Teaching computers to read first step in developing consciousness

And we don't need to know.

Bringsjord, the director of the Artificial Intelligence and Reasoning Lab at Rensselaer Polytechnic Institute, knows the deal in a totally non-conceptual way. He let something slip once. And he got The Call.

We have detected mention on your vitae of a dollar sign associated with us. You must remove that immediately.

Bringsjord explained to the voice that he wasn't presently at a computer. The voice insisted. Bringsjord, a tall, slender man who is far more down-to-earth than his erudite resume might lead one to believe, complied. (Argument: Reasoning for or against not recommended when dealing with government agencies.)

Being someone who trafficks in figuring things out and then sharing them with the world, keeping information on the down-low isn't a natural state for Bringsjord. His resume, for instance, is 23 pages long, filled with accomplishments and knowledge bestowed. He's written six books, countless book reviews and research papers with titles like "The Zombie Attack on the Computational Conception of Mind." He's written opinion pieces for The Troy Record and The Los Angeles Times. He is sought out by reporters nationwide every time Hollywood releases a new robot flick or when anyone anywhere makes a new advance or claim in the field of AI.

Oftentimes, of course, that advance is being made by Bringsjord and his team at RPI. As was the case when the lab developed a robot -- psychometric experimental robotic intelligence, or PERI — that can pass portions of a standard IQ test. And then there was Brutus.1, a program that could write original short stories (albeit with limited, sinister themes), which created global buzz. Bringsjord, in fact, is now trying to define evil so Brutus.1, whose characterizations were "terrible," can fashion a more believable bad guy.

"This means defining evil, defining lying, which isn't as straightforward as it would seem," Bringsjord said.

Last December, Bringsjord and co-investigator Konstantine Arkoudus were awarded a $400,000 grant from the Defense Advanced Research Projects Agency, the central research and development organization for the Department of Defense, to investigate issues about learning and reasoning for a project called Poised for Learning. The grant may extend for three years and total $1.2 million.

In short, Bringsjord and his team will try to produce a computer that can read, comprehend and reflect on what it has learned. It's a crucial step in the advance toward expanding artificial intelligence.

"RPI's work will investigate learning and reasoning, both areas that are key to achieving the vision of cognitive systems," said Jan Walker, with DARPA's external relations department. "In addition, while learning and reasoning are generally important, it is also important to be able to measure when a cognitive system has learned. The RPI projectwill develop ways to help measure when a system has truly learned something."

Bringsjord is 46, a skier and a father of two teens who are now gleefully surpassing their father, athletically.

It wasn't math or computers, but a fascination with logic and natural theology that drew Bringsjord into the field he now dominates. As a child, seeking a good argument, he canvassed his mom and other churchgoers as to why they believed in God. As he grew, people urged him to become a trial lawyer. Admittedly not mechanically or mathematically inclined, he earned his Ph.D. in philosophy from Brown University before his love of logic — a la Sherlock Holmes, Mr. Spock — carved a path straight toward the growing field of AI in 1980. He came with some ideas of his own.

"I'm a massive skeptic about robots ever reaching human intelligence, which makes my own life kind of paradoxical," said Bringsjord, who later moved to RPI to head up its philosophy and cognitive science departments and then the AI and reasoning lab.

The idea of consciousness is a heated subject in AI circles: It's known as the "C" word.

"People take such great lengths to avoid that word in AI," Bringsjord said. "If you use the 'C' word you should be able to tell us in terms of working computer programs or algorithms what it is, and you can't, so shut up."

Discussing consciousness in terms of computers is, at present, a waste of time, Bringsjord said. "The truth is that how could a machine ever experience skiing at 40 miles per hour, when you could die at any moment? It can't."

Yet Bringsjord soldiers on, striving to prove himself wrong — an achievement he doubts will come in his lifetime, but one whose fundamental building blocks appear to be firmly within his grasp.

An avid fan of sci-fi films (favorite movies are "Blade Runner" and, of course, "2001: A Space Odyssey"), Bringsjord noted with amusement how, on the big screen, humans seem obsessed with dark and highly cinematic visions of computer domination (see "The Matrix," "Terminator") and of omnipresent super computers gunning for humanity. Sci-fi in real life is far more mundane, tedious and unsexy, say those who know.

"Things that are easy for humans are really difficult for robots," said Bettina Shimanski, a doctoral student in computer science who's working under Bringsjord's supervision (and who, by the way, does believe robots someday will achieve human abilities). "Things that are hard for humans, like math, are easy for robots."

Reading is one of the best and most efficient ways to learn, yet, "the brute fact is that machines, though often touted as learning this and that, can't read," Bringsjord said. "If you can really teach a computer logic and reasoning then you can really teach it to do anything."

Looking supremely professorial in cords, turtleneck and blazer, Bringsjord sat in his office on RPI's campus, the office with the panoramic view overlooking Troy, at dusk. He pulled an elementary school-level astronomy book from a bookshelf.

He flipped to a page of text, topped by a picture of a night sky laced with contrails.

"This shows humans who read the book how planets can look like they're traveling backwards when they really aren't," Bringsjord said. "When humans read, it's a self-reflective process where the reader is thinking about how they may be called on to use the information they're reading about."

A human reading the text would consider possible questions and devise answers, even if the specific question wasn't addressed in the text. Teaching a computer to do the same means first breaking language down into something computers can understand: math, algorithms.

"Selmer wants to be able to ask a computer something and get both an answer and a justification. That sort of thing hasn't been done successfully before," said Sunny Khemlani, a junior cognitive science and psychology major who takes proofs the system understands and turns them into English, acting as a sort of translator between human and computer.

Now, getting a computer program to provide a humanlike response is a wholly inorganic process: It means programming in every single detail that would determine the outcome. Ultimately, if the project is successful, this laborious step wouldn't be necessary.

"We want the computer to be able to learn that on its own, by reading," Khemlani said. "Essentially the computer would be doing its own research."

And once that is possible, the applications would be limitless. One could, for instance, build artificial agents that can reason about the ethics of their actions. One could stage realistic war games, where computers reading manuals or plans can instantly adjust military maneuvers. One could ...

"If you write that, they'll find it," Bringsjord broke in, all too aware of the omnipresent watchful eye of something that's not yet a supercomputer, but that, strangely, doesn't seem entirely human either.

Stephanie Earls may be reached at 454-5761 or at searls@timesunion.com.