Baidu, a China-based search company, said that it developed a supercomputer that ups the ante for artificial intelligence.
According to Baidu, the supercomputer was able to identify 95.42 percent of images in a batch of 1 million pictures, beating software by Microsoft, Google and even humans in the endeavor.
Baidu's supercomputer, which is named Minwa, features 72 powerful processors and 144 GPUs, which are high-performance processors that usually deal with visual information.
To be able to test the capabilities of Minwa, Baidu released the supercomputer onto ImageNet, which is a database that houses over 1 million images. The supercomputer then taught itself how to arrange the pictures into about 1,000 pre-defined categories. This task required the supercomputer to analyze and find differences in similar-looking images, such as two pictures of dogs but of different breeds.
To carry out the assignment, Minwa used a "neural network" for the supercomputer to be able to recognize the pictures, then trained its software with high-res versions of the images to understand the characteristics that it needed to look for.
Minwa also received the images in altered forms, with the pictures being cropped and distorted in different ways, to make sure that the artificial intelligence was picking up the necessary details of the images. This training developed the supercomputer to recognize images even when the picture was printed out, viewed at a slanting angle or photographed again.
"Our company is now leading the race in computer intelligence," said Baidu scientist Ren Wu, whose claim is supported by the fact that Minwa beat the image-recognizing record of Google, which had a 95.2 percent success rate.
Wu added that the computational power held by Minwa could land it among the 300 most powerful computers all over the world, if only it was not developed for deep learning purposes.
Deep learning is seen as one of artificial intelligence's most potent types, involving algorithms that have just found their way to the technology industry from the academe. It is also seen to be working best with bigger networks and data sets.
According to Wu, previous learning techniques of artificial intelligence did not show improvements when going beyond certain points.
Minwa uses deep learning with an artificial neural network that has billions upon billions of connections, which is hundreds of times more compared to any previously built neural network.
Wu, along with colleagues Shengen Yan, Yi Shan, Qingding Dang and Gang Sun, published the results of the image recognition experiment using Minwa in a paper titled Deep Image: Scaling up Image Recognition.
Photo: Julien Gong Min | Flickr
