Andrew Ng and Baidu Push Computer Vision to New Heights With Deep Image Supercomputer

Nobody talks about “supercomputers” anymore. The term almost sounds quaint, bringing to mind those giant 1960’s computers built from dozens of reel to reel tape recorders, vacuum hoses and switches and levers, and of course they run for five noisy hours before cranking out a punchcard that you’ll need a technician to decipher.

However, supercomputers are still relevant in technology fields. Computers like the Tianhe-2 at Sun Yat-sen University in China can process data at millions of times the speed of an iPhone or a home PC. NASA uses a supercomputer known as the Pleiades, with nearly two hundred thousand cores and a total memory of over six hundred terabytes. This is more processing power than the average consumer will ever need, but if you’re running a large halidron collider or sending a probe to Mars, all of those extra “petraflops” will come in handy.

Supercomputers in the 21st Century are primarily used for some very high-level hard-science stuff: quantum physics, climate research, molecular modeling, and incredibly precise physics simulations used to explore everything from nuclear fusion to recreations of the Big Bang. When the findings are finally published in journals like Popular Science, you’ll notice that it’s always broken down into simile and analogy, because to make sense of the data itself would take years of technical training.

Baidu’s new supercomputer, Minwa, uses all of this processing power for something a little easier for the layman and the armchair scientist to grasp: image recognition. The computer was built to house Deep Image, Baidu’s computer-vision software.

When we say “image recognition” you may be thinking of something like reverse-image searching with Google. That’s the tip of the iceberg. The ultimate goal of computer-vision is to allow a computer to see the real world better than even human beings are able to see it.

Prior to Deep Image, Google led the way with a 6.66% rate of error. Deep Image clocks in at 5.98%. That’s a small difference if we’re talking about how big a slice of birthday cake you want, but in computer-vision, it’s a tremendous leap.

The software itself uses deep learning, which effectively means that it can, well, think , and learn from its mistakes.

Robotics right now demand a routine that has already been laid out for them, such as on an assembly line, or for a human being to be at the controls, as in a bomb-disposal robot. When machines can see as well as a human, or better, and recognize what they’re looking at, they can also respond and react more quickly than a human being can. Right now, a bomb disposal bot is only as smart and as fast as its human operator. 20/20 computer vision inches us closer to bomb disposal bots that can react to its actual, physical situation as quickly as its software can process it.

In short, Minwa blurs the line between digital and physical: when a computer can process its surroundings even better than a human being can do the same, the possibilities are multiplied thousands of times over.

Andrew Ng and Baidu Push Computer Vision to New Heights With Deep Image Supercomputer