Facebook to use "public" videos to train its AI

Facebook has just announced its latest project called Learning from Videos. The project focuses on AI systems that can learn from "publicly available" data. The company says that it will help improve the platform’s content recommendation feature and policy enforcement along with creating a completely new experience. Let’s learn more about the Learning From Videos program in detail in this post.

Facebook "Learning From Videos" AI Project

It is not the first time that Facebook has come with an ambitious AI project. The company has its own facial recognition system that can automatically recognize people in the photos and recommend tags. Back in 2018, the company even started using images from Instagram to improve its AI algorithm. However, it is probably the first time the company is using videos to train its AI system. Since Facebook is working with a self-supervised AI system, the system is less dependent on human and pre-processed data. In fact, you don’t need labelled data sets here as the system can mimic a human’s behaviour of learning. For the system to give reliable results, it needs to be fed with a large volume of data. As mentioned earlier, it will come from the videos that are uploaded and shared publicly by users/pages on Facebook. Also Read: Facebook is reportedly looking to enter the smartwatch industry.

Applications

The company has already experimented with such a system. According to the company, it has been successfully using a similar AI system for recommending Reel’s videos on Instagram. Similarly, the company is planning to use AI for detecting foul languages and hate speech across its platform. According to the company, the usage of a self-supervised AI algorithm has resulted in a reduction of speech recognition errors by around 20%. Facebook is also proposing the use of the system in a multimodal video system to extract key moments from videos. Facebook is also looking to use the Learning from Videos to its upcoming AR glasses. The company has released demos of Project Aria and has already confirmed its partnership with Rayban for Smart AI Glasses. In its latest blog, the company has proposed Smart AR glasses that can capture videos automatically and make certain moments accessible when required using the Learning from Videos AI system.

Potential Controversies

While Facebook has proposed a number of application of the Learning From Videos the company has kept the public in the dark on how it is planning to control and oversee the use of public data. For instance, when The Verge questioned the company about asking users for their consent, the company replied that their data policy allows the use of uploaded content for “product research and development”. Hence, it looks like Facebook will be using public videos without notifying the uploader about it. It could also mean that there will be no option to opt-out of it. Similarly, the company has not outlined its strategy regarding the volume of data that it requires and how it is planning to supervise the whole process to reduce misuse. On the other hand, self-supervised algorithms come with their own downsides. The ones used in the past for content recommendation systems have shown to boost misinformation. A report from MIT Technology Review has shed light on how these kinds of AI systems have boosted misinformation, controversy extremism, and political polarization for the sake of engagement.