Facebook’s AI software is getting better at detecting inappropriate and dangerous content on its site.
To demonstrate this, Facebook’s chief technology officer, Mike Schroepfer, put a photo of marijuana and broccoli up next to each other on stage at the firm’s F8 developer conference.
He said the social media giant’s ‘state of the art’ computer vision is now capable of detecting the difference between food and drugs in a matter of seconds.
The hope is that this kind of software can help Facebook clean up the site more quickly and with greater accuracy than humans ever could.
Scroll down for video
Facebook’s CTO Mike Schroepfer (pictured) put a photo of marijuana and broccoli up next to each other on stage at the firm’s F8 developer conference to show how its AI is getting smarter
Facebook’s AI was 93.7 percent sure that it was looking at an image of marijuana, while it was 88.3 percent sure it was looking at an image of tempura broccoli.
But it wasn’t always this smart.
‘The most surprising thing is in 2014, computer vision, the study of AI to understand images, was so far away from where it is today that it had no idea what [marijuana] was,’ Schroepfer said.
Over time, by supplying its AI with more examples, the system has ‘adapted’ to be able to pick up on more signals of content that violates Facebook’s policies.
For example, people may use code words or obscured imagery to try to fool content moderators, but Facebook’s AI can still flag these posts.
Facebook’s AI was 93.7 percent sure that it was looking at an image of marijuana, while it was 88.3 percent sure it was looking at an image of tempura broccoli, the company said
The AI not only looks at images, but also obvious indicators like text, in addition to coded language, and accounts behind the content, when deciding whether it violates its policies
‘As soon as our adversaries adapt…within an hour or a few hours, we’re able to catch very similar things,’ Schroepfer said.
In one example, he showed a post where a user was advertising a ‘potent batch’ of rice krispie treats.
This coded language of a ‘potent batch,’ coupled with other signals, like the account behind the post, in addition to other context, have allowed Facebook’s AI to take down more offending posts.
Going forward, Facebook said it will be leaning on AI more and more as a means for curtailing misinformation, election interference, hate speech and other problems.
‘AI is our best bet to keep the community safe across our platforms,’ Manohar Paluri, a director of artificial intelligence at Facebook, said on stage.
The company is now working toward a future where its AI can understand content without human supervision.
Facebook’s AI software is getting better at detecting inappropriate and dangerous content on its site. The company’s hope is that this software can help it clean up the site more quickly
Facebook’s AI is getting better at flagging and removing problematic content. In a recent three-month period, it was able to remove a billion spam accounts and 700M fake accounts
As part of this effort, it has devised a method called ‘self-supervised learning,’ where a computer can continuously learn on its own by studying labeled and unlabeled data.
These systems have already had clear results, with Facebook taking down more than a billion spam accounts, 700 million fake accounts and millions of pieces of content that contain nudity and violence over the course of a three-month period.
The next step is to make its AI just as good at flagging inappropriate content in videos as it can in photos.
‘It is clear that video understanding is in its infancy,’ Paluri explained.
Facebook was criticized for its failure to quickly remove a livestream of the Christchurch shootings, which left 50 worshippers dead and wounded 50 more at two mosques last month.
In that event, Facebook’s AI was not able to detect the violent video at all.
Facebook is also looking at the need for safety in spaces outside the real world.
Facebook is also looking at the need for safety in spaces outside the real world, particularly as more and more users are experimenting with augmented and virtual reality
Facebook has created ways for users to block, report and mute others in VR spaces to prevent harassment, Lindsay Young, a member of Facebook’s AR and VR team (pictured), said on stage
As more and more people are using augmented and virtual reality applications, there’s been an increased need for policies and software that ensures their safety.
Facebook has created ways for users to block, report and mute others in VR spaces to prevent harassment, Lindsay Young, a member of Facebook’s AR and VR team, said on stage.
One feature, called a ‘safety bubble,’ lets users create a virtual boundary for themselves in VR environments, so that abusive or annoying players can’t come ‘closer than you like’ in your personal space.
If an avatar enters your safety bubble, the game will automatically block them.
Facebook has also installed live moderators in some virtual reality spaces, akin to body guards at a concert, that can enforce policies on the spot, Young said.
‘We believe VR is the next frontier of human interaction so it’s an area where we need to be incredibly mindful when creating it,’ she said.
‘As we bring people together in VR, they should have access to tools that make them feel safe.’
The firm is now considering ways to protect users’ headsets and personalized avatars from being abused, by using facial recognition and fingerprint verification technology as a way to authenticate your identity when logging into a VR application.
HOW DOES FACEBOOK PLAN TO IMPROVE PRIVACY?
In a March 6 blog post, Facebook CEO Mark Zuckerberg promised to rebuild based on six ‘privacy-focused’ principles:
- Private interactions
- Reducing permanence
- Secure data storage
Zuckerberg promised end-to-end encryption for all of its messaging services, which will be combined in a way that allows users to communicate across WhatsApp, Instagram Direct, and Facebook Messenger.
This he refers to as ‘interoperability.’
He also said moving forward, the firm won’t hold onto messages or stories for ‘longer than necessary’ or ‘longer than people want them.’
This could mean, for example, that users set messages to auto-delete after a month or even a few minutes.
‘Interoperability’ will ensure messages remain encrypted even when jumping from one messaging service, such as WhatsApp, to another, like Instagram, Zuckerberg says.
Facebook also hopes to improve users’ trust in how it stores their data.
Zuckerberg promised the site ‘won’t store sensitive data in countries with weak records on human rights like privacy and freedom of expression in order to protect data from being improperly accessed.’