More artificial intelligence it is still built on a human labor base. Explore an AI algorithm and you will find something built using data that has been selected and labeled by an army of human workers.
Now, Facebook has shown how some AI algorithms can learn to do useful work with far less human help. The company built an algorithm that learned to recognize objects in images with little help from labels.
Facebook’s algorithm, called Seer (from SElf-supERvised), fed on more than a billion images taken from Instagram, deciding for itself what objects look like. Images with mustaches, fur and pointed ears, for example, were assembled in a pile. Then, the algorithm received a small number of labeled images, including some labeled “cats”. He 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 AI and computer vision. “Making self-supervised learning work is very challenging, and advances in this space have important consequences for improved visual recognition.”
Russakovsky says it is remarkable that Instagram images were not handpicked to facilitate independent learning.
Facebook research is a milestone for an AI approach known as “self-supervised learning,” says Facebook’s chief scientist, Yann LeCun.
LeCun was a pioneer in the machine learning approach known as deep learning, which involves feeding large artificial neural networks. About a decade ago, deep learning emerged as the best way to program machines to do all sorts of useful things, such as image classification and speech recognition.
But LeCun says the conventional approach, which requires “training” an algorithm by feeding it a lot of labeled data, is simply not scalable. “I have been defending this whole idea of self-supervised learning for a long time,” he says. “In the long run, progress in AI will come from programs that just watch videos all day and learn like a baby.”
LeCun says that self-supervised learning can have many useful applications, for example, learning to read medical images without having to label so many exams and x-rays. He says a similar approach is already being used to generate automatic hashtags for Instagram images. And he says that Seer technology could be used on Facebook to combine ads with posts or to help filter out unwanted content.
Facebook research is based on steady progress in adjusting deep learning algorithms to make them more efficient and effective. Self-supervised learning was previously used to translate text from one language to another, but it was more difficult to apply to images than to words. LeCun says the research team has developed a new way for algorithms to learn to recognize images, even when part of the image has been altered.
Facebook will launch part of the technology behind Seer, but not the algorithm itself, because it was trained using data from Instagram users.
Aude Oliva, who heads MIT’s Computational Perception and Cognition laboratory, says the approach “will allow us to perform more ambitious visual recognition tasks”. But Oliva says the size and complexity of cutting-edge AI algorithms like Seer, which can have billions or trillions of neural connections or parameters – far more than a conventional image recognition algorithm with comparable performance – also presents problems. These algorithms require huge amounts of computational power, overloading the available chip stock.
Alexei Efros, a professor at UC Berkeley, says the Facebook article is a good demonstration of an approach he believes is important for the advancement of AI – getting machines to learn for themselves using “gigantic amounts of data” . And, as with most progress in AI today, he says, it builds on a series of other advances that have emerged from the same Facebook team, as well as from other research groups in academia and industry.
More great stories from WIRED