Building the Next Generation of Intelligent Computers to Resemble the Human Brain
Maybe you have heard of robots that can fly. But how about a robot that is able to learn through its own experiences to drive itself to the airport?
Stefano Fusi, an associate professor of neuroscience, is working to build an artificial brain that might one day allow robots and other machines to do that and more.
For the last three years, Fusi has been part of a $42 million project funded by the Pentagon’s Defense Advanced Research Projects Agency (DARPA). The goal: to build a new kind of intelligent computer—one with hardware and software that draws inspiration from the architecture of the human brain.
“What we are building is not a processor that can understand a predefined set of instructions and execute them,” he says. “This is a self-organizing system that learns autonomously. It’s a dynamic system based on the brain.”
Why do scientists need a new kind of computer? In addition to its unparalleled ability to learn certain types of behavior, the human brain—small enough to fit into a 2-liter soda bottle—consumes about the same amount of energy as a 20-watt lightbulb.
Today’s computers, by comparison, are bulky energy hogs. The central processing unit is physically separated from the computer’s memory, and data is shuttled between the two components through a communications channel, or “bus.”
But the brain’s memory and processing units are in the same place. Its billions of neurons—think of them as processors—are surrounded by cell membranes that carry an electrical charge. They’re connected to one another by trillions of synapses—think of the varying strengths of those synaptic connections as memory.
When a neuron’s charge exceeds a certain threshold, it fires, releasing chemicals into its synapses. If the synaptic connection is strong enough, the neighboring neuron will also fire. This biological structure is one of the main reasons why the brain is so fast and efficient.
The project is being led by a team of engineers at IBM’s Almaden Research facility in Silicon Valley, who are tackling the hardware challenge. They have created a new kind of low-powered digital computer chip that contains 256 imitation neurons and thousands of synaptic connections. Eventually, they will create a massive brainlike computer out of the new chips.
Fusi and his collaborators at the University of Wisconsin, University of California, Merced and Cornell are developing mathematical models that represent the dynamic connections between neurons in the prefrontal cortex, the part of the brain associated with decision making and social behavior.
They will build electronic systems in the chips that mimic the equations of the mathematical models. The models dictate the rules governing the connections between small populations of neurons but will someday be scaled up on a huge network capable of simulating a massive number of neural systems.
“We know a lot about the neural systems of animals—vision, decision making—all the sensory modalities have been studied in detail,” he says. “We also know a lot about memory, learning and reinforcement learning. What nobody has done is try to integrate them.”
Key to that effort is mimicking the way the brain encodes learning. When we sense something, it is because electrical signals are passing from receptors to other neurons and causing the neurons to fire. But neurons can also cause one another to fire—that is how associative thinking works.
The more often outside stimuli cause neurons to fire together, the stronger their synaptic connections grow—and the more likely they are to activate one another in the future even if the original stimulus is absent. Consider, for instance, how we learn to associate a fragrance with a rose. The experience of smelling a rose builds connections between smell-related neurons and sight-related neurons.
If the connections are strong enough, even a picture of a rose will excite the area of the brain where smell is recorded. Even if we don’t actually smell a fragrance, the picture of the rose may cause us to remember the smell.
In his programs, Fusi has mathematically represented these learning rules. “Most of computation in the brain emerges from the interactions between neurons,” Fusi explains.
Fusi isn’t sure when his project will produce actual thinking robots. But he believes some level of functionality isn’t far away.
“There will be many applications,” he said. “You could control a flying robot. It will be able to learn autonomously and to adapt to a new situation in a flexible way—everything that computers can’t do that humans can.”
—by Adam Piore
|Columbia scientists talk about the new Jerome L. Greene Science Building, interdisciplinary hub of the Mortimer B. Zuckerman Mind Brain Behavior Institute and scheduled to open in 2016. (7:05)|