“Microsoft Researchers Win ImageNet Computer Vision Challenge”, 2015-12-10 ():
Microsoft researchers on Thursday announced a major advance in technology designed to identify the objects in a photograph or video, showcasing a system whose accuracy meets and sometimes exceeds human-level performance.
Microsoft’s new approach to recognizing images also took first place in several major categories of image recognition challenges Thursday, beating out many other competitors from academic, corporate and research institutions in the ImageNet and Microsoft Common Objects in Context (MS COCO) challenges.
…The researchers say even they weren’t sure this new approach was going to be successful—until it was. “We even didn’t believe this single idea could be so important”, said Jian Sun, a principal research manager at Microsoft Research who led the image understanding project along with teammates Kaiming He, Xiangyu Zhang and Shaoqing Ren in Microsoft’s Beijing research lab…Sun said researchers were excited when they could successfully train a “deep neural network” system with 8 layers 3 years ago, and thrilled when a “very deep neural network” with 20–30 layers delivered results last year. But he and his team thought they could go even deeper.
For months, they toyed with various ways to add more layers and still get accurate results. After a lot of trial and error, the researchers hit on a system they dubbed “deep residual networks.”
[compare this description of the discovery with the one in the published paper, et al 2015; note eg. 1989, invention of Transformers]
…The major leap in accuracy surprised others as well. Peter Lee, a corporate vice president in charge of Microsoft Research’s NExT labs, said he was shocked to see such a major breakthrough. “It sort of destroys some of the assumptions I had been making about how the deep neural networks work”, he said.
…None of this means that computers are getting smarter than humans, in a general way. The researchers say what it shows is that computers are getting very good at very narrow tasks, like identifying images in a database. Still, that has big implications for how computers could eventually help people in any number of ways, like recognizing the difference between a tree and a car in your side view mirror or the frustrating task of sorting through photos for specific things, like a great picture of your dog.
“We don’t believe we’re anywhere close to the limit of the ultimate improvement in data classification accuracy for any of these tasks”, Lee said.