Ever wish your computer could think like you do or perhaps even
understand you?
That future may
not be now, but it’s one step closer, thanks to a Texas A&M University-led
team of scientists and engineers and their recent discovery of a
materials-based mimic for the neural signals responsible for transmitting
information within the human brain.
The multidisciplinary team, led by Texas A&M chemist Sarbajit Banerjee
in collaboration with Texas A&M electrical and computer engineer R. Stanley
Williams and additional colleagues across North America and abroad, has
discovered a neuron-like electrical switching mechanism in the solid-state
material β’-CuxV2O5 — specifically,
how it reversibly morphs between conducting and insulating behavior on command.
The team was able to clarify the underlying mechanism driving this
behavior by taking a new look at β’-CuxV2O5, a remarkable
chameleon-like material that changes with temperature or an applied electrical
stimulus. In the process, they zeroed in on how copper ions move around inside
the material and how this subtle dance in turn sloshes electrons around to
transform it. Their research revealed that the movement of copper ions is the
linchpin of an electrical conductivity change which can be leveraged to create
electrical spikes in the same way that neurons function in the cerebral nervous
system — a major step toward developing circuitry that functions like the human
brain.
Their resulting paper, which features Texas A&M chemistry graduate
students Abhishek Parija (now at Intel Corporation), Justin Andrews and Joseph
Handy as first authors, is published Feb. 27 in the Cell Press journal Matter.
In their quest
to develop new modes of energy efficient computing, the broad-based group of
collaborators is capitalizing on materials with tunable electronic
instabilities to achieve what’s known as neuromorphic computing, or computing
designed to replicate the brain’s unique capabilities and unmatched
efficiencies.
“Nature has
given us materials with the appropriate types of behavior to mimic the
information processing that occurs in a brain, but the ones characterized to
date have had various limitations,” Williams said. “The importance of this work
is to show that chemists can rationally design and create electrically active
materials with significantly improved neuromorphic properties. As we understand
more, our materials will improve significantly, thus providing a new path to the
continual technological advancement of our computing abilities.”
While smart
phones and laptops seemingly get sleeker and faster with each iteration, Parija
notes that new materials and computing paradigms freed from conventional
restrictions are required to meet continuing speed and energy-efficiency
demands that are straining the capabilities of silicon computer chips, which
are reaching their fundamental limits in terms of energy efficiency.
Neuromorphic computing is one such approach, and manipulation of switching
behavior in new materials is one way to achieve it.
“The central
premise — and by extension the central promise — of neuromorphic computing is
that we still have not found a way to perform computations in a way that is as
efficient as the way that neurons and synapses function in the human brain,”
said Andrews, a NASA Space Technology Research Fellow. “Most materials are
insulating (not conductive), metallic (conductive) or somewhere in the middle.
Some materials, however, can transform between the two states: insulating (off)
and conductive (on) almost on command.”
By using an
extensive combination of computational and experimental techniques, Handy said
the team was able to demonstrate not only that this material undergoes a
transition driven by changes in temperature, voltage and electric field
strength that can be used to create neuron-like circuitry but also
comprehensively explain how this transition happens. Unlike other materials
that have a metal-insulator transition (MIT), this material relies on the
movement of copper ions within a rigid lattice of vanadium and oxygen.
“We essentially
show that a very small movement of copper ions within the structure brings
about a massive change in conductance in the whole material,” Handy added. “Because
of this movement of copper ions, the material transforms from insulating to
conducting in response to external changes in temperature, applied voltage or
applied current. In other words, applying a small electrical pulse allows us to
transform the material and save information inside it as it works in a circuit,
much like how neurons function in the brain.”
Andrews likens
the relationship between the copper-ion movement and electrons on the vanadium
structure to a dance.
“When the copper
ions move, electrons on the vanadium lattice move in concert, mirroring the
movement of the copper ions,” Andrews said. “In this way, incredibly small
movements of the copper ions induce large electronic changes in the vanadium
lattice without any observable changes in vanadium-vanadium bonding. It’s like
the vanadium atoms ‘see’ what the copper is doing and respond.”
Transmitting,
storing and processing data currently accounts for about 10 percent of global
energy use, but Banerjee says extrapolations indicate the demand for
computation will be many times higher than the projected global energy supply
can deliver by 2040. Exponential increases in computing capabilities therefore
are required for transformative visions, including the Internet of Things,
autonomous transportation, disaster-resilient infrastructure, personalized
medicine and other societal grand challenges that otherwise will be throttled
by the inability of current computing technologies to handle the magnitude and
complexity of human- and machine-generated data. He says one way to break out
of the limitations of conventional computing technology is to take a cue from
nature — specifically, the neural circuitry of the human brain, which vastly
surpasses conventional computer architectures in terms of energy efficiency and
also offers new approaches for machine learning and advanced neural networks.
“To emulate the
essential elements of neuronal function in artificial circuitry, we need
solid-state materials that exhibit electronic instabilities, which, like
neurons, can store information in their internal state and in the timing of
electronic events,” Banerjee said. “Our new work explores the fundamental
mechanisms and electronic behavior of a material that exhibits such
instabilities. By thoroughly characterizing this material, we have also
provided information that will instruct the future design of neuromorphic
materials, which may offer a way to change the nature of machine computation
from simple arithmetic to brain-like intelligence while dramatically increasing
both the throughput and energy efficiency of processors.”
Because the
various components that handle logic operations, store memory and transfer data
are all separate from each other in conventional computer architecture,
Banerjee says they are plagued by inherent inefficiencies regarding both the
time it takes for information to be processed and how physically close together
device elements can be before thermal waste and electrons “accidentally”
tunneling between components become major problems. By contrast, in the human
brain, logic, memory storage and data transfer are simultaneously integrated
into the timed firing of neurons that are densely interconnected in 3-D
fanned-out networks. As a result, the brain’s neurons process information at 10
times lower voltage and an almost 5,000 times lower synaptic operation energy
in comparison to silicon computing architectures. To come close to achieving
this kind of energetic and computational efficiency, he says new materials are
needed that can undergo rapid internal electronic switching in circuits in a
way that mimics how neurons fire in timed sequences.
Handy notes that the team still needs to optimize many parameters, such
as transition temperature and switching speed along with the magnitude of the
change in electrical resistance. By determining the underlying principles of
the MIT in β’-CuxV2O5 as a
prototype material within an expansive field of candidates, however, the team
has identified certain design motifs and tunable chemical parameters that
ultimately prove useful in the design of future neuromorphic computing
materials, a major endeavor that has been seeded by the Texas A&M X-Grant
Program.
“This discovery
is very exciting because it provides fertile ground for the development of new
design principles for tuning materials properties and also suggests exciting
new approaches to researchers in the field for thinking about energy efficient
electronic instabilities,” Parija said. “Devices that incorporate neuromorphic
computing promise improved energy efficiency that silicon-based computing has
yet to deliver, as well as performance improvements in computing challenges
like pattern recognition — tasks that the human brain is especially
well-equipped to tackle. The materials and mechanisms we describe in this work
bring us one step closer to realizing neuromorphic computing and in turn
actualizing all of the societal benefits and overall promise that comes with
it.”
No comments:
Post a Comment