Sex, Bombs and Burgers (36 page)

Scientists at Columbia University in New York, for example, are working on using magnetic brain stimulation to lessen a person’s need for sleep. Defence contractor Honeywell is using electroencephalographs to detect neural spikes in satellite analysts’ brains before they consciously register what they are seeing, which is resulting in faster action. Boeing is using near-infrared technology to monitor pilots’ brains with the hopes of eventually allowing them to fly several planes at once. At the University of Alabama, scientists have successfully kept lab mice alive with 60 percent of their blood gone through
injections of estrogen, a process they believe can be replicated on humans.

And that’s just the stuff we know about. While many of the publicly known programs stop short of full-out genetic engineering, there’s little reason to believe that some cloning research
isn’t
being done with military applications in mind. Congress has raised some concerns over this sort of biological research, but the net effect so far has been the delay of funding for some projects or the simple renaming of others. “Metabolic Dominance,” for example, was changed to the less ominous-sounding “Peak Soldier Performance.”
¹⁷

Found in Translation

Like any good business, the military is looking to streamline operations, cut costs and introduce efficiencies. That’s the impetus for the U.S. military’s “network-centric operations,” or a fully networked battle force. The plan is to get all those robots in the field to communicate better with their human masters, and also with each other. It also means coming up with better ways to deal with the ever-increasing amounts of data pouring in from the battlefield. Part of the solution is better communication systems, such as the delay-tolerant network Vint Cerf is working on. Another idea is “cognitive computing,” an effort to reduce the amount of data humans have to sift through by letting computers do it for them. The computer then only passes on the most pertinent details to its human master, perhaps with a suggested course of action. This rudimentary artificial intelligence is intended to reduce the military’s “tooth-to-tail” ratio—the number of support staff it has to field compared to its actual fighting forces.

In 2009 Robert Leheny, DARPA’s acting director, made the case for smarter computers in a speech to a House of Representatives committee on terrorism. “Without learning through experience or instruction, our systems will remain manpower-intensive and prone to repeat mistakes and their performance will not improve,” he said. The Department of Defense “needs computer systems that can behave like experienced executive assistants, while retaining their ability to process data like today’s computational machines.”

DARPA’s Personalized Assistant That Learns program, or PAL, is doing just that in military hospitals. The computer system is capable of crunching large amounts of data and then taking action by itself. Receptionists, not programmers, are teaching the system to find vacant appointment slots and make referrals.

If that sounds like the beginning of a
Terminator
-style apocalypse, the work being done on translation is where things really get scary. Franz-Josef Och looks like your average mild-mannered computer programmer, although his German accent might frighten some into believing he is the quintessential mad scientist. He grew up in a small town near Nuremburg and discovered a passion for computer science early on. Around 1997, while attending the University of Erlangen-Nuremberg, Och became interested in something called statistical machine translation, a method of understanding languages using algorithms rather than grammatical rules.

The idea of using computers to translate languages has been around since the beginning of the Cold War, when the United States was focused on understanding Russian and vice versa, but very little quality progress was made over the intervening fifty
Sex, Bombs and Burgers years. The problem, Och explains, was twofold. The grammatical approach, in which a computer is programmed with the rules of two languages, say English and Russian, didn’t work very well, because there are too many little differences, slang uses and idiosyncrasies to provide an accurate translation. The statistical approach, in which a computer algorithm analyzes patterns in the languages and then compares them, was potentially more promising, but it also had big issues. First, Cold War–era computers didn’t have the processing power to analyze reams of digitized data. Second, those reams of data, which are all-important in producing the statistical sample for algorithms to analyze, simply didn’t exist. By the late nineties, however, both problems were no longer issues as computer processors were packing impressive horsepower and the internet had made digital data plentiful.

In 2002 Och went to work at the University of Southern California’s Information Sciences Institute. The same year DARPA sponsored a contest, not unlike its robot car races, to develop new algorithms for statistical machine translation, particularly ones that could translate Arabic into English. Och entered the contest in 2003 and built a translation engine using publicly available documents from the United Nations, which are automatically translated by actual humans into the six official languages: Arabic, Chinese, English, French, Russian and Spanish. With this gold mine of millions of comparable digitized documents, Och’s algorithm scored an impressive accuracy rate and won the DARPA prize. The following year he was snagged by search-engine giant Google which, like the military, had a significant interest in computerized translation. The two actually have reverse interests: while the military wants
to translate languages, particularly Arabic and Chinese, into English so that it can understand its real and potential enemies, Google wants to translate English into other languages to open up the largely Anglo web to the rest of the world, which will dramatically increase the size of its advertising market.

Google launched Translate in 2001 and refined it with Och’s methods when he came on board. As of 2009 this online tool, where the user simply pastes in the foreign text and presses a button to have it translated into the language of their choice, handled more than forty languages. And it works—unlike those other gibberish-spouting attempts that came before it, Google Translate generally gives you the gist, and a little bit more, of the text. As a result, the search company tends to score at the top of annual machine-translation tests run by the National Institute of Standards in Technology. Still, the tool isn’t perfect and the trick now is to get its success rate up to 100 percent. To that end, the company in 2009 announced “translation with a human touch,” in which actual human translators can suggest improvements to the algorithm’s results. The success rate is also bound to improve, Och says, now that the tools to create better systems are freely available. “It’s so easy now because some PhD student somewhere can download all the data and some open-source tools that various people have written and build an end-to-end system. It was virtually impossible before.”
¹⁸

Google is also applying the statistical approach to voice translation. In 2007 the company launched a 411 phone service that people could call to find businesses they were looking for. The caller spoke their query into the phone and Google computers would either speak the answer back or send a text message with the information. The point behind the service
was to amass a database of voice samples, similar to Och’s U.N. documents, from which Google’s algorithm could work. The experiment has borne fruit, with the company launching a voice search service for mobile phones that lets users speak their query into the device, rather than typing it.

DARPA is now testing similar technology in Iraq. Soldiers there are being equipped with iPod-sized universal translator machines that can interpret and speak Arabic for them. The devices have the basics down, like discussions about infrastructure or insurgents, and are slowly improving in other areas as well. “We knew that we couldn’t build something that would work 99 percent of the time, or even 90 percent of the time,” says Mari Maeda, who runs the translation program for DARPA. “But if we really focused on certain military use cases, then it might be useful just working 80 percent of the time, especially if they don’t have an interpreter and they’re really desperate for any kind of communication.”
¹⁹

Some linguists believe that computers, which have already become better chess players than humans, will eventually surpass our ability to translate languages as well. “Human translators aren’t actually that great. When humans try to figure out how to translate one thing, they drop their attention as to what’s coming in the next ‘graph,” says Alex Waibel of Carnegie Mellon, who was also born in Germany and does translation work for DARPA. “And they’re human. They get tired, they get bored.”
²⁰

Welcome Our Robot Overlords

The potential for machine translation far outstrips simple military and commercial uses. The ability to understand all languages may take the internet’s equalizing and empowering
abilities to a higher level, providing a greater chance at world peace than we’ve ever known. Since the advent of mass media, the public has had its opinions of other people in other parts of the world shaped largely by third parties: newspapers, books, radio, television, movies. While the internet opened up direct links between the peoples of the world and theoretically cut out those middlemen, the language barrier still prevents real communication. With truly accurate and instantaneous text and voice translation only a matter of years away, that final obstacle is about to fall. In a few years time Americans, for example, will no longer have to take the media’s portrayals of Middle Eastern Muslims at face value. They’ll be able to read, watch and understand Arabic news as easily as they view the
New York Times
’ website or CNN. And people from around the world will be able to communicate and interact with each other directly, one on one. Pretty soon, we’ll be getting friend requests on Facebook and Twitter from people in China, Tanzania and Brazil. Our social circles are about to broaden massively and we’re going to learn a lot more about people who have been alien to us thus far. As Och puts it, “There’s a real possibility to affect people’s lives and allow them to get information they otherwise couldn’t get. Machine translation can be a real game-changer there. That seems to me to be a good thing.”

Indeed, with this greater communication will come a greater understanding of other people, which will make it more difficult to go to war against them. If governments—the democratically elected kind, anyway—find it difficult today to muster public support to attack another country, it will only be harder when there are direct communications between the people of those two countries.

Where things really get interesting is in the application of statistical machine translation to more than just languages. Because the algorithm is designed to identify patterns, its potential uses in artificial intelligence are mind blowing. Google has identified as much and is taking baby steps toward the idea. In 2009 the company announced plans for a computer vision program that will allow machines to identify visual patterns. The project, still in its research phase, uses the same sort of statistical analysis as Translate. Google fed a computer more than forty million GPS-tagged images from its online picture services Picasa and Panoramio and came up with a system that could identify more than fifty thousand landmarks with 80 percent accuracy. In announcing the project the company said, “Science-fiction books and movies have long imagined that computers will someday be able to see and interpret the world. At Google, we think computer vision has tremendous potential benefits for consumers, which is why we’re dedicated to research in this area.”
²¹

Science fiction is in fact proposing the next direction that statistical machine translation could take.
Caprica
, the prequel to the hit series
Battlestar Galactica
—the best show ever, I might add—explores the idea of using pattern-identifying algorithms to create an artificially intelligent (AI) personality. In
Caprica
’s two-hour pilot, a teenager named Zoe uses such an algorithm to create a virtual AI of herself by feeding it with all the personal digital data she has produced in her lifetime. After Zoe’s death her father, roboticist Daniel Graystone, discovers the AI in a virtual world created by his daughter. The AI, a perfect replica of Zoe, explains to him how it was done:

AI Zoe: You can’t download a personality, there’s no way to translate the data. But the information being held in our heads is available in other databases. People leave more than footprints as they travel through life. Medical scans, DNA profiles, psych evaluations, school records, emails, recording, video, audio, CAT scans, genetic typings, synaptic records, security cameras, test results, shopping records, talent shows, ball games, traffic tickets, restaurant bills, phone records, music lists, movie tickets, TV shows, even prescriptions for birth control.

Daniel: A person is much more than usable data. You might be a good imitation, a very good imitation, but you’re still an imitation, a copy.

AI Zoe: I don’t feel like a copy.

As we learned with sex robots in the previous chapter, the lines between real, thinking and feeling human beings and well-programmed machines are likely to blur in the future. If, as in Zoe’s case, a computer can be programmed to statistically infer how individuals might act based on everything they’ve done before, we may very well be forced to treat them as real people. Och, who has watched
Battlestar Galactica
, isn’t sure that his algorithms will eventually result in killer Cylon robots, which is what Zoe’s AI eventually becomes, but he does think they will enable smarter machines. “Many people see different things in the term ‘artificial intelligence,’” he says, “but it will definitely lead to more intelligent software.”

The Future Is Invisible

While many of the technologies outlined in this chapter are already having an impact on the world outside the military, there
are also some way-out-there lines of research that could lead us in directions we’ve never even dreamed of. In that respect, Sir John Pendry, a theoretical physicist at London’s Imperial College, may one day be viewed as the “godfather of invisibility.”