Facebook will train its AI by using public videos
The company says this will improve accuracy in artificial intelligence systems and help create new applications.
Facebook on Friday said it's launching a project that'll automatically learn audio, textual and visual representations from public videos uploaded to the site. The social network says the project, called Learning from Videos, will help improve its core AI systems for uses like content recommendation and policy enforcement, and will enable new applications.
"By learning from global streams of publicly available videos spanning nearly every country and hundreds of languages, our AI systems will not just improve accuracy but also adapt to our fast-moving world and recognize the nuances and visual cues across different cultures and regions," Facebook wrote in a blog post. "And by helping AI researchers break away from the reliance on labeled data, we can improve AI-powered products and create entirely new experiences."
Facebook says self-supervision could help AI systems automatically learn to find videos that match search phrases like "Show me every time we sang happy birthday to Grandma," for example. This involves teaching systems to match "happy birthday" with things like candles, cakes and people singing birthday songs.
The social media giant, which has come under fire for several privacy and security issues, says it's placing an emphasis on enforcing privacy using "automated solutions." It notes that already, semi- and self-supervised learning on public Facebook videos has improved its computer vision and speech recognition systems.
Within six months of creating a "state-of-the-art, self-supervised framework for video understanding," Facebook says, it's deployed an AI model for Instagram Reels' recommendation system. The company added that early experiments have shown applying self-supervised learning to videos leads to a 20% drop in speech recognition errors, which could improve applications like auto-captioning and tasks that help flag hate speech.
Facebook has been trying to level up its AI efforts lately. Earlier this month, the company's AI team said it achieved a "breakthrough" when its self-supervised computer vision model known as Seer was able to learn from a billion random, unlabeled and uncurated public Instagram images. Seer was then able to correctly identify and categorize the dominant object in photos with an accuracy rate of 84.2%. Facebook also said late last year that thanks to improvements to its AI technology, it's catching more hate speech before users report it.