There is a running joke in China about the image search engine of Google and that of its Chinese counterpart Baidu. The joke says if you search for an image of Jupiter's moon Io, Google will show you a whole lot of other images of Io. Baidu will show you a load of pancakes. That might be just a joke, but now Baidu says its image recognition technologies have achieved higher accuracy than Google.
Baidu says its supercomputer Minwa is robust in identifying subjects from images despite extreme transformation. [Photo courtesy of Baidu] |
Baidu published a paper this week to explain the new record achieved by its supercomputer Minwa. It reported a top-5 error rate of 4.58 percent in recognizing images, for example, telling a specific breed of dog.
The top-5 error rate is measured by analyzing how often the system's top five guesses for a given image miss the right answer.
Previously, Google held the best record with a top-5 error rate of 4.82 percent, while Microsoft managed to make it to 4.94 percent. These artificial intelligence systems all did better than humans, as the average error rate of humans is 5.1 percent.
Baidu says its supercomputer Minwa is powered by 72 Intel Xeon processors, and has 144 graphic processing units. Baidu scientists have been training it to learn essential characteristics of subjects, and so Minwa has learned to recognize a subject even if the image is distorted or colored.
In the paper, Baidu offers test results to show how powerful Minwa is when it recognizes subjects from an image off a printed photo, even if the photo is held at an oblique angle.
Baidu says it has earlier used a similar approach for speech recognition, and "also achieved state-of-the-art results."