In an effort to oust Google from its the top spot in image recognition, Baidu, Inc. has dedicated a supercomputer to what it claims to be the world’s most accurate computer vision system.
Using the ImageNet object classification benchmark, the Chinese search engine company claims that their system managed a 5.98 percent error rate, which compares to Google’s rate of 6.66 that won it the top spot in the 2014 ImageNet competition.
Humans, meanwhile, are estimated to have a 5.1 percent error rate on the ImageNet tests. So if Baidu’s results prove accurate, the company has succeeded in not only besting Google, it’s closer to humans in its ability than it is to other AI technologies.
Taking center stage in this achievement is the Minwa supercomputer that Baidu built in house for its Deep Image computer vision system. Comprised of 72 Intel Xeon E5-2620 processors and 144 Nvidia Tesla K40m GPUs, the system packs 432 cores for a theoretical peak of .6 petaflops. Minwa also employs FDR InfiniBand to help overcome interconnect bottlenecks that have hindered deep learning algorithms in the past.
According to Baidu, this enabled researchers to work with more sophisticated images to develop or “train” the system. Rather than a 256×256-pixel image, higher-res 512×512 graphics were chosen and then altered through various filters such as color-casting, vignetting and lens distortion that helped to teach Deep Image to see past common editing tweaks.
And this isn’t the first foray Baidu has taken in the pursuit of Google. Deep Speech, the company’s speech-recognition system went public in December in anticipation of greater adoption of voice and image search technologies.
Both Deep Image and Deep Speech fall into a greater category of machine learning algorithms called deep learning, which encompasses abilities such as natural language processing and audio recognition that come to humans naturally while tending to elude computers.
Meanwhile, Google still holds the official ImageNet record and recently announced new Google Translate features that will take advantage of image recognition technology. The next ImageNet competition is scheduled to take place in late 2015.
Additional details about Deep Image, Minwa and Baidu’s deep learning research can be found in the company’s official paper on the topic.