Facebook is in the process of developing its own chips to help it filter video contents, said chief AI scientist Yann LeCun at the Viva Technology industry conference in Paris.
The social media giant concluded that conventional methods are no longer sufficient as they require too much energy and compute power, according to LeCun.
Using systems based on conventional computers to monitor every video, live and recorded, would require Facebook "a huge amount of compute power," and that cost a lot in both bills and environmental impact.
Facebook uses Intel CPUs for many of its AI services, to do things like remove content by terrorist organizations like ISIS. And a transition to a more specialized chips that are more energy-efficient could help the company in filtering videos a lot faster.
And all that in real-time.
"Let’s imagine someone uses Facebook Live to film their own suicide or murder. You’d like to be able to take down that kind of content as it happens," explained LeCun.
"There’s a huge drive to design chips that are more energy-efficient for that," said LeCun. "A large number of companies are working on this, including Facebook. You’ve seen that trend from hardware companies like Intel, Samsung, Nvidia. But now you start seeing people lower in the pipeline of usage having their own needs and working on their own hardware."
With the rapidly increasing amount of unlabeled data like images and videos in the real world, chipmakers like AMD, Intel, Qualcomm and Nvidia are doing all they can to keep up with the pace.
But large tech companies are looking to supply themselves, because their needs are more selective. For that matter, they want to depend less on those chipmakers.
For example, Google has its own machine-learning hardware that better fits its own requirements, allowing it to accelerate its AI training and deployment with its tensor processing unit (TPU). Google has also launched the even more powerful TPU 2.0, which prioritized training over inference; then the TPU 3.0, which promises another big increase in performance.
Microsoft also developed one: a field programmable gate array (FPGA) chip for its Project Brainwave.
Others also have their own requirements and hardware to do the job. All that with a goal of speeding up AI process to benefit using their own means.
More and more smartphones are equipped with powerful chips that let users take advantage of speech recognition, augmented and virtual reality, as well as image and video processing directly on their devices, said LeCun. That trend will grow, and it’s prompting more software companies to think about hardware, he said.
The ability to filter contents in real-time has had some drawbacks. There are times when AI cannot interpret an image correctly, and flagging it. However, the technology is not a bad idea since it should allow Facebook in taking down violent videos, hate speech, extremist propaganda, fake accounts and more, all that in speed faster than any human.
Using AI and algorithms, Facebook can teach its systems to understand the patterns of such posts, enabling it to take them down those contents as soon as they're shared.
What this means, fewer people would be aware of those videos' existence. Proving this should make Facebook a better social media, and also a better place for people to share things that matter to them the most.
"Facebook has worked on hardware before: It makes its own server design, motherboards, its own communications chips for data centers,” LeCun said. "So this is not completely new for Facebook."