A new brain-inspired chip uses 10,000× fewer calculations by copying the cerebellum, not the cerebrum

A new chip from Northwestern University flagged irregular heartbeats with 98% accuracy on simulated heart-rhythm data, using 10,000× fewer calculations than conventional AI, according to a paper published July 10 in Nature Communications and re-reported by LiveScience. The team didn't borrow a vague "brain-inspired" label. They copied a specific part of the brain: the cerebellum, the small region at the back of the skull that handles coordination and reflex, rather than the "thinking" cerebrum most neuromorphic research has targeted.

That choice is the point of the paper. The cerebellum's job is to filter routine signals and only fire on surprises; the cerebrum's job is to reason broadly. For a chip designed to wake up only when something goes wrong, the cerebellum is a better template, and the McCormick School's engineering account frames the work as the first time that bet has been tested on a benchmark at this scale.

Lead author Mark Hersam, a Northwestern professor of materials science and engineering, has called the cerebellum "a prime, untapped candidate" for neuromorphic design, because its competing excitatory and inhibitory circuits naturally do novelty detection, flagging the unexpected and ignoring the background. The team's hardware is built around that balance.

The hardware is a memtransistor, a transistor that remembers past signals and weights new ones against them. Conventional AI runs every input through the same dense math. The memtransistor skips most of that, prioritizing the small share of inputs that don't match the running pattern. On a simulated ECG (electrocardiogram) trace, that meant flagging an arrhythmia (an irregular heartbeat) within one-fifth of a heartbeat, twice as fast as conventional AI, and with 98% accuracy, per the team's release mirrored on EurekAlert.

The 10,000× number sits inside that specific test. It is not a general AI efficiency claim; it is what the team's setup measured on the ECG task, against a conventional AI baseline. That distinction matters: neuromorphic chips are not yet running production AI workloads, and this paper does not claim they are. The result is a research-stage demonstration that a small, surprise-first architecture can match conventional accuracy on a narrow problem while spending a fraction of the compute.

The use cases the team points to are the standard low-power-AI wish list: always-on health monitors, self-driving cars, autonomous robots. These are settings where latency and battery are the binding constraints, and where routing every signal to a data center is impractical. None of them is a product today. Hersam's framing, in the Northwestern release, is that the paper is a step toward that class of device, not a delivery of one.

What's still open is concrete. The ECG test is simulated, on a benchmark trace, in a controlled setup. The team has not yet reported the chip running on real hardware in a wearable, on a car's sensor stack, or on a robot. The memtransistor fabrication is research-grade. The 10,000× gap will likely narrow when the comparison is moved off a benchmark and onto a general task, because conventional AI is not standing still either. The honest read is that this paper shows the cerebellum move works once, on a specific test, with a specific chip design, and that the next question is whether it survives contact with the messier data those use cases actually produce.

For a reader new to the field, the working mental model is this: neuromorphic computing is the project of building chips that process less, react faster, and run on smaller devices, by taking cues from how neurons actually fire. Most of that project has copied the "thinking" brain. This paper argues, with a Nature Communications result, that the noise-filter brain is at least as promising, and shows one specific way to make that argument count on a benchmark today.