Facebook today announced a major breakthrough: Facebook’s AI, named SEER (SElf-supERvised AI), was able to correctly identify and categorize objects in photos without the aid of humans with a high degree of accuracy.
Yan Lecun, Facebook’s chief artificial intelligence scientist, wants the company’s AI to be able to learn similar to how humans do when they’re babies: through looking at what is in front of them.
“We’d like artificial intelligence systems to learn how the world works by observation because that will have a huge implication,” he told CNET. “It would allow machines to have some level of common sense.”
The group came a lot closer to that goal today. SEER was able to learn from a billion random, unlabeled, and uncurated public Instagram images and from that information was able to identify and categorize the dominant object in photos with an accuracy rate of 84.2%, which outperforms existing self-supervised systems by one percentage point. That may not seem like a lot, but it’s significant enough for the group to be very excited about it.
“SEER’s performance demonstrates that self-supervised learning can excel at computer vision tasks in real-world settings,” Facebook writes in a blog. “This is a major breakthrough that ultimately clears the path for more flexible, accurate, and adaptable computer vision models in the future.”
AI that is able to teach itself how to recognize objects and correctly categorize them would help a variety of Facebook products as well as assit other social networks. Currently, AI is used to rank content in feeds and flag images and videos that may violate the rules of the platform (like hate speech or nudity). AI is also leveraged in cars to help avoid collisions and in medical devices to make diagnosing symptoms faster. The number of possible applications of AI is limitless, which is why a self-supervising version is so exciting.
“The advantage of self-supervised learning is that you can train very big networks and it will still be accurate,” LeCun said.
Images online tend to not be the best quality. They could be blurry, out of focus, or taken at an unusual angle. If an AI can learn to recognize these things and still pick out the subject of the image and adapt to it, that would make for a vastly more useful and flexible system. Additionally, self-supervizing AI can learn without the intrinsic biases that come from human intervention.
For example, there are some studies that have shown that facial recognition systems have a harder time correctly identifying minorities which could be due to the photo sets that feature predominantly white people that are fed into the algorithm. Though LeCun says this idea is “speculative,” the hope is that by removing the errors humans cause in data sets, the end resulting AI may have reduced bias and better final results.
Image credits: Header image via Facebook