Sound Waves Give Neuromorphic Chips a Brain-Simulating Edge
IEEE Spectrum Charles Q. Choi
Scientists built a brain chip that computes with sound instead of electricity. It's way more efficient and could let one small circuit handle many different jobs, like real neurons do.
Based on reporting by IEEE Spectrum, Charles Q. Choi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Neuromorphic chips are supposed to copy the brain's trick of merging memory and computing in one place, saving energy that regular chips waste shuttling data back and forth. The problem is that most of these chips are still crude compared to biology. A human neuron might have thousands of synapses, sometimes over 100,000 in the case of Purkinje cells in the cerebellum. Most artificial neuromorphic devices, by contrast, act like a single synapse, so stacking enough of them to do anything complex means piling on wires, power draw, and cost.
A team led by Xiaodong Yan at the University of Arizona thinks sound waves might get around that bottleneck. Instead of encoding information as ones and zeros in silicon, they built a device using three aluminum rods, each about 60 centimeters long, glued together with epoxy and fitted with ultrasonic transmitters and sensors attached using a layer of honey. Sound waves traveling through the rods carry what the researchers call phi-bits, units that encode multiple values in a wave's phase rather than a single binary state. These aren't quantum bits, just classical stand-ins that borrow some of the same math, but they let one physical device juggle several computations at once instead of needing a separate chip for each.
The payoff showed up when the team tested their acoustic synapse against a standard neural network on a classic benchmark: sorting 150 iris flowers into three species. The sound-based system hit 96.7 percent accuracy using just 39 parameters and got there 20 percent faster than a conventional multilayer perceptron, which would have needed nine neurons and more parameters to match it. The researchers estimate their setup uses at most a tenth of the power of today's best electronic neuromorphic hardware.
What's more interesting than the speed gain is the flexibility. Real synapses get tuned by neuromodulators like dopamine and serotonin, chemicals that can make a connection more sensitive, faster, or slower depending on context, sometimes with ten different modulators acting on one synapse at once. Replicating that in electronic hardware usually means redesigning the whole circuit. With the acoustic version, the team found they could mimic several neuromodulatory effects, both fast dopamine-like responses and slow stress-like ones, just by bolting on an extra rod. Sandia researcher Brad Aimone, who wasn't involved in the work, points out that this could eventually mean smaller networks that reconfigure themselves for different tasks rather than needing a separate trained model for everything.
The work, published June 12 in Science Advances, is still lab-bench stuff involving 60-centimeter metal rods and ultrasonic sensors, nowhere near a shippable chip. But the core idea, using physical wave dynamics to do parallel computation cheaply, is the kind of unconventional approach that neuromorphic research has needed for a while.
My take — AI-written commentary, not fact-checked reporting
I like this precisely because it's weird: gluing metal rods together with honey to beat silicon at efficiency is the sort of left-field engineering that actual progress tends to come from, not another slightly bigger transformer. The neuromodulator trick is the real story here, since one adaptable circuit beating a zoo of specialized networks is exactly the direction AI hardware needs to go if we want efficiency gains that don't just come from throwing more GPUs at the problem.
Read more about this at: IEEE Spectrum