Big black screens full of code litter a secure corner office in a WeWork building in Montreal, Canada. The monitors are piled in side-by-side, one on top of the other, and there’s barely enough room for the 20 or so research scientists and engineers, who work for the Facebook Artificial Intelligence Research (FAIR) group. “We’re moving to a new office soon,” says Joelle Pineau, head of FAIR’s Montreal lab and an associate professor at McGill University. Pineau’s lab has grown from four people to 20 since it was established a little over a year ago, and it isn’t the only FAIR lab expanding rapidly.
The FAIR group as a whole — tasked with advancing the field of AI — has grown to almost 200 researchers worldwide since it was founded by Facebook’s chief AI scientist Yann LeCun in 2013, and is set to double by 2020. Its mission: to develop the smartest machines possible. The research group, which openly publishes almost all of its work, wants to build AI that can see, hear, and communicate with humans seamlessly. If you have a quick chat with Siri, Alexa, Google or Cortana, you’ll quickly realise that there’s a long way to go. Facebook’s own virtual AI assistant, “M”, was shut down in January.
FAIR’s researchers — now spread across Menlo Park, New York, Seattle, Pittsburgh, Montreal, Paris, London, and Tel Aviv — are focusing on areas such as robotics, computer vision, natural language processing, language translation, and games. The view is that every incremental advance in each of these fields helps towards developing AI with human level intelligence. Recently, researchers at FAIR have taught an AI to create a recipe and list a set of ingredients from simply looking at a photograph of a meal. FAIR researchers are also exploring how AI can be used to speed up MRI scans up by to 10 times.
It all started with a dinner. In 2013, Facebook CEO Mark Zuckerberg realised he needed to develop better AI systems to enable more sophisticated product features on its social network. Initially, he looked into acquiring a discrete London AI lab called DeepMind, which had been backed by tech billionaire and Facebook board member Peter Thiel, as well as Elon Musk and Skype co-founder Jaan Tallinn. But there was another tech heavyweight interested in DeepMind: Google.
Facebook and Google both wanted DeepMind to relocate to Silicon Valley but DeepMind refused, arguing there was more untapped talent on offer in Europe. In the end, Google acquired DeepMind in 2014 for around £400 million. And Facebook opted for building an AI research organisation from the ground up – and invited LeCun over for dinner. The scientist was intrigued. “The next day I visited Facebook and at the end of the day [Zuckerberg] says: ‘Ok, so now can you help us?’”
LeCun, known as one of the godfathers of AI (along with Google’s Geoff Hinton and Element AI’s Yoshua Benjio), didn’t want to leave his home in New York, a city he adores, nor his job at New York University, where he has been a professor since 2003. The 58-year-old Frenchman told Zuckerberg his conditions. “He said ‘Yes’ and I said ‘Ok. Where do I sign?’”
Today, DeepMind is FAIR’s biggest rival and the two constantly battle it out for the most talented people in the industry, offering big salaries in the process. Several top AI researchers at the senior level are paid seven figure salaries, LeCun says. Samim Winiger, a Berlin-based AI researcher, decided to join Google instead. “While Facebook has world class AI researchers and infrastructure, the actual outputs (research papers and deployed projects) are all over the map,” he says – meaning that FAIR is active in practically every area of AI, but not leading in any – making it feel disconnected and not focused.
“Of course we do fight a lot. We have huge battles with [Google] for the best talent,” says Rob Fergus, head of FAIR’s New York lab. “Sometimes we win, sometimes we lose.”
Both groups are considered among the best AI research teams in the world. When FAIR makes a breakthrough, Facebook’s Applied Machine Learning (AML) group looks at how it can use the technology to build a new product or feature for Facebook and Facebook’s other platforms: Messenger, Instagram, Oculus, and WhatsApp.
Zuckerberg sees AI as a vital technology and Facebook already uses AI to power many of the core features on the main Facebook platform. The Facebook News Feed, for example, is underpinned by AI that predicts what content each person may want to see. “Everybody basically has a trained system that has learned their preferences and it uses all kinds of signals to do with the kind of content they like to interact with and the people they like to interact with,” says LeCun.
Google DeepMind, which has over 700 people, has made more headlines thanks to its AlphaGo AI system, which successfully defeated the top human Go player in the world, Lee Sedol in 2016. But it’s arguably quite difficult for DeepMind to push new technology into Google products because of the geographic isolation, the rivalry between them and Google Brain, and the isolation of research and code, says LeCun. “As a consequence, several years from now, someone will ask the question at Google: Why are we spending all this money? This is not a situation I would like to be in.” (Google has used DeepMind’s AI in its data centers and Android operating system).
For Facebook, the applications for AI are perhaps more obvious. AI underpins Facebook’s very precise ad-targeting software, which makes it billions of dollars every year, in addition to other user features like automatic photo tagging and automated translation services. Elsewhere, Facebook is attempting to use AI software to identify fake news that’s going viral and harmful content (such as hate speech videos posted by terrorist organisations). It’s also trained an AI system to spot posts from people who might be contemplating suicide so that it can reach out to them and offer help.
Most of Facebook’s existing AI features have been built using a form of machine learning known as supervised learning. “Over the last five years it made complete sense to just focus on supervised learning because there were so many applications where it was economically feasible to collect data, have it labelled, and train a control net to solve the problem in whatever it is … translation of French and Chinese to English and vice versa, or classifying the topic of a piece of text, or recognising images, or detecting objects in images etc,” says LeCun.
But now Facebook wants to develop applications where the data doesn’t always exist. If Facebook wants to translate Pashto to Swahili, for example, there’s no parallel text for machines to learn from, so Facebook needs to use alternative methods. AI researchers have recently been trying to tackle some problems that supervised learning can’t deal with, a technique known as reinforcement learning — a popular branch of AI that gives agents positive and negative rewards for their behaviour in order to train them.
This has proved successful for helping machines to master Atari games such as Space Invaders and complex board games like Go, but the trial and error method has its limitations. “To get a reinforcement learning system to learn to play Atari games it needs to play for the equivalent of about 100 hours with the best algorithms to reach a level of performance that a human can reach in a few minutes,” LeCun says. “So it tells you that we’re missing something.”
One of the factors here is that humans are able to learn certain tasks quickly and safely because they have a degree of background knowledge. “The question I’m interested in is how did they get to learn this?” LeCun adds. “How is it that if you wanted to use reinforcement learning to get an autonomous car to drive, it would have to kill 10,000 pedestrians and run off a cliff a few thousand times before it figures out how not to do it? Whereas we seem to be able to learn how to drive in about 30 hours of training without any accident. So what’s the difference? That’s the big question.”
His hypothesis is that people have predictive models of the world, having acquired a lot of background knowledge about how the world works – in other words, common sense. “That allows us to predict beforehand if we turn the wheel to the left and the car runs off the cliff, nothing good is going to come out of this. So we can plan ahead and don’t actually do it. Whereas a classical reinforcement learning system has to actually try it to realise it’s bad,” says LeCun.
And that’s what Facebook is trying to do next – to build machines that can run models of the world. There’s a complication, though, says LeCun: the world is not entirely predictable.
More great stories from WIRED
– Why you don’t need to skip sleep to be successful
– Scientists explain why Hyperloop is so dangerous and difficult
– The suspicious Facebook page pushing Brexit ads to millions
– Why India is turning off the internet to fight fake news
– These photos show the devastating impact of human progress
Don’t miss out. Sign-up to WIRED Weekender to get the best of WIRED in your inbox every weekend