Facebook Collects 2,200 Hours Of First-Person Footage To Make AI Much More Human
Image via Facebook
Facebook has announced that it has managed to compile a collection, 2,200 hours’ worth, of first-person footage from around the world as artificial intelligence training material in a project called ‘Ego4D’.
With this, the company has set out to equip AI with the capabilities to “interact with the world like we do, from a first-person perspective.” Although AI often learns through being fed existing data like photos and videos, the “next generation” of AI will instead use “videos that show the world from the center of action.”
700 participants consented to their daily lives being recorded in this way, resulting in the astounding amount of information collected by 13 universities and labs across nine countries. Facebook claims that it is 20 times larger than any other collection of footage.
Image via Facebook
Some university partners include the Carnegie Mellon in the US, the UK’s University of Bristol, the National University of Singapore, the University of Tokyo, and the International Institute of Information Technology in India, according to CNBC.
This data and the resultant investigation “is going to catalyze progress for us internally but also widely externally in the academic community and [allow] other researchers to get behind these new problems but now be able to do it in a more meaningful way and at a greater scale,” Kristen Grauman, Facebook lead research scientist, explains to CNBC.
Image via Facebook
“Smarter” and “more useful” AI assistants are being developed as we speak, according to the social media giant. Five benchmark challenges are referenced in its news release, which include episodic memory, forecasting, object manipulation, audio-visual identification, and social interaction.
Image via Facebook
The “egocentric” data collected will, therefore, aim to be able to answer questions like, “where did I leave my keys?” and, “what was the main topic during class?” Futuristic assistants will also be capable of delivering when told “help me better hear the person who’s trying to talk to me in this noisy place.”
To protect participants’ privacy, Grauman stresses that the universities that carried out the data collection had undergone a “pretty intensive and important process to create a policy for proper collection.”
Some steps taken include erasing personally identifiable information from the videos like bystanders’ faces and car license plates. Audio from many clips was also removed.
[via CNBC, image via Facebook]
Also check out these recent news
