon 9 earlier this summer, the companyremains steadfastin its goals to better connect the world. It ispursuing a number of infrastructure projects, made possible by machine intelligence, aimed at augmenting both urban and suburban internet connectivity.

Project Aquila, which has been well documented, is designed to bring connectivity to suburban areas. The carbon fibre, solar powered, planes fly in a constellation at altitudes higher than a commercial airliner.Planes remain in communication with each other via laser and communicate with the ground via radio frequency. Facebook is using trained neural networks to identify population centers across Africa that could be targeted by the planes.

For urban connectivity, Facebook is using LiDAR, the same technology behind many self-driving car projects, to map out utility polls. Identifying the location of these polls is critical to optimizing connectivity. Teams at the company are building virtual network graphs on top of the data to createthe best data mesh.


Using LiDAR data to enable urban connectivity in San Jose

Art first, AI revolution later

For consumers,the easiest way to visually interact with Facebooks ML work is through Style Transfer. Built on top ofa new mobile deep-learning platform called Caffe2Go, the feature lets users capture artistically stylized video footage in real-time. Chris Cox, Facebooks chief product officer, previously demoed the technology at WSJDLive showing the audience how they can make their own content look like it was created byVan Gogh himself.

When engineers first began to build Style Transfer, the best they could do with the computational resources of a smartphone was to stylize a small frame within a full-frame video. The version we got to see not only altered the entire frame in real time, but didnt display any noticeable signs of lag. Its set to be integrated into the FB app soon and is already being tested in a few countries.

Caffe2Go wont remain limited to Style Transfer it holds the key to deploying convolutional neural nets across Facebooks suite of mobile apps.

For anything we build on the server, we now have a vehicle to ship itto mobile devices, noted Schroepfer.

Bringing machine intelligence to the smartphone is just step one. For the time being, the world still doesnt have an answer for training neural networks on mobile devices. The prospect, coupled withFacebooks long-term roadmap, makes a compelling argumentfor a future wherethe average person coulddesign and train a customneural net on their own smartphone for daily use.

Betting big on virtual reality

Using todays cutting edgetechnology, researchers still had to generate hundreds of artistic stylizations and painstakingly hone in attributes to deliver the best possible Style Transfer experience. Withcontinued research, Facebook hopes to someday leverage our own facial expressions to give us another way to interact with its technology.

Near term, this could manifest itself as a surprise filter being laid over a selfie takenwith wide-eyes, but improvements in facial tracking couldsomeday let use seamlessly share our emotional state with the technologies around us perhaps most interestingly while invirtual reality.

Of course, all of this becomes more interesting once corded VR is no more. The next set of VR problems will require leaps forwardin computer vision. Facebook has already rolled out new stabilization technology for 360 videos, but there is more to be done. Related technologies likeinside-out tracking and speech recognition will also improve the realismof VR experiences.

We have talked about doing things in the long term in augmented reality, added Schroepfer. We would absolutely build devices for it because I think that will be a way you will communicate with AIs in a real-time basis.

Dont count sheep, label training sets


Image captioning struggles to identify that the plane in the right image is crashing

Even with a concretevision about how investment in artificial intelligence can bolster Facebooks products and services, its important to remember that our cutting edge AI research leaves a lot to be desired.

Machine intelligence thrives on patterns, and unfortunately our world is full ofan almost limitless number of outliers. Regardless of the barriers, Facebook has little choice in prioritizing AI as its competitors pour billions into beating the company to the next great breakthrough though the mere fact that everyone is all in on the race is perhaps what makes it so interesting.

Read more: