DARPA aims for machine cognition

The Defense Advanced Research Projects Agency wants to develop computers that can think and learn from past experience.

Defense Advanced Research Projects Agency Web site

ANAHEIM, Calif. -- The Defense Advanced Research Projects Agency wants to develop computers that can think and learn from past experience.

"Today's computer systems are creatures of ignorance," said Ronald Brachman, director of the information processing technology office, speaking yesterday at the DARPAtech 2004 conference.

Science fiction has had many visions for intelligent machines, from the computer HAL in "2001: A Space Odyssey," to Stephen Spielberg's "A.I." DARPA isn't stretching that far, but the agency hopes to develop the first cognitive personal assistant, a machine that can learn from past experience and reasonably anticipate what its user's needs will be based on that information.

"They need to be able to learn about their environment over time so that once they survive an attack or a breakdown, they can immunize themselves against future occurrences," Brachman said. "They need to be able to adapt to their users so that humans can do their jobs more effectively, and they need to be sensitive to those users' missions and priorities."

Brachman said his office is currently determining the best way to get from the dumb machines of today to where they want to go. The goal is not to create a machine that is human-like, but one that can advise and alert based on its past experience, in the same way humans do.

For instance, if a smart radio knows from past experience that a certain section of a platoon's patrol area doesn't have good reception, it can alert the user to complete any transmissions before they enter the area. Or, a machine that notices an enemy is attacking in a certain pattern or formation can determine if it has seen that formation used before and extrapolate a reasonable guess as to what the enemy's tactics might be.

To achieve cognitive systems, the information processing technology office has determined that five elements are needed to create a truly intelligent machine:

* Real world learning.

* Real world reasoning.

* Natural language.

* Episodic memory.

* Self-aware reflection.

* Perception and cognition.

The information processing technology office has issued a broad agency announcement for ideas as to how to create an architecture for cognitive information processing.

"The new cognitive agenda of the information processing technology office will bridge the gap between the cognitive side and the computing side of our office," Brachman said. "We have the beginnings of a new cognitive system in place, but we're just getting started."

NEXT STORY: States slow DHS grant disbursement