Facebook announces AI that learns from videos

As a platform hosting a great deal of content, Facebook developers are creating an AI project called Learning from Videos that can learn from publicly available videos uploaded to the website. This AI aims to use audio, textual and visual data to add to its lexicon of content from users spanning the globe.

The goals of this new artificial intelligence center on providing content recommendation for users, content policy enforcement and enhancing the ability of AI to learn like humans. In fact, the AI would ideally be able to recognize any type of video content, as opposed to specific content curated by expert developers.

Recently, AI systems have advanced in their recognition of language, speech and vision. Such increased sophistication helps AI to rely less on procured datasets to build their knowledge base and more on everyday experiences like humans do.

Now, Facebook developers hope to use Learning from Videos as a way to help the platform absorb a vast array of data from different cultures and regions of the world. In this way, they seek to create an AI capable of anticipating user habits while navigating Facebook.

In order to ensure ethical deployment of Learning from Videos, the AI team has adopted transparent privacy policies regarding user content. By implementing guidelines surrounding user trust and safety at the infrastructure level, developers hope to make content creators feel secure using this new feature. Security measures work based on technical safeguards throughout the data lifecycle.

Facebook has already had success with semi- and self-supervised AI products. Indeed, self-supervised learning products already show a 20% reduction in speech recognition errors. Mitigating such errors could improve auto-captioning as well as decrease instances of hate speech by flagging problematic material present on the platform.

The first example of this feature has gone live in Instagram Reels' recommendation system. Reels works as a great proof of concept for self-supervised learning AI, as the system can analyze common themes shared among trending videos. At this point, developers are equipping Learning from Videos to suggest content related to trending videos while filtering out duplicate material users have already viewed. This sort of discretion allows users the kind of nuance needed to save time browsing Facebook videos while enabling the AI to learn the difference between similar pieces of content.

Currently, the next major challenge arises with programming the Learning from Videos feature to draw upon memorized audio and visual input and then correlate the two according to a common theme.

© 2021 Science X Network