If all goes well, 2009 will be the year that
bioengineering Associate Professor Kwabena
Boahen unveils a unique computer that he
has been dreaming of for decades. Modeled
on the brain but built in silicon, it will have
the equivalent of a million brain cells, or
neurons, giving it the intellectual power of,
well, a honeybee. Not impressed? Step
back and reconsider that it will be a silicon
brain, implementing nature as an information
technology to an unprecedented degree.
“It’s not going to be very smart,” Boahen
acknowledges. “The more important point
is that it’s large enough for us to be able to
model areas of cortex and have them talk
to each other in enough detail that we can
identify what these neurons are doing and
how that affects how they interact.”
That’s the beginning stage of replicating
cognition on a chip, an achievement that
could not only yield new insight into
the fog that sits between neuroscience
and psychology, but also produce a very
efficient and inexpensive supercomputer.
Boahen estimates that one of the world’s
faster supercomputers, IBM’s Blue Gene,
processes information at roughly the same
speed as the human brain, but uses 100,000
times as much power. Nature has a much
more efficient architecture.
And so, he strives to make computers based on neurobiology by arranging conventional transistors—the binary electrical switches on which computer chips are based—into groupings that emulate the workings of nuerons.
In August 2008 Boahen’s research group
celebrated a milestone. They commissioned
the manufacturing of a new chip, called
NeuroCore, with 65,536 silicon neurons.
Each NeuroCore is made to be directly
connected to up to three others, forming
a 16-chip system called a NeuroGrid. That
million-neuron computer will be, by far, the
most sophisticated neurologically based,
or “neuromorphic,” computer ever made.
Once assembled, the NeuroGrid will allow
Boahen to run experiments that will advance
the goal of using a neural architecture for
computation.
Building up the nerve
Boahen first became interested in how the
nervous system processes information while
a junior studying electrical engineering at
Johns Hopkins University. He had come
to Baltimore from his hometown of Accra,
Ghana, where as a teenager he was both
fascinated by and disappointed in his first
computer, a Radio Shack TRS-80. The
way the computer handled information,
with a rigidly timed series of 0s and 1s, seemed inelegant to him. When Boahen attended a lecture by Johns Hopkins neural network
researcher Terrence J. Sejnowski, who in 1986 had co-developed a system called NETtalk that
could read text aloud, Boahen became inspired to think that neuroscience provided a better
computational model.
“I thought that was very elegant,” Boahen said. “That got me very interested in the biological
approach.”
Before coming to Stanford in 2005, Boahen was a professor at the University of Pennsylvania.
That’s where he began building neural circuitry in silicon. His first real success wasn’t with the
brain, but with another better-studied member of the nervous system: the retina. Using what
amounted to a “blueprint” developed by neuroscientists, he figured out how to build it with
transistors. The silicon retina successfully captures images much like a real part of the retina
does, by tracking the moving edges of objects (the eye looks at things many different ways,
leaving the brain to reconcile the information to produce the images we perceive). In recent
years Boahen has also made a chip that acts as an artificial inner ear.
From binary to brain-like
In some ways, transistors and neurons couldn’t be more different. Transistors are precise,
punctual, and reliable, but neurons are sloppy and slow. Microprocessors are methodical,
serial, and fixed, while the brain is spontaneous, parallel, and constantly remaking itself. It
turns out that to design a circuit that can properly mimic the behavior of even a simplified
neuron, Boahen needs about 340 transistors. That means each NeuroCore chip employs more
than 22 million transistors to create 65,536 neurons.
Most of the transistors compose an electrical analogue of the ion channels that allow real
neurons to send and receive signals electrochemically.
“The motivation for this work is to build a model that has a lot of biological fidelity,” Boahen
says. “Each neuron has a repertoire of ion channels and that determines their specific
behavior.”
Boahen’s design allows for his neurons to have four such “channel populations,” or to combine
those into two more complex ones. This flexibility allows the neurons to be programmable,
taking on different types and behaviors and establishing or breaking connections, much as real
neurons in the brain do.
NueroCore dedicates about a million transistors to creating a network infrastructure that allows Boahen to study how individual
nuerons are working and to reporgram them. That infrastructure is what makes the chip a useful research platform.
This is not the first neuromorphic chip, as
both Boahen and others in his small field
have made some before. But Boahen’s is
the biggest and best architected one so
far; the previous best system included
four chips for a total of 45,000 neuron
equivalents. And while the prior systems
were all simply strung together like beads
on a necklace, the NeuroGrid will arrange its
16 chips in a much more tightly integrated
“tree” structure, in which each chip is
connected to one above and two below.
The tree structure and the integrated
network infrastructure on the NeuroCore
chips combine to allow the NeuroGrid to
achieve its 20-fold improvement in scale.
Ultimately many more innovations and
probably decades will likely be necessary
before Boahen or anyone else could create
a machine that rivals the human brain. But
for now, NeuroGrid, with its honeybee brain,
seems certain to strike many researchers as
the “bee’s knees.”
|