Baidu Inc. says it has posted the world’s best results on a closely watched artificial intelligence benchmark. It had a secret weapon, according to researchers at the Chinese search giant: Minwa.
The company’s Minwa supercomputer scanned ImageNet, a database of just over one million pictures, and taught itself how to sort them into a predefined set of roughly 1,000 different categories. This meant learning the difference between a French loaf and a meatloaf, but also trickier challenges such as distinguishing a Lakeland terrier from a wire-haired fox terrier.
Five years ago, the possibility that computers would surpass humans at this work appeared remote. But computers run by Microsoft Corp., Google Inc., and now Baidu have all done better than the best human results in the past few months.
With practice, humans correctly identify all but about 5 percent of the ImageNet photos. Microsoft’s software had a 4.94% error rate; Google achieved 4.8%. Baidu said that it had reduced the error rate further to 4.58%.
The so-called deep learning algorithms that Baidu and others are using to ace these tests have only recently made the leap from academia to Silicon Valley. But they’re starting to have an impact in daily life.
Two years ago, Google used deep learning to dramatically boost Android’s voice recognition system.
Baidu is using an even larger supercomputer to analyze 14,000 hours of speech data to improve the search company’s Chinese- and English-language speech recognition.
“I am very excited about all the progress in computer vision that the whole community has made,” said Andrew Ng, Baidu’s chief scientist. “Computers can understand images so much better and do so many things that they couldn’t do just a year ago,” he said.
In the coming 18 months, Baidu plans to build an even larger machine, one that can perform 7 quadrillion calculations per second. That would be enough to rank the system as one of the 10 most powerful supercomputers in the world, although it performs less complex calculations than the world’s top-ranked supercomputers.
Deep learning is also becoming a sport of kings, increasingly played by deep-pocketed companies that can hire top AI experts, amass gargantuan data sets, and stand up the computing resources needed to analyze them.
“It’s interesting that the top three teams processing ImageNet all appear to be large tech companies with considerable computational resources,” Mr. Ng said. His company drew on supercomputing expertise from China and its Silicon Valley laboratory to build Minwa, he added.
Yann LeCun, a noted artificial intelligence researcher who leads a similar effort at Facebook Inc., said his company was deploying big-time computational resources, although he wouldn’t comment on the size of Facebook’s computer system.
With all the computing firepower being aimed at deep learning, the ImageNet test that Google, Microsoft and Baidu aced is starting to become “passe as a benchmark,” Mr. LeCun said. “People are focusing on much larger data sets and more challenging tasks that involve object recognition, such as object detection and localization,” he said.