Engineers at Northwestern University have developed an artificial intelligence device that identifies abnormal heart rhythms within one-fifth of a heartbeat with more than 98% accuracy. Inspired by the cerebellum, the brain region responsible for reflex responses, the device achieves this near-instant detection with about 10,000 times fewer computer operations than traditional AI. This energy efficiency comes from mimicking how the cerebellum prioritizes unexpected changes and reserves resources for new events, rather than processing a constant flow of normal data. “In our study, we developed a device that mimics the cerebellum, which controls reflex responses seemingly without thinking,” explains Mark C. Hartham of Northwestern University, who co-led the study. This breakthrough enables low-power, always-on AI systems for a variety of applications, from wearable health monitors to autonomous robots.
Cerebellum-inspired memtransistor design enables efficient novelty detection
The pursuit of energy-efficient artificial intelligence has shifted the focus from the cerebrum to the cerebellum. A new device developed by researchers at Northwestern University enables near-instant novelty detection with significant power savings. This performance improvement stems from a fundamental redesign of the AI hardware, which goes beyond mimicking the brain’s processing centers to emulate the cerebellum’s specialized role in filtering information and responding to change. The team’s innovation is centered around the memtransistor, a device that combines memory and computation into a single unit, significantly reducing energy consumption. This new iteration requires approximately 10,000 times fewer computer operations than standard AI approaches. “With brain-like computing, researchers seek to mimic the cerebrum, which is typically considered the brain’s ‘thinking center,'” explains Mark C. Hartham of Walter P. Kennedy.
The asymmetric design of the memory transistor, made from molybdenum disulfide, an atomically thin semiconductor, allows it to seamlessly switch between excitation and inhibition modes based on voltage direction. This allows the AI to ignore routine data to save power and react instantly to unexpected events, such as irregular heartbeats identified within a fifth of a heartbeat. Researchers envision applications ranging from wearable health monitors and self-driving cars to cybersecurity systems that would benefit from this low-power, always-on capability. Hersam’s team will continue to build on this early success and refine the design to further emulate the cerebellum’s adaptive learning capabilities.
Memory transistors inspired by our cerebellum detected irregular heartbeats within seconds, before the heartbeat had finished.
This design allows the device to function as both excitatory and inhibitory synapses and reflect competing signals within cerebellar circuits. The research team was able to recreate the dynamics of the cerebellum, where signals remain balanced during normal activity, but change when unexpected events occur. This is achieved by memory transistors that switch between modes simply by reversing the applied voltage. Hartham, co-lead author and Walter P. Murphy Professor of Materials Science and Engineering at Northwestern University, explained that by ignoring routine data and focusing only on anomalies, the device can significantly reduce energy consumption. Researchers are now focused on enabling devices to learn and adapt over time, further refining their ability to distinguish between truly new and familiar events.
Although we have demonstrated parts of the cerebellar neural circuitry, there are still parts that we have yet to emulate.
Rapid ECG analysis demonstrates identification of sub-heartbeat abnormalities
Researchers are increasingly looking to biological systems for inspiration for artificial intelligence, and a team led by Mark C. Hartham at Northwestern University has demonstrated an efficient approach by modeling AI hardware after the cerebellum. Unlike traditional brain-inspired computing, which focuses on the cerebrum, Hersam’s group has developed a device that mimics the cerebellum’s ability to quickly detect unexpected changes, achieving significant reductions in energy consumption. This efficiency comes from a fundamental change in design. The team aggregated memory and computation into a single component called a memtransistor, reducing the need for constant data transfers between separate processing units. A previous study published in Nature Electronics in 2023 showed that AI classification tasks that traditionally required more than 100 transistors could be performed with just two memory transistors, reducing energy usage by a factor of 100.
Current research builds on this foundation by designing memory transistors to mimic cerebellar circuits responsible for novelty detection. This device distinguishes between expected and unexpected events by utilizing both excitatory and inhibitory signals, reflecting the dynamic equilibrium found in the cerebellum. To test the device, researchers analyzed electrocardiogram recordings and found that it was able to ignore normal heartbeats and identify abnormal heart rhythms within one-fifth of a heartbeat. This speed highlights the potential for real-time condition monitoring and other applications requiring immediate response.
In the world of brain-like computing, researchers typically try to mimic the cerebrum, which is considered the brain’s “thinking center.”
