The human brain is astonishingly efficient. It runs on about 20 watts, roughly the power required to keep a dim bulb lit. Over the last three decades, computer scientists in a field called neuromorphic computing have sought ways to replicate that energy efficiency in computing devices modeled on how networks of synapses process information in the brain. Recent years have even seen the production of neuromorphic computing chips from companies like Intel and IBM.
But there’s a trade-off: That efficiency gain in neuromorphic chips comes at the cost of slower computing speed, says Shantanu Chakrabartty at Washington University in St. Louis. “We need to break this barrier,” he says. “Ideally, you want to get more bang for the buck for the amount of energy you spend.”
Stochastic thermodynamics, a subfield of physics, may help. Researchers in the field study the energy cost of computation in systems out of equilibrium, which includes modern computers. Bringing it into the world of neuromorphic computing is a natural pairing, says SFI Professor David Wolpert. “These fields have so much to say to one another.”
Wolpert and Chakrabartty, who met two years ago, organized a working group, held at SFI from December 10–12, designed to give experts from both fields the opportunity to learn from each other and exchange ideas. The 13-person working group was supported in part by a grant from the National Science Foundation, which had requested the meeting.
After decades of work, said Chakrabartty, researchers have investigated many architectures in pursuit of what’s called the “neuromorphic advantage.” Even so, it’s still an aspirational approach, and there’s plenty of uncharted territory — especially in finding ways to harness the random noise and small-scale fluctuations that naturally arise during computation. “We just haven’t figured out how to exploit them efficiently,” says Chakrabartty. “And that’s where stochastic thermodynamics comes in.”
Current neuromorphic approaches, says Wolpert, focus on the phenomenology — the observed performance — of the device, rather than the underlying physics. Stochastic thermodynamics points to a way to gain a deeper understanding. “This meeting was required for neuromorphic to really move forward,” he says.
One of the goals of the working group, Chakrabartty says, was to start a conversation in which researchers in neuromorphic computing could give the stochastic thermodynamics community an overview of the state of the art in their field — and vice versa.
“This was a first date between the two communities,” says Wolpert — one that went very well. As soon as participants from each field learned what the others were doing, they grasped the connection so quickly it seemed obvious, he says. The meeting concluded with plans for future collaborations and meetings.
Read more about the working group "NeST: Neuromorphic Stochastic Thermodynamics."
This meeting was supported by the National Science Foundation under Grant No. 2529902 entitled "Conference: NeST: Neuromorphic Stochastic Thermodynamics Workshop at SFI".