Researchers developed a brain-inspired memtransistor that detects unusual events faster while consuming far less energy, paving the way for efficient always-on artificial intelligence systems.
Northwestern University researchers have developed a memtransistor that mimics the brain’s cerebellum, enabling faster and significantly more energy-efficient anomaly detection for artificial intelligence (AI) applications. The device, described in Nature Communications, identified abnormal heart rhythms within one-fifth of a heartbeat with more than 98% accuracy while requiring roughly 10,000 times fewer computing operations than conventional AI approaches.
Unlike traditional AI systems that continuously analyse incoming data, the new hardware remains largely inactive until it detects an unexpected event. Inspired by the cerebellum, which filters routine information and reacts rapidly to surprises, the device conserves energy by ignoring predictable signals and processing only meaningful changes.
The researchers recreated the cerebellum’s balance between excitatory and inhibitory neural responses using an asymmetric transistor architecture built from molybdenum disulphide (MoS₂). By reversing the applied voltage, the device switches between response modes, allowing it to distinguish ordinary activity from genuinely novel events without constant computation.
To evaluate the technology, the team processed electrocardiogram (ECG) recordings containing both normal and abnormal heartbeats. Rather than analysing every heartbeat in detail, the memtransistor ignored routine signals before detecting irregular rhythms within milliseconds. This approach more than doubled the detection speed of conventional AI while dramatically reducing energy consumption.
The research builds on earlier work that integrated memory and computation into a single device, eliminating the energy-intensive transfer of data between separate processors and memory units. According to the researchers, the latest design moves beyond efficient AI classification by enabling rapid recognition of unexpected events.
The technology could support a new generation of always-on AI systems, including wearable health monitors, autonomous robots, self-driving vehicles and cybersecurity platforms. Future research will focus on enabling the hardware to learn from repeated events, allowing it to adapt over time in a manner closer to the human cerebellum.
