Heads Up!

With apologies to the film Armageddon, I've set up a countdown timer. It's counting down to the end of 1999. I've committed to finishing my Ph.D. before this clock runs down. If I'm late, the entire WWW will know. Click on the timer to get an update of how my work is progressing.

Research Vision

Building Human-Level Intelligence From Scratch May Be Impossible, but...

Natural intelligences (people) take years to learn the kind of skills that make others regard them as "smart." The task of the artificial intelligence designer is even harder than that of the parent, because computers lack the ability to learn from a wide variety of input and using various methods. It's a lot harder to build a system than to teach a system. In fact, I think it may be impossible to build a system with human-equivalent intelligence "out of the box".  However I think there is a better chance of building a system with significantly sub-human abilities that can learn to become as intelligent as a human being, and perhaps even beyond. Therefore, it is far more important to find out how a computer might learn to do a task than to program it to do so.

The most powerful form of learning is learning the skills of learning itself.

"Catching Fire"

This phrase, "catching fire" is a metaphor for what happens when an artificial intelligence can learn to learn. This sets up a powerful feedback loop. The better an artificial intelligence can learn, the more it can learn, and some of the knowledge it acquires could again improve its ability to learn, and the cycle repeats. The artificial intelligence grows and expands, consuming and integrating more and more of the knowledge around it like a flame consuming fuel, its rate of growth increasing the more it grows.

Bio

Right now I'm a Ph.D. student at The AI Lab at Georgia Tech in the College of Computing. I received my B.A. in Computer Science / Mathematics and Psychology from Emory University in 1994, graduating with Highest Honors. My honors thesis investigated human perception of the intersection of planes in 3-D space.

Current Research Project

Planning to Learn

In any realistic learning situation, there are a near-infinite number of concepts that could be learned. To learn effectively, an agent must pick and choose which concepts to learn first. Sometimes learning one concept depends on learning others, e.g., learning to multiply large numbers requires an understanding of multiplication tables. For this reason, learning can be viewed as a planning task, and traditional planning techniques can help an agent create a sequence of actions aimed at learning a particular concept.

Publications

AlgoNet2 papers from the ONR Workshop and ICLS

These Are Some People in my Neighborhood

Ashwin Ram

My research advisor and all-around guru. Learned artificial intelligence directly from Roger Schank. Wrote the AQUA system, which reasons about typical news stories such as terrorist bombings in the middle east Also responsible for some small bits of the ubiquitous EMACS text-editor.

Anthony Francis

Renaissance man. This guy knows something about everything, from computer science  to animal locomotion to web site design. He does Artificial Intelligence for a living and writes science fiction just for fun. His current project is NICOLE, which features a very human-like memory system.

Kenneth Moorman

The reading machine. Ken is a big SF fan too (Heinlein in particular). Has been known to read a lengthy novel in a single day. No big surprise, then, that his research interests are in natural language. He has just finished his PhD dissertation on ISAAC, his science-fiction-story-reading AI system.

Marin Simina

The Denny's of research. This guy never stops working (or at least we've never seen him stop). He studies creativity, focusing on the discoveries of Alexander Graham Bell.