Facebook's latest AI project learns from videos and improves products
Social media giant Facebook announced a project called Learning from Videos recently. The project has multiple objectives, including improving Facebook's Artificial Intelligence (AI) engine, content recommendation, and policy enforcement. The project stands out as it doesn't rely on manually labeled data and learns like a human does. Manually labeled data is the weakest link in AI and Machine Learning (ML) training.
Manually labeling data for ML training is tedious and error-prone
Manual labeling of data is where a human classifies data into categories and assigns it identifiers called labels. The data is then fed to ML algorithms as data sets. The algorithms then learn to label similar instances of the data from newer data sets. However, the process is tedious, time-consuming, and prone to human error making it the weakest link in the chain.
Learning from Videos suggests better Instagram Reels to you
Learning from Videos will use publicly available videos on Facebook to understand and learn audio, textual, and visual representations of data. Facebook has deployed the project to improve Instagram Reels recommendations. The enhanced algorithm will suggest videos of people doing the same dance to the same music as the Reel you're watching. We believe the project could also be behind Instagram Stories' auto-captioning feature.
The project could help you find digital memories more conveniently
Facebook is working on a smart glasses project as well. The company's announcement suggested that Learning from Videos could allow AI to help you find digital memories, including those captured by the smart glasses. For example, Facebook says you could ask the system to show you "every time we sang happy birthday to grandma" and then view those memories.
Facebook is sampling videos from all around the world
The project is challenging as algorithms have difficulty understanding and classifying images and video. This is due to varying lighting conditions, difficult subject tracking, and languages abruptly changing mid-video. The project is even more challenging without manual labeling. Facebook says the project is sampling videos in multiple languages from almost every country to "recognize visual cues across different cultures and regions."
Facebook's monetization model, poor data handling track record cast doubts
While Facebook has leveraged AI and ML technologies, Learning from Videos shows that Big Tech will stop at nothing to use your data to train AI under the pretext of convenience. The concern is partly due to Facebook's poor track record handling user data and the risk associated with monetizing AI tools. Advertisers getting access to videos opens a whole new can of worms.
