Going Deep: Employing AI Learning Variants to Predict Machine and Human Behavior

'That's really deep,' has always meant, figuratively, that an idea is thought provoking. Now, it may well be literal as well.

Tech firms using deep learning, a variant of artificial intelligence, are finding they can predict with 85 percent accuracy when hard drives will fail.

They can also determine how people will respond to certain types of ads based on how what sort of behavior similar ads have stimulated in the past. Both benefits save money by permitting more efficient allocation of resources and make money in the long run by improving returns on investment in related initiatives.

We have become so jaded by advances in technology and knowledge that little, if anything surprises us anymore. We yawn at the realization that smartphones can perform calculations and that provide guidance to in our lives that a roomful of mainframes might have struggled to do a generation ago. But the reality is that these predictive capabilities may have implications for health, welfare - and, yes, institutional control - that could once only be dreamed about. The real question is to what degree that knowledge is used for the benefit of those who are providing the data versus those who are analyzing it. JL

Cade Metz reports in Wired:

Deep learning better analyzes data describing how people have responded to digital ads in the past and adjust new ad campaigns accordingly. “Serving ads is basically running a recommendation engine, which deep learning does well.”
Deep learning can do many things. Tapping the power of hundreds or even thousands of computers, this new breed of artificial intelligence can help Facebook recognize people, words, and objects that appear in digital photos. It can help Google understand what you’re saying when you bark commands into an Android phone. And it can help Baidu boost the bottom line.
The Chinese web giant now uses deep learning to target ads on its online services, and according to Andrew Ng—who helped launch the deep learning operation at Google and now oversees research and development at Baidu—the company has seen a notable increase in revenue as a result. “It’s used very successfully in advertising,” he says, sitting inside the company’s U.S. R&D center in Sunnyvale, California. “We have not released revenue numbers on the specific impact, but it is significant.”
Originally developed in the academic world, deep learning attempts to more closely mimic the behavior of the human brain with computer hardware and software, operating “neural networks” that process information using models inspired by nature’s biological neurons. In essence, these neural nets use enormous amounts of digital data to train themselves on certain tasks, from recognizing images and natural language to predicting how our bodies will respond to certain chemicals. Everyone from Google, Facebook, and Baidu to Twitter and Yahoo is now using this technology in one form or another.

‘It’s used very successfully in advertising. We have not released revenue numbers on the specific impact, but it is significant.’

Led by Ng and a researcher named Kai Yu, Baidu has been particularly aggressive in its use of the technology, even before Ng joined the company six months ago. “Baidu, more than any other company, has aggressively moved deep learning into products, things at the heart of the company,” Ng says.
In addition to targeting ads on web services, deep learning powers the Baidu Eye prototype, a Google Glass-like wearable computer that seeks to automatically identify objects in your line of sight, and it even provides a way for the company to identify when computer hard drives inside its massive data centers are on the verge of failure. According to Ng, this deep learning system can predict hard drive failure with about 85 percent accuracy.
“We know, one day ahead of time, when a hard drive is about to fail,” he says, explaining that engineers can reroute computing tasks to other places if a disk is about to fail. “This means we can improve the reliability of the data center—and decrease costs.”
The salient question is just how much the technology is juicing ad revenues. Though Ng won’t say, a major boost would not be surprising, according to Adam Gibson, a software engineer who aims to bring deep learning algorithms to the wider tech world through a startup called Skymind. Deep learning, he explains, better analyzes data describing how people have responded to digital ads in the past and adjust new ad campaigns accordingly. “Deep learning [is] able to handle more signal for better detection of trends in user behavior,” he says. “Serving ads is basically running a recommendation engine, which deep learning does well.”
In April, during the company’s first quarter financial earnings call, CEO Robin Li indicated that deep learning was helping boost bottom line. Li’s candor surprised Bryan Catanzaro, who helped explore deep learning at chip maker nVidia and has now joined Baidu to work on the technology. “The closer that you get to the financial engine that powers these companies,” he says, “the most closely guarded the secrets are.”
Ng, who not only worked on deep learning at Google but is a central part of larger and rather close-knit deep learning community, says he’s not aware of other companies using deep learning to target ads. But there’s one notable possibility: Google.

According to Ng, this deep learning system can predict hard drive failure with about 85 percent accuracy.

Google spokesperson Jason Freidenfelds won’t say whether the company is using deep learning for advertising, but he points out that Google’s deep learning tools can be used across the company. Currently, the company uses deep learning not only to drive Google Now, the voice-powered search tool included with Android phones, but to identify images on its Google+ social networking services, and it has at least experimented with a system that translates information from one language to another.
Certainly, deep learning seems to be evolving inside Google and Baidu in similar ways. Each has built a central deep learning platform that can be used by engineers and projects across the company. And both are now running deep learning algorithms atop machines packed with hundreds of GPUs, or graphics processing units, a type of computer chip that was originally designed to process digital images but is also suited to other tasks. Deep learning algorithms require a large network of chips running in parallel, and a network based on GPU is potentially more efficient, because the chips are designed to perform the kind of mathematical calculations that are the bread and butter of deep learning, and you can stuff more of them into a single machine.
Odds are, Google is also using all this technology to target ads—and has been for quite a while. After all, that’s where the money is.