Charles Peck and James Kozloski of IBM's Biometaphorical Computing team say they have created a mathematical model that mimics the behavior of neocortal minicolumns, thin strands of tissue that aggregate impulses from neurons. Further research could one day lead to robots that can "see" like humans and/or make appropriate decisions when bombarded with sensory information.A research paper on the model is expected to come out this week.
The brain consists of roughly 28 billion cells, Peck explained. The 200 million minicolumns essentially gather sensory data and organize it for higher parts of the brain. The minicolumns also communicate with each other through interconnections. Minicolumns are roughly 1/20 of a millimeter in diameter and extend through the cortex.
The mathematical model created at IBM simulates the behavior of 500,000 minicolumns connected by 400 million connections. With it, "we were able to demonstrate self-organization" and behavior similar to that seen in the real world, Peck said.
"What we are trying to do is study the brain at the highest level of abstraction without masking the underlying function," he said.
In a test outlined in the upcoming paper, the system was able to solve a pattern recognition problem that will cause errors on ordinary computers.
Ideally, the algorithm could one day help scientists more fully understand the underlying processing that takes place when people see things. In a nutshell, an image is received, decomposed into color, shape, texture and other attributes and then reassembled, prompting the animal to change its behavior. Not all parts of the process are fully understood, Peck said.
Over the past two years, researchers have increasingly looked towardto emulate. Some companies, such as Cambrios, are trying to develop new compounds by exploiting proteins secreted by biological viruses. PalmOne founder , meanwhile, is creating a company that will sell systems that use the same thought processes as the human brain. Intel co-founder Gordon Moore recently said that computers won't likely be able to think like humans unless they are redesigned.
Brains typically think by making predictions about future events by looking at a vast array of past experiences, Hawkins said in a speech Monday at an event unrelated to IBM. Hawkins showed off a prototype application that can recognize shapes it has "seen" in the past.
IBM is presenting the paper at the International Conference on Adaptive and Natural Computing Algorithms in Coimbra, Portugal.