Deep Learning: Gaze into the Web Abyss, and It Gazes Also into You

REUTERS/Kacper Pempel/Files
REUTERS/Kacper Pempel/Files

Americans are accustomed to the dominance of Google, Microsoft, and Yahoo as search engines, but on the global stage, a Chinese service called Baidu is now second only to Google in popularity.

Baidu hasn’t announced any firm plans to move into the U.S. market yet, but, in addition to the huge market in China, they’ve expanded services to countries such as Egypt, Thailand, and most recently Brazil. They’ve also recently hired away one of Google’s top researchers, Andrew Ng, a specialist in artificial intelligence who has taught courses at Stanford University.

As Ng explained in an interview with VentureBeat, Baidu’s claim to fame is their search engine’s reliance on “deep learning” algorithms. In other words, their products learn what users want by analyzing their requests and what they do with the information. To some extent, every search engine and social media platform is trying to learn from its users, producing a certain degree of discomfort among those who feel their Internet tools are spying on them.

The deep learning movement wants to take this idea of “living” software further, creating systems that digest enormous amounts of data very quickly, without much human intervention, producing highly customized experiences for each individual user.

It is hoped that the benefits from this level of swift, automatic customization, and the fact that it’s all being done by faithful, computerized servants instead of intrusive, human programmers, will overcome consumer unease. One might also suppose that much of the early work is being carried out with customers who are, for better or worse, less nervous about having their online activities monitored than Americans and Europeans.

If companies like Baidu can realize the ambitious plans of visionary designers like Ng and bring polished, incredibly useful living software perfected overseas to American audiences, consumer resistance could be minimal. As the old saying goes, nothing succeeds like success.

The big American tech companies are also interested in developing this kind of technology, and while power players like Google are certainly swimming in resources, it doesn’t take much reading between the lines to get the idea that Ng left Google because they’ve become too hidebound and bureaucratic.

Ng talks about receiving needed resources much faster than his unit at Google provided, getting things done without sitting through tedious committee meetings. He can frolick through Chinese markets with much faster birth-death cycles for online products than is typical in the West, and he has a field day hiring away other top artificial intelligence researchers to join his team.

To get an idea of what this deep learning software would be like, the VentureBeat article on Ng references the recent movie “Her,” a near-future science fiction film in which Scarlett Johansson voiced an artificially intelligent digital assistant whose tireless efforts to relate with her user, and enhance the quality of his life, blossomed into an actual romance. (Not to spoil anything for those who have not seen the film, but hopefully deep learning researchers are pondering the ultimate resolution of that romance at great length.)

The key feature of these next generation systems over existing smart search engines and voice-activated smartphones is the causal grace of their relationship with users. Computer systems have grown steadily easier to use with each passing decade, long ago reaching the point of widespread consumer acceptance by growing friendly enough for average, non-techie folks to use reflexively on a daily basis. One indicator of how user-friendly consumer systems have become is that no one really talks about “user friendliness” any more. It was the sizzling-hot buzz phrase of the computer industry not very long ago, but it is now well-understood that if an application isn’t simple enough for most people to figure out with a casual glance and a few curious mouse clicks, it’s not going to break through to a mass audience.

Cryptic text-based systems have given way to graphical user interfaces that didn’t really work—they tended to cripple the machines they were running on. Eventually graphical interfaces were perfected, and now they’re ubiquitous, running on handheld devices with incredibly convenient touch screens. Voice command is the next step, but even the most gee-whiz voice apps today, like Apple’s famed Siri, are just a shadow of what they could be.

Greater accuracy is the key to making smart search engines and voice applications work, and that’s what Baidu hired Andrew Ng to work on. At the moment, the ability of online systems to adjust themselves automatically to suit user preferences is fairly crude. Much of the learning is based on what users request of a particular system… but people don’t always know what they really want. A deep learning system would study what they actually do with the data they request, build a network of preferences from many different data sources, consider the preferences of similar users, and learn to interpret the personal subtleties of spoken language.

The personal assistant envisioned in “Her” displays these qualities—she can anticipate what her user wants based on very vague requests, she adjusts her behavior to meet his demonstrated preferences, she understands the nuances of spoken conversation (there are some cleverly written early interactions where she asks questions of her user to figure out when he’s being sarcastic, what he really means when he uses certain figures of speech, and so forth.)

The ability to accurately pull meaningful data from the random noise of human behavior is crucial to making such features work. Living human beings don’t actually handle these tasks with a very high degree of accuracy, unless they know each other very well. The speed and power of artificially intelligent computer systems, combined with the endless patience of machine intelligence, could make them much better at becoming everyone’s close personal friend.

With these capabilities perfected, computer systems would cross the final user-friendliness bridge and begin doing more than half of the work to maintain a relationship with humans. As it stands, even the most easy-to-use interfaces require us to learn how the system thinks—we have to figure out where the menus are, perhaps learn a few shortcut keys, learn the peculiarities of each system’s tools for such routine tasks as file uploads, and learn how we can configure the interface to suit our personal tastes.

One reason we tend to think of modern systems as more user-friendly is that the typical user is far more comfortable with learning his way around an interface than his father was; an online generation is coming of age, and it’s more adventuresome and patient than the eighties and nineties executives who howled in frustration at the inscrutable but indispensable machines on their desks. When computer systems are able to shoulder most of the interaction burden, and we can use them as casually as we would request help from a trusted human assistant, the next computer revolution will be at hand.

Artificial intelligence is still an argument—there are those who believe it will never be anything but an illusion, that talk of “neural networks” is just slick marketing for really fast computers. We currently have search engines that work extremely well; both user experiences and uninvited advertising have been tailored to individual preferences. Such tasks as piloting an automobile, routinely accomplished by millions of humans without extensive training or individual genius, remain beyond the capability of the smartest computer system; the ability of an incredibly complicated machine to briefly fool a panel of human judges into thinking they are holding a conversation with a human child is celebrated as a landmark A.I. triumph. The practical application of such technology remains elusive, and yesterday’s promised miracles remain science fiction.

But then again, many electronic conveniences we currently take for granted were science fiction just twenty years ago. It’s interesting to watch where high-rolling tech companies are placing their bets, and they all seem convinced that deep learning systems are worth investing sizable sums of money in. The problem for American investors is that it doesn’t seem like our domestic giants are nimble and adventurous enough to keep up with the work being done by the big Chinese companies. At least, that’s how Andrew Ng placed his bets.