March 21, 2005 12:15 PM PST
Hawkins start-up takes brainy challenge
Hawkins, who created the Palm handheld, is forming a company that will specialize in systems able to recognize patterns, make predictions about familiar phenomena and in general function like human gray matter.
An early software application he has helped create can recognize drawings, a challenge that has bedeviled scientists to date. "There is no (available) software where you can show a picture to a computer and say, 'What is it?'" Hawkins told an audience at PC Forum, taking place here this week. (PC Forum is owned by News.com publisher CNET Networks.)
Hawkins has written a book about how the brain functions, called "On Intelligence," and has the nonprofit Redwood Neuroscience Institute to research human intelligence. Information about Hawkins' new company will come out in an e-mail from RNI.
"You will get one e-mail, and e-mail only, and that is when the company comes out," he said.
Hawkins declined, however, to say when that e-mail would be sent or whether his prototype application would be part of the company's offerings.
Human intelligence is largely a matter of memory, Hawkins said. The neocortex, the part of the brain that handles more complex functions, gathers data and then organizes it into patterns for future use.
"The hierarchical structure of the cortex actually maps the hierarchical structure of the world," he said, or at least our perception of the world.
To give an example, he clapped his hands. Everyone in the audience, he said, predicted that the hands would touch. "You did not expect them to turn into potatoes," he said. When they clapped, the usual sound emitted. "You didn't hear a pig squeal, which would have been a surprise."
Reptiles don't have the same sort of memory hierarchy. "You can anticipate things better than a crocodile or an alligator," he said.
RNI has come up a with a prototype software application that enables a computer to recognize 90 black-and-white drawings. Once the computer is shown a doodle that resembles a dog or helicopter, the machine will recognize the same drawing again, even if the size of the drawing is shrunk or slightly distorted. In one dramatic example, the software recognized the dog drawing a second time, though it was now facing in the opposite direction.
Pattern recognition is based, in part, on probabilistic techniques, which have gained popularity in recent years among search engine specialists. Although probability was largely scoffed at 15 years ago, it is now widely considered the most promising solution for artificial intelligence.