CMU at forefront in building thinking machines

Just as Carnegie Mellon University produced the prototype computer that eventually checkmated World Champion Garry Kasparov in 1997, it now leads efforts to create robots that will defeat the world's best 11-man squad on grass by 2050.

There's serious science behind this. If robots are going to fight forest fires or build skyscrapers, they first have to learn how to work as a team -- on their own, under pressure, when every second counts.

CMU's small robot squad, four-time world champions, astounds Eric Horvitz, president of the Association for the Advancement of Artificial Intelligence.

"Watch them shoot," said Horvitz, a Microsoft senior researcher. "You smell the future."

Also impressed is retiring Microsoft founder Bill Gates, who has donated $20 million to CMU to build a computer science complex. In fact, he wrapped up his "farewell tour" in February at the Oakland campus with a speech touting the future of artificial intelligence.

Experts understand that robot soccer might seem bizarre, but it is one of many paths being blazed at CMU to replicate the human brain.

A computing titan

CMU is easily on par with Stanford University and Massachusetts Institute of Technology, critics say.

Some argue it is better.

"How (CMU) compares to those other two, I think it has a reputation for getting AI to work in practice," said Toby Walsh, editor-in-chief Journal of Artificial Intelligence Research.

That the university's robot SUV won a federally sponsored suburban road rally in California last fall was no surprise, Walsh said.

"Just sitting here in my office at Microsoft, looking around, I'm surrounded by young CMU researchers we've hired because they're the best and the brightest in AI," said Horvitz.

With $6.9 million in National Science Foundation grants for AI-related projects over the last three years, CMU gets nearly as much money for its experimental work in the field as MIT ($4.1 million) and Stanford ($2.9 million) combined.

Seeing with silicon

Computer science Professor Alexei Efros, 33, is part of the university's next generation. The Russian immigrant is working on one of AI's most daunting puzzles: vision.

Cruise missiles and drones pilot themselves using terrain landmarks. Robots quickly charge a soccer ball when they spot it. Computers can pick out shapes and edges, but they don't understand the world their cameras show them.

Until a robot understands what it sees, it won't be able to follow a direction such as "turn left at the church," or know that the child it was supposed to watch is behind the jungle gym and did not vanish.

"Yes, (computers) can beat the best chess player, but they can not do all the things that even the stupidest person can do, like finding your way home, or walking down steps or buying spaghetti at the supermarket," said Efros.

This summer, Efros is expected to introduce a program that guesses where a picture was taken by comparing the photo to millions of others posted on an Internet sharing site, where users tag their images with location coordinates.

A computer with the visual sense of a common animal, such as a rat, would be an immense accomplishment, Efros said. A little progress can yield big dividends.

Efros, who is legally blind, has unique insight.

"I think I'm more optimistic, because given my vision, I should not be doing as well as I'm doing. Yet most people don't even know that I don't see well, because it seems a human can adapt even given poor input."

Computer science professor and artist Yang Cai, 45, is trying to teach computers how to be creative.

"Robot art and computer art is unbearable, robot music is unbearable, because they're created by engineers, by scientists, not by real (artists and) musicians," said Cai, whose etchings on stone will be exhibited in a Manhattan gallery this year.

Cai is starting with some basic building blocks. Working with an English professor, Cai compiled a list of descriptions of classic nose shapes -- Roman, button, etc. -- and showed computers their silhouettes.

He's also trying to teach computers how to recognize people's nether regions so they can draw digital underwear. The reason is to give some privacy to people being scanned by the latest airport security devices, but Cai said a side benefit is teaching machines to appreciate the human form.

Building a brain

Researchers are still trying to simulate a brain.

Tom Mitchell, head of the university's machine learning department, and Marcel Just, director of its cognitive brain imaging center, won a $1.1 million private grant this year to map which neurons in our brains fire when we think about individual objects.

They want to map the neural pattern for every English common noun, such as "tomato," plus abstract nouns, such as "justice." Their prototype has done about 60 words.

"Imagine that, some day, somebody figures out the wiring diagram for all the neurons in the brain, and the rules for neurons firing. You can then imagine people building a simulation in a computer of all the individual neurons firing. I'm guessing if you did that you could have something that behaves like the brain behaves," Mitchell said.

Jaime Carbonell and Anatole Gershman met at Yale in 1975, as doctoral students in natural language processing. For each, English is a second language. Carbonell is from Uruguay; Gershman, from Russia.

Carbonell stayed in academia and Gershman went into corporate research. Reunited at CMU, they are collaborating on a super-Google that doesn't just look for matching words on Web pages, but reads and understands the pages and writes a summary.

They have dubbed it "Extracting Knowledge from the Web for the Purpose of Being Able to Answer Non-Factoid Questions."

"We haven't come up with a snazzy project name, but we're open to suggestions," Carbonell said.

Even projects that don't pan out can bring great rewards. One of Carbonell's students created Lycos, a pre-Google search engine that had a brief, shining moment during the Internet stock bubble.

With the money it earned from selling Lycos shares, CMU built a computer science building, named Newell-Simon Hall in honor of AI pioneers Allen Newell and Herbert Simon.

What comes next

Understanding human speech was one of the first goals of artificial intelligence.

"Almost any speech recognizer people are dealing with -- whether they're calling directory assistance or, for better or for worse, airline reservations lines -- really, it is fair to say, has its modern roots in core research at Carnegie Mellon," said Microsoft's Horvitz.

Professor Rob Rutenbar has helped design a custom microchip that can understand about 95 percent of the words spoken by someone reading the Wall Street Journal -- played back at 10 times normal speed. The goal is 1,000 times faster, which could make blanket surveillance of telephone calls feasible.

"But technology is good and it's bad," he said. For example, fast speech recognizers could help archive decades of newscasts to make them as easily searchable as Google.

The steady pace of advancement makes Mitchell optimistic about reaching the ultimate goal.

"My guess is that there's nothing in principle that makes it impossible for computers to be intelligent. It's just our own stupidity and our inability thus far to figure out how to do it," he said.

It's still a long road ahead.

Shoot and score

Professor Manuela Veloso is preparing her team of graduate students and the paint can-sized robots they program to defend CMU's world soccer title, starting with the U.S. championship in Pittsburgh in May.

Her work is attracting notice. A paper she and graduate student Colin McMillen wrote on robot teamwork and strategy was voted the best submitted to last year's conference of the Association for the Advancement of Artificial Intelligence.

Though she claims to hate soccer, the native of Portugal native spends hours watching tape, analyzing opponents, trying to find ways to tweak the code to keep the winning streak going.

"I tell you, I myself, one of these days I will have a heart attack if I'm not very careful," she said.

Veloso obsesses over a video of the first soccer match between humans and large rolling robots last year. When a professor on the human team tries a shot, the robot goalie charges and blocks him. So moments later, the professor just shoots from farther out. Easy score.

"Humans are fantastic. It almost seems like a lost cause to be this smart, even in something so simple," she says with a groan.

In the video, Veloso coaches the humans from the sidelines. When her player scores, she does something no artificially intelligent coach ever would.

She benches him.

• Most Read
• Most E-mailed
• Latest News