If
you doubt the power of the human brain, ponder this for a moment: It
takes today’s state-of-the-art supercomputer eight-and-a-half minutes to
simulate just five seconds of normal human brain activity. Meanwhile,
that supercomputer will consume 140,000 times as much electricity as the
brain—1.4 million watts to ten to be exact—to do the work. For
sheer processing power and efficiency, nothing quite compares to the
human brain.
In a recent paper in the online edition of Nano Letters,
a team of Stanford University engineers has demonstrated a new nanoelectronic
device that emulates human synapses, the brain’s computing mechanism. It is a breakthrough that might one day lead to portable,
energy-efficient, adaptable, and interactive computer systems that can
learn rather than merely respond to given programs.
The
Stanford team, led by Professor H.-S. Philip Wong, post-doctoral
scholar Duygu Kuzum, and graduate students Rakesh Jeyasingh and Byoungil
Lee, has been working in a new field known as “brain-inspired
computing,” which seeks to mimic in computer chips the neurological
signaling mechanism of the human synapse.
The
researchers are not the first to venture down this path, but they are
the first to succeed at creating synaptic devices small enough, with a
low enough energy consumption, and created with a mature technology so
as to anticipate commercial viability down the road.
“This
development could lead to electronic devices that are so small and so
energy efficient that we might be able to make nanoelectronic versions
of certain parts of the brain to study how they work,” says Wong, a
professor of electrical engineering. “While you can’t alter a biological
brain, a synthetic device such as this would allow researchers to
change the device parameters to reveal how real brains function.”
How computers work
To
understand why this device is such a departure from what went before,
it is necessary to understand how computer systems store and compute
information. Within the nano-scale circuitry of today’s computer chips
are billions of tiny electrical components—transistors—that convey
information using binary logic. That is, their logic is based on two
numbers, either 1 or 0. In electrical terms, a transistor is either “on”
or “off.”
With
enough transistors packed into each chip, programmers can manipulate
electrical circuits, turning the billions of transistors on or off as
necessary to store and process information—to “compute.” The speed and
size of computer chips has largely been determined by our ability to
create faster transistors and to pack them into smaller spaces.
Synapses
are the smallest computational units in the brain, but different from
transistors in at least two very important ways. First, they can vary in
strength. In other words, synapses can convey far more information than
a transistor. Second, synapses can change over time.
“Synapses
change based on learning,” says Jeyasingh, “something conventional
computers cannot do. Once most computer chips are made, you cannot
change them easily.”
Practice makes perfect
In
neuroscience, these two advantages are combined in a concept known as
“synaptic plasticity,” one of the leading theoretical foundations for
how our brains learn, remember and compute.
Like
transistor circuits, neurons and synapses are small and packed tightly
together, but their circuitry is based on the varying strength of the
synapses. The repetition of electrochemical signals traveling the same
path will reinforce the synapses in the path and make them more or less
likely to fire in the future. As neuroscientists like to say, “Neurons
that fire together, wire together.”
Synaptic
plasticity explains why practice makes perfect. Repeating an electrical
pattern through practice strengthens the pattern; therefore, the brain
“learns.”
“The
brain is an amazing machine. Its circuits are far more complex, far
more capable, far more energy efficient and far more powerful for
performing certain tasks than even the very best computer chips based on
binary logic,” says Kuzum.
The
Stanford team’s device emulates synaptic plasticity using a technology
known as “phase-change material,” the same technology that allows DVDs
and CDs to store information. When juiced with electricity, these
materials change their physical characteristics and therefore their
electrical conductivity in tiny increments—more electricity, more
change.
Rather
than the two states of a transistor, however, the Stanford team has
demonstrated an ability to control the synaptic device in 1%
increments—like a lightbulb on a dimmer—meaning each phase-change
synapse can convey at least 100 values.
The device can be manufactured using existing commercial equipment with readily available materials.
“Using
well-understood manufacturing processes, we can construct a cross-point
architecture allowing three-dimensional stacking of layers that could
one day approach the density, compactness, and massive parallelism of the
human brain,” says Kuzum.
The
researchers do not, however, foresee their new chips replacing existing
ones. Instead, they say, they will lead in promising and exciting new
directions that are currently out of reach.
“Our
long-term goal is not to replace existing chips, but to define a
fundamentally distinct form of computational devices and architectures.
These new devices and architectures will excel at distributed,
data-intensive algorithms that a complex, real-world environment
requires, the sort of algorithms that struggle through today’s
processing bottlenecks,” says Kuzum.
Thinking in parallel
Among
the most intriguing possibilities of these synaptic devices is greater
parallelism. The brain is very good at juggling many types of sensory
information simultaneously, something computers do very poorly. A
supercomputer, by comparison, does not owe its great power to the speed
of its processors so much as to splitting up big problems among many
processors, each working on a small part of the problem. A more
brain-like architecture might allow much smaller chips to think in
parallel on many things at the same time.
And
where might this lead us? It could lead to real-time brain simulations
for use in neuroscience that may augment our understanding acquired from
biological measurements. Brain-inspired computers may prove
particularly adept at making decisions based on probability, as well.
“This
work is a promising step forward in our ability to emulate brain
functions using nanoelectronic devices and circuits. We can now
contemplate a new direction of research which utilizes nanoelectronics
for the study of neuroscience,” says Wong.
“Beyond
neuroscience, more brain-like systems could find use at the
intersection of sensing and computation,” says Kuzum. “Such applications
would be able to process huge amounts of sensory data in parallel,
meaning computers that can process visual information, recognize images,
or aid in navigation.”
On
a more fundamental level, the work is likely to produce a deeper
understanding of the physics of gradual control of phase-change
materials, allowing for additional fine tuning of the synaptic devices
and even greater processing ability.
“This is a significant development,” says Wong. “And we are excited to see where it leads.”
This
work is supported by DARPA SyNAPSE through a collaboration with IBM
Research, the National Science Foundation, and the Nanoelectronics
Research Initiative of the Semiconductor Research Corporation.
Nanoelectronic Programmable Synapses Based on Phase Change Materials for Brain-Inspired Computing
