Chip Modeled on the Brain’s “Reflex” Region Spots a Bad Pulse in a Fifth of a Beat — Using 10,000x Fewer Operations

Right now, tucked at the back of your skull, a fist-sized lump of tissue is doing almost nothing. Your cerebellum is watching, not thinking. It sits idle while you walk, reach, blink, and breathe, holding its fire until the moment your foot catches an unseen curb or the coffee cup turns out heavier than you guessed. Then, in a few thousandths of a second, it acts.

That habit of staying still until something surprising happens is, it turns out, a good blueprint for a computer chip. A team at Northwestern University has built one that borrows the trick, and in early tests it flagged an abnormal heartbeat within about a fifth of a single beat, using roughly 10,000 times fewer computing operations than a conventional AI would need.

The problem they were chewing on is a familiar one to anyone who has watched a data center’s electricity bill. Modern AI is brilliant at spotting patterns, but it does so by grinding through every scrap of incoming data, over and over, whether or not anything has changed. “Today’s AI is remarkably good at recognizing patterns, but it often spends enormous amounts of computing power to continuously analyze streams of data,” says Mark Hersam, who co-led the work. As he puts it, the machine “burns energy on unnecessary analysis.”

Most brain-inspired computing has looked to the cerebrum, the wrinkled outer bulk we tend to think of as the seat of thought. Hersam’s group went somewhere less fashionable.

The lopsided little transistor

“In our work, we developed a device that mimics the cerebellum, which controls reflex reactions seemingly without even thinking,” says Hersam. The cerebellum, he adds, “is excellent at ignoring the expected and reserving its resources for reacting to the unexpected.” That indifference to the ordinary is exactly what makes it cheap to run, in energy terms, and it’s the property the team wanted in silicon (or rather, in something thinner than silicon).

Their device is a memtransistor, a component that folds memory and computation into one place instead of shuttling data back and forth between separate chips, which is where a lot of ordinary computing energy leaks away. The team built theirs from a polycrystalline film of molybdenum disulphide barely a single atom thick, then did something slightly odd to its geometry: one electrode was made to overlap the semiconductor through a thin insulating layer, leaving the whole thing deliberately asymmetric.

That asymmetry is the clever bit. In the cerebellum, two signals push against each other, one excitatory and one inhibitory, and most of the time they cancel out into a kind of watchful equilibrium. When something unexpected arrives, the balance tips and a cell called a Purkinje neuron briefly goes silent, and that pause is the brain’s flag for novelty. The Northwestern device reproduces both halves of the tug-of-war in a single lopsided component. Push voltage one way and it behaves like an excitatory synapse, its response building slowly as signals repeat, with a decay time of about 2.43 seconds. Reverse the voltage and the same device flips to inhibition, answering strongly at first and then fading fast, in roughly 0.28 seconds. One device, two opposite personalities, chosen by which way the current runs.

To see whether the idea held up, the researchers fed the system a stream of electrocardiogram recordings drawn from a public database, healthy rhythms interleaved with dangerous ones. The chip largely ignored the normal beats, as designed.

Faster than the beat itself

Then the arrhythmia hit. “Our cerebellum-inspired memtransistor detected an irregular heartbeat within a fraction of a second, before the heartbeat even ended,” says Hersam, and by his account that is “more than twice as fast as conventional AI.” Across the tests the approach hit better than 98 percent accuracy inside a fifth of a beat, a window in which the standard rivals (transformers and various neural networks among them) were still stuck somewhere between 70 and 80%.

Only the memtransistors themselves were physically built and measured; the surrounding network, the part that encodes signals and makes the final call, was simulated. The much-quoted 10,000-fold saving counts computing operations, not the full electricity draw of a finished gadget, which would also have to reckon with memory access and wiring and all the fiddly peripheral circuitry.

There is also a limit baked into the biology they copied. This is short-term memory only; the device holds its state for a few hours, no more. The real cerebellum can also learn, gradually filing a once-startling event under “expected” if it keeps happening, and the chip cannot yet do that. Which is the point of what comes next.

If the approach scales, the appeal is obvious enough for anything that has to stay switched on and alert without a power cord or a distant server to lean on. Think wearable heart monitors that catch the first flutter of trouble, or a robot that notices a person stepping into its path, or a security system twigging a strange pattern on a network before it curdles into an attack. The common thread is a machine that mostly dozes and only wakes for the new, which is roughly how animals have managed to think on a calorie budget for a very long time.

“We have demonstrated one part of the cerebellum neural circuit, but there is more that we have not yet emulated,” says Hersam. The next stretch of road, he reckons, runs toward a chip that can learn what to stop being surprised by.

Frequently Asked Questions