Facebook Builds AI That Learns to See Using One Billion Instagram Photos
Facebook develops a new artificial intelligence program named SEER that learns to identify objects by analyzing one billion random, unlabeled Instagram images. The self-supervised learning model outperforms existing AI systems and promises future improvements in accessibility and content moderation.
Facebook announces a major breakthrough in artificial intelligence with a new computer vision program called SEER. The company trains this AI system by feeding it over one billion public, unlabeled images from Instagram. Unlike traditional AI models that rely on carefully curated and labeled data sets, SEER uses a self-supervised learning approach to figure out what objects are in the photos entirely on its own.
The AI program achieves an impressive 84.2% classification accuracy score on the standard ImageNet object recognition test. This high score shows that SEER outperforms existing AI models, proving that machines learn effectively from random, uncurated real-world data. Facebook researchers state that this breakthrough clears the path for more flexible, accurate, and adaptable computer vision models in the future.
Although SEER remains a research project, Facebook sees broad potential applications for the technology. The company suggests the AI improves automatically generated text descriptions for visually impaired users, enhances item categorization on Facebook Marketplace, and helps keep harmful images off the platform. However, the massive use of public Instagram photos raises inevitable questions about user privacy and data collection practices.