Feb 24, 2022 - Technology

Meta plans AI-driven universal translator

screenshot of Mark Zuckerberg

Meta CEO Mark Zuckerberg, speaking at an online event Wednesday. Screenshot: Axios

Facebook parent company Meta on Wednesday detailed several research efforts that it insists are key to its project of building a metaverse.

The big picture: While some of the work, like a new type of haptic gloves, is closely tied to virtual reality, a number of the efforts Meta talked about on Wednesday predate the company's metaverse push and have wide use beyond it.

Of note: Meta touted self-supervised machine learning as a key technology to helping create and make sense of a future with many digital worlds.

  • Unlike other approaches, AI systems of this type learn the rules of the realm they are trying to understand through observation, rather than based on human classification.

What they're saying: "The road to the metaverse goes through AI," Meta AI research managing director Antoine Bordes told Axios.

  • Bordes also said that computer vision systems based on self-supervised AI have the potential to be less biased, since humans aren't adding their own labels to the data.

Among the projects that Facebook touted was a language translation system capable of converting texts directly among 100 languages in real-time without needing to use English as an intermediary step.

  • That means being able to solve for 10,000 language-to-language pairings and allowing the technology to go "beyond the populations that have been well served with technology," Bordes said — something that wasn't possible two years ago.
  • Overall, Meta says that currently more than 20% of the world's population uses a language not covered by commercial language translation systems. "We don't have to wait for the metaverse to happen to have impact," Bordes said.

Between the lines: Facebook tried to paint its work as a contribution to the broader field of artificial intelligence. The company pledged to be transparent about its research as well as to share some of the code through open source releases.

Go deeper