UPDATED 10:31 EDT / JANUARY 16 2015

Baidu CEO Robin Li NEWS

China’s Baidu builds its own supercomputer to beat Google at image search

Baidu CEO Robin Li

Baidu, Inc. is upping the ante in its fight with Google for image-recognition supremacy with what it hails as a record-setting computer vision system capable of recognizing different variations of the same image better than any other artificial intelligence on the planet. The secret? A dedicated supercomputer.

The Chinese search giant’s homegrown neural network runs on a specially assembled cluster of 36 Linux servers each packing two six-core Intel Xeon E5-2620 processors, which can individually support up to 12 threads at a peak clock speed of 2.5Ghz. But the main source of the supercomputer’s image scanning power are the four Nvidia Tesla K40m graphic processing units included in every node, which manage a combined maximum of 617 trillion floating-point operations per second.

That’s roughly 20 percent more than the amount of computational capacity the US National Oceanic and Atmospheric Administration (NOAA) had at its disposal until a recent upgrade, horsepower that the researchers at Baidu employed to increase the quality of the images fed to their model for training purposes. The reasoning behind the decision is surprisingly straightforward.

The overall goal behind their investment in deep learning is to provide more accurate results for a growing number of image-based searches, said Andrew Ng, chief scientist at Baidu in an interview with Bloomberg. This particular supercomputer was built for the sole purpose of training the model.

According to the official paper describing the experiment that GigaOm picked up on Wednesday, the 256×256-pixel image resolution most commonly used for computer vision often leads to situations in which smaller object lose too much resolution for the system to recognize. Increasing the level of detail preserves more information and thereby makes it possible to exploit the added level of detail for better accuracy and image identification.

Baidu says that the use of higher-resolution images enabled its system to reach an error rate of 5.98 percent on the ImageNet benchmark, which puts it 0.7 percent below Google’s record, or about way closer towards the human average of 5.1 percent. If the accuracy of computer vision continues on the rapid upward curve it has followed in the five years since the test was created, 2015 could see artificial intelligence could score another major victory over mankind.


A message from John Furrier, co-founder of SiliconANGLE:

Your vote of support is important to us and it helps us keep the content FREE.

One click below supports our mission to provide free, deep, and relevant content.  

Join our community on YouTube

Join the community that includes more than 15,000 #CubeAlumni experts, including Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and many more luminaries and experts.

“TheCUBE is an important partner to the industry. You guys really are a part of our events and we really appreciate you coming and I know people appreciate the content you create as well” – Andy Jassy

THANK YOU