Most artificial intelligence it is still built on a foundation of human labor. Look inside an AI algorithm and you will find something built with data that was selected and tagged by an army of human workers.
Now, Facebook has shown how some artificial intelligence algorithms can learn to do useful work with much less human help. The company created an algorithm that learned to recognize objects in images with little help from labels.
Facebook’s algorithm, called Seer (by SElf-supERvised), drew on over a billion images pulled from Instagram, deciding for itself which objects look alike. Pictures with whiskers, fur, and pointed ears, for example, were gathered in a pile. The algorithm then received a small number of tagged images, including some tagged “cats.” It was then able to recognize images, as well as a trained algorithm using thousands of labeled examples of each object.
“The results are impressive,” says Olga Russakovsky, an assistant professor at Princeton University who specializes in artificial intelligence and computer vision. “Making self-directed learning work is very difficult, and advances in this space have important downstream consequences for improving visual recognition.”
Russakovsky says it’s notable that the Instagram images weren’t hand-picked to facilitate independent learning.
Facebook’s research is a milestone for an artificial intelligence approach known as “self-supervised learning,” says Facebook’s chief scientist Yann LeCun.
LeCun pioneered the machine learning approach known as deep learning that involves feeding data into large artificial neural networks. About a decade ago, deep learning emerged as a better way to program machines to do all sorts of useful things, like image classification and speech recognition.
But LeCun says the conventional approach, which requires “training” an algorithm by feeding it a lot of tagged data, just won’t scale. “I’ve been advocating for this whole idea of self-supervised learning for quite some time,” he says. “In the long term, progress in AI will come from shows that just watch video all day and learn like a baby.”
LeCun says that self-supervised learning could have many useful applications, for example learning to read medical images without the need to label as many scans and X-rays. He says a similar approach is already being used to automatically generate hashtags for Instagram images. And it says Seer technology could be used on Facebook to match ads to posts or to help filter out unwanted content.
Facebook’s research is based on constant progress in tweaking deep learning algorithms to make them more efficient and effective. Self-directed learning has previously been used to translate text from one language to another, but it has been more difficult to apply to pictures than to words. LeCun says the research team developed a new way for algorithms to learn to recognize images even when part of the image has been altered.
Facebook will release some of the technology behind Seer, but not the algorithm itself because it was trained using the data of Instagram users.
Aude Oliva, who heads the MIT Computational Cognition and Perception Laboratory, says the approach “will allow us to take on more ambitious visual recognition tasks.” But Oliva says that the size and complexity of cutting-edge AI algorithms like Seer, which can have billions or trillions of connections or neural parameters – many more than a conventional image recognition algorithm with comparable performance – also raises problems. Such algorithms require enormous amounts of computational power, depleting the available supply of chips.
Alexei Efros, a UC Berkeley professor, says the Facebook document is a good demonstration of an approach that he believes will be important to the advancement of AI: making machines self-learn by using “gigantic amounts of data”. And as with most advancements in AI today, he says, it builds on a number of other advancements that emerged from the same team at Facebook, as well as other research groups in academia and industry.
More great stories from WIRED