Neurogrid board replicates human brain circuitry
A circuit board features a prodigious boost in performance, running several times faster while using much less power than a typical computer—all because it replicates the human brain. Developed by biotechnology engineers at Stanford University, the board is akin to an Apple iPad in size and power consumption, and uses 16 custom-designed "Neurocore" chips to simulate the brain processes.
It was designed specifically to mimic the processes of human synapses as part of a National Institutes of Health effort to study the workings of the human brain by building tools that can read or can mimic the activity of one part of the brain at a time. The Neurocore board has potential for more than just neurological research, however, according to Kwabena Boahen, associate professor of bioengineering at Stanford and senior researcher, in a paper describing the neuromorphic circuit board published in the current issue of Proceedings of the IEEE.
Boahen: The human brain, with 80,000 times more neurons than Neurogrid, consumes only three times as much power. Achieving this level of energy efficiency while offering greater configurability and scale is the ultimate challenge neuromorphic engineers face.
The boards could also be developed into control systems for prosthetic limbs, using the combination of low energy demand and high-efficiency computing to run software that could identify the wearer's intentions by interpreting the motion of muscles or input from nerves in the leg and converting either into digital commands quickly enough to operate a robotic leg or arm almost as naturally as the original. Like most technical breakthroughs, however, that requires the creation of standards that would allow someone other than the inventor to understand the system's capabilities and write applications to expand them.
"Right now, you have to know how the brain works to program one of these," Boahen is quoted as saying in a Stanford announcement of the paper's publication. "We want to create a neurocompiler so that you would not need to know anything about synapses and neurons to able to use one of these."
The neurogrid is actually a combination of hardware emulation and software simulation. Rather than hard-wiring each of the millions of potential connecting points on a neuron, the Stanford team emulates the behaviour of those connections using as few as eight transistors each of which has a unique address that can be used to "softwire" transistors together to simulate the behaviour of a synaptic connection.
Each Neurogrid neuron has a total of 61 graded and 18 binary programmable parameters, allowing a single board to emulate many different areas of the brain, depending on how the board is programmed. The Neurocore chips themselves are CMOS processors built using a 180nm process, paired with a daughterboard containing 32MB of SRAM memory and a field programmable gate array to make and maintain connections among the virtual synapses.
(Video by Kurt Hickman)
The whole setup cost about $40,000 to build, but shrinking the circuits down to something more modern than the 15-year-old 180nm architecture it uses now would make each one far more power efficient and—if manufactured in bulk—far cheaper. Boahen believes it may eventually be possible to make the whole Neurogrid board for about $400, which would make the boards practical as supporting systems enhancing or replacing damaged biological circuitry. Neutrogrid circuits could, for example, be power- and space-controllers for prosthetic arms or legs, which could read human muscle- or nerve-signals and use them to operate the prosthetics as fast and transparently as the originals.
Boahen and fellow Stanford Professor Krishna Shenoy are working on ways to read brain signals about physical movement with the goal of creating a Neurocore-like chip that could be implanted in the brain to run prosthetics that would allow amputees or paralyzed people to use artificial limbs like the real thing. The key to applications like that is energy efficiency far beyond even what the current version of Neurogrid can manage.
Neurogrid is 100,000 times more energy-efficient than a PC would be in simulating the activity of a million neurons, but is still far more of an energy hog than an actual brain. "The human brain, with 80,000 times more neurons than Neurogrid, consumes only three times as much power," Boahen writes. "Achieving this level of energy efficiency while offering greater configurability and scale is the ultimate challenge neuromorphic engineers face."
- Kevin Fogarty
|Related Articles||Editor's Choice|