DoD Eyes 'Brain-Like' Supercomputers
August 11, 2017
The human brain is remarkably efficient. Using a few dozen watts of
energy, it performs information-processing functions that would take a
conventional computer millions of watts to replicate. But a new Defense
Department-funded IBM computer chip could enable machines to start
Qing Wu, an Air Force Research
Laboratory principal electronics engineer, holds a TrueNorth computer
chip at Department of Defense Lab Day, a biennial technology exhibit, at
The TrueNorth computer chip is a "neuromorphic"
chip that mimics human neurons and performs unusually advanced
computations using far less energy than conventional chips, said Qing
Wu, principal electronics engineer at the Air Force Research Laboratory
at Wright-Patterson Air Force Base, Ohio. The technology could be a huge
boost for artificial intelligence-based defense systems and the human
data analysts who use them, he added.
"This is about building more intelligent machines that will work with
humans to make human operators and analysts be more effective and
efficient when dealing with data," Wu said. "The major advantage of this
chip is it runs machine learning algorithms -- the same ones as we run,
the same functionality, same accuracy, but with much less power
In June, IBM announced that it will build a new supercomputer powered by
64 TrueNorth chips for the Air Force Research Lab, which lab officials
plan to use for analytics involving pattern and object recognition and
“sensory processing” -- converting audio, video and other forms of data
received by sensors into symbols that the computer can process. Though
conventional computers can perform these tasks, they require huge
numbers of processors, which consume heavy amounts of electricity.
TrueNorth brings the wattage down. Dharmendra Modha, the lead researcher
of IBM’s brain-inspired computing group, reported in an article posted
on the company's website that the TrueNorth chip uses no more than 70
milliwatts of power, which he said is “four orders of magnitude lower"
than a conventional computer chip’s power consumption.
"The architecture … has the potential to revolutionize the computer
industry by integrating brain-like capability into devices where
computation is constrained by power and speed," Modha said in a
statement in 2014, when IBM built the first chip.
A conventional chip has a central processing unit, but TrueNorth
contains a million "neurons" that transmit data back and forth in a way
similar to human brain neurons, Modha explained. The neurons communicate
throughout the system using patterns of pulses similar to the way human
neurons use electrochemical pulses, he said.
TrueNorth's neurons are packed in clusters inside interconnected "cores"
across the chip, IBM officials said in the June 23 news release
announcing the company's collaboration with the Air Force Research Lab.
Each core also holds components for information storage, processing, and
Dharmendra Modha, an IBM engineer,
presents two basic neuromorphic computers that have 16 TrueNorth chips
each at the IBM Almaden Research Center in San Jose, Calif., March 24,
2016. IBM is now building a scaled-up 64-chip computer for the Air Force
Research Laboratory, equivalent to four of these pictured computers. It
will be the largest array of TrueNorth chips yet built.
By contrast, a conventional chip
stores information in a memory drive and processes it in the central
processing unit, he said. It constantly shuttles the data back and forth
between them and burns energy along the way. Because TrueNorth's cores
do both data storage and data processing, these energy-intensive data
swaps are eliminated.
The human brain also integrates thought and memory, Modha said, which
partly explains why the average brain consumes only 20 watts of energy.
The IBM Sequoia, one of the fastest conventional supercomputers, has
less computing power than the brain and consumes 7.9 megawatts.
An AI Data-Analysis Partner
William Halal, a George Washington University professor of management,
technology and innovation, and founder of the technology-forecasting
think tank TechCast, said a neuromorphic computer like TrueNorth can
"think" in ways few conventional computers can: It excels at parallel
processing -- that is, running multiple calculations at once -- and
interpreting, finding patterns or drawing conclusions from data, he
Halal pointed out that while these data-reading skills come naturally to
humans, they're difficult for conventional computers. Where conventional
computers store and process data, but need human users to tell them
which data is most important and what to do with it, TrueNorth has no
such limitations, he said. It could discern in advance what the user
might want to know and gather data accordingly -- or connect sets of
data to spot a trend all on its own.
"The real advantage to this computing is that it operates differently,"
Halal said. "It's more intelligent. It does things the way humans think.
It offers the prospect of modeling these complex human cognitive
processes that have resisted being developed with the present
TrueNorth's heightened capabilities could help human defense analysts
comb through data and spot vital information more quickly, said Mark
Barnell, a senior computer scientist with the Air Force Research Lab's
information directorate. He said he hopes that this, in turn, could
enable military planners at all levels to make better-informed decisions
in less time.
"The computer could look through the data quickly and tell us if there
is something interesting and if there is something worth looking more
at," Barnell said. "It would close the timeline of collecting the data
and disseminating information."
TrueNorth's story starts in 2008, when the Defense Advanced Research
Projects Agency launched its Systems of Neuromorphic Adaptive Plastic
Scalable Electronics program. The mission was to build computer systems
whose functioning resembles a living mammal's brain -- including the
brain's learning and problem-solving abilities. SyNAPSE contracted the
research work to IBM and HRL Laboratories.
This program ran until 2014, by which time IBM had pioneered TrueNorth.
And the research and development continues today at the Air Force
Research Lab's facilities. The lab's upcoming computer will be the first
time that so many chips will work together on one system, Barnell noted.
DoD and AI
said energy efficiency is one reason that he and colleagues are looking
forward to this computer, but that he's anticipating something even
greater: This new system, with its brain-like architecture, will be a
very real step toward true artificial intelligence, he said, becoming a
powerful way of doing computation that in some ways mimics biological
Artificial intelligence is already integral to many DoD operations, said
Craig Arndt, a Defense Acquisition University professor of systems
engineering. He cites facial-recognition systems, unmanned vehicles, and
"predictive maintenance" systems that identify internal mechanical
problems and alert human operators of them as several examples. DoD
supported development of each one and found defense-relevant purposes
for each, he said.
"AI has been around for a long time, and is a very broad area of
research, and DoD has been involved in that research at its service labs
pretty much the entire time," Arndt said. "It has been an important area
of computer science for us because of the problems it tries to solve."