Northwestern University engineers have built a brain-like chip that mimics the cerebellum's ability to ignore routine events and react only to novelties, cutting AI power consumption by orders of magnitude. The device detected abnormal heart rhythms in a fraction of a heartbeat with more than 98% accuracy while using roughly 10,000 times fewer computer operations than conventional AI methods, according to a study published July 10 in Nature Communications.
The work, co-led by Mark C. Hersam, aims to move AI hardware beyond pattern classification toward always-on systems that instantly flag unexpected changes-without relying on energy-hungry data centers. Potential applications include wearable health monitors, self-driving cars, autonomous robots, and cybersecurity sensors.
Mimicking the cerebellum's reflex-like efficiency
Unlike typical brain-inspired computing that replicates the cerebrum, the new device copies the cerebellum, which handles reflex reactions and filters out predictable input. "The cerebellum is excellent at ignoring the expected and reserving its resources for reacting to the unexpected," Hersam said. "That approach ultimately translates into lower energy consumption, and that is where we achieve orders of magnitude improvement."
Hersam's lab previously demonstrated that just two memtransistors-devices that combine memory and computation-could perform classification tasks that would need over 100 conventional transistors. The new design goes further by recreating a cerebellar circuit that balances excitatory and inhibitory signals to detect novelty.
How excitatory and inhibitory signals detect novelty
In the cerebellum, neural circuits maintain a steady equilibrium between excitatory and inhibitory signals. When something surprising happens, that balance shifts briefly. The Northwestern team built a memtransistor that can switch between two modes: an excitatory mode that strengthens its response with repeated signals, and an inhibitory mode that fires strongly at first then fades. This dual behavior lets the device distinguish routine data from genuinely novel events.
To achieve this, the engineers used molybdenum disulfide, an atomically thin semiconductor, and designed an asymmetric transistor architecture. One electrode partially overlaps the semiconductor through a thin insulating layer, altering current flow. Reversing the applied voltage toggles the memtransistor between its excitatory and inhibitory functions.
Heart rhythm detection at millisecond speeds
In tests, the system received electrocardiogram recordings containing both normal heartbeats and arrhythmias. It ignored the normal beats entirely but flagged an irregular heartbeat within milliseconds. "Our cerebellum-inspired memtransistor detected an irregular heartbeat within a fraction of a second, before the heartbeat even ended," Hersam said. "That is more than twice as fast as conventional AI."
The device achieved this by not wasting energy on continuous analysis of unchanging data. "Today's AI is remarkably good at recognizing patterns, but it often spends enormous amounts of computing power to continuously analyze streams of data-even when nothing has changed," Hersam said. "Therefore, it burns energy on unnecessary analysis."
For those working in AI for IT & Development, the research points to new possibilities for energy-efficient edge computing, where low-power, always-on inference could reshape system architectures.
Why this matters for IT and Development
This memtransistor-based design demonstrates that novel hardware can cut AI power consumption by thousands of times while improving real-time responsiveness. For IT infrastructure teams managing continuous monitoring or edge device fleets, the breakthrough suggests that silicon-like efficiency gains are achievable without waiting for cloud data centers. Development teams can begin exploring how event-driven, low-power inference might replace constant stream processing with change-triggered analysis, reducing bandwidth and compute loads in IoT and security deployments.
The researchers plan to expand the design to mimic the cerebellum's ability to adapt over time-learning to treat repeated events as non-novel-potentially making these systems even more efficient.
Your membership also unlocks:
