Social-Media
Facebook's next major AI project is to train its machines on public videos submitted by users.
One of the most difficult challenges — and greatest possible breakthroughs — in the field of machine learning is teaching AI systems to understand what's going on in videos as thoroughly as a person can. Facebook revealed a new initiative today that it hopes will give it an advantage in this critical work: training its AI on public videos posted by Facebook users.
Access to training data is one of the most strategic advantages in AI, and tech giants like Facebook, Google, and Amazon have been able to push ahead in different areas by gathering this resource from millions and millions of their users. Although Facebook has already used billions of images from Instagram to train machine vision models,Although Facebook has already educated machine vision models on billions of photos collected from Instagram, it has yet to announce similar video comprehension ventures.
By learning from global streams of publicly accessible videos covering nearly every country and hundreds of languages, our AI systems can not only boost accuracy but also adjust to our fast-paced environment and understand the nuances and visual cues across cultures and regions, the company wrote in a blog post. Learning from Videos is part of Facebook's broader efforts toward creating machines that learn like humans, according to the company.
According to Facebook, the resulting machine learning models would be used to develop new content recommendation systems and moderation tools, but they could be used for much more in the future. Artificial intelligence that can understand video content may give Facebook unparalleled insight into its users' lives, enabling them to evaluate their hobbies and interests, brand and clothing preferences, and a slew of other personal data. Of course, Facebook already has access to this data through its current ad-targeting operation, but being able to parse video using AI will add an extremely rich (and intrusive) data source to its stores.
Facebook has remained tight-lipped about its ambitions for AI models educated on user videos in the future. The company told The Verge that such models could be used for a variety of purposes, including captioning videos and designing advanced search features, but it did not respond to a query about whether they would be used to collect data for ad targeting. When asked whether users had to agree to their videos being used to train Facebook's AI or whether they could opt out, the company simply stated that users' uploaded content can be used for "product research and development" in its Data Policy. Facebook also declined to answer questions about how much video would be obtained for AI training.
Facebook also didn't respond to questions about how much video would be collected for AI training and how access to this data by the company's researchers would be controlled.
However, in its blog post announcing the project, Facebook did mention one potential future application: using AI to recover "digital memories" captured by smart glasses.
This year, Facebook plans to introduce a pair of consumer smart glasses. The device's specifications are sketchy, but it's possible that these or future glasses will have built-in cameras to catch the wearer's perspective. If AI systems can be equipped to understand video content, users will be able to search for it.
Facebook also didn't respond to questions about how much video would be collected for AI training and how access to this data by the company's researchers would be controlled.
However, in its blog post announcing the project, Facebook did mention one potential future application: using AI to recover "digital memories" captured by smart glasses.
This year, Facebook plans to introduce a pair of consumer smart glasses. The device's specifications are sketchy, but it's possible that these or future glasses will have built-in cameras to catch the wearer's perspective. If AI systems can be equipped to understand video content, users will be able to search for it.
Users would be able to search for past recordings if AI systems can be trained to understand video content, similar to how many photo apps enable people to search for specific places, items, or people. (This is, interestingly, knowledge that has frequently been indexed by AI systems that have been trained on user data.)
People should be able to remember specific moments from their large bank of digital memories just as easily as they catch them, Facebook says, as capturing video with smart glasses becomes the norm. It uses the example of a user searching for Show me every time we sang happy birthday to Grandma before being served relevant clips. As the firm points out,
As the company points out, such a search will necessitate AI systems learning to match the term ‘happy birthday' to cakes, candles, people singing various birthday songs, and more, as well as creating links between different types of data. AI, like humans, will need to comprehend complex concepts involving various forms of sensory information.
