Cerebellum-Inspired AI Chip Detects Unexpected Events With Ultra-Low Power

Scientists have developed a new artificial intelligence (AI) chip inspired by the human brain’s ability to respond rapidly to unexpected events. The technology mimics the cerebellum, the brain region responsible for balance, coordination and precise motor control.

Unlike conventional AI systems that continuously analyze every piece of incoming data, the new chip is designed to filter out routine information and react only when it detects something unusual. This approach could lead to faster, more energy-efficient AI for health monitoring, autonomous vehicles, robotics and cybersecurity.

In simulated tests using electrocardiogram (ECG) data, the cerebellum-inspired device identified irregular heartbeats, or arrhythmias, within one-fifth of a heartbeat and with 98% accuracy. According to the researchers, it operated more than twice as fast as conventional AI systems while using approximately 10,000 times fewer computational operations.

The research team published its findings July 10 in the journal Nature Communications.

The technology could help create always-on AI systems that identify and respond to anomalies locally, without depending on powerful data centers or cloud computing. Potential applications include wearable health monitors, self-driving cars, autonomous robots and other edge-AI devices.

A New Approach to Neuromorphic Computing

Computer systems modeled on the human brain are known as neuromorphic computers. Researchers are developing these systems to make AI more efficient by imitating how biological neurons communicate and process information.

The human brain does not process every incoming signal with the same intensity. Instead, its neural circuits prioritize important information and ignore predictable background activity. This filtering system helps the brain conserve energy while remaining ready to respond to sudden changes.

Many neuromorphic computing systems are modeled primarily on the cerebrum, the brain’s largest region and its main center for conscious thought. In this study, however, the researchers focused on the cerebellum, a smaller brain structure associated with coordination, balance and instinctive motor responses.

Neural circuits in the cerebellum contain both excitatory and inhibitory signals. These opposing signals normally balance one another. When an unexpected event occurs, that balance changes, alerting the brain that it needs to respond.

This ability to ignore predictable activity while focusing on novelty makes the cerebellum a promising model for low-power AI systems, said study co-author Mark Hersam, a professor of materials science and engineering at Northwestern University.

“The cerebellum is excellent at ignoring the expected and reserving its resources for reacting to the unexpected,” Hersam said in a statement. “That approach ultimately translates into lower energy consumption.”

Memtransistors Combine Memory and Processing

Modern AI is highly effective at recognizing patterns, but it often requires substantial computing power to analyze continuous streams of information. This constant processing can increase energy consumption and place greater demands on data centers.

Another challenge comes from the traditional separation between computer memory and processing units. Moving data between these components creates delays and contributes to a limitation known as the von Neumann bottleneck.

The new AI chip addresses this problem by combining memory and computation in a single component called a memtransistor — a device that combines the functions of memory and a transistor. By reducing the need to move data between separate components, the system can process information more quickly and efficiently.

The memtransistor is made from an atomically thin semiconductor called molybdenum disulfide. The material forms a channel between two electrodes. One electrode directly contacts the semiconductor, while the other is positioned partially above it and separated by a thin insulating layer.

This uneven structure changes the way electricity flows through the device. Reversing the direction of the voltage allows the memtransistor to switch between excitatory and inhibitory modes, imitating the competing neural signals found in the cerebellum.

The Output Layer of a Spiking Neural Network

The researchers designed the device to serve as the core of the output layer in a larger spiking neural network (SNN), Hersam explained in an email to Live Science.

Spiking neural networks process information through brief electrical pulses, or spikes, in a way that resembles communication between biological neurons. The researchers measured how individual memtransistors responded to repeated electrical pulses and used those results to model a network capable of distinguishing normal ECG patterns from arrhythmias.

They compared the cerebellum-inspired network with a standard transformer model, the AI architecture used by many large language models. The memtransistor-based system detected abnormal heart rhythms more than twice as quickly and required roughly 10,000 times fewer calculations, according to the study.

Although these results suggest that the technology could support faster and more energy-efficient AI hardware, real-world performance will depend on the size, speed and scalability of the devices.

“We have not scaled memtransistors to the level of commercial silicon chips, but in principle, 2D materials and memtransistors can be scaled to comparable sizes and operating speeds,” Hersam told Live Science.

Energy-Efficient AI at the Edge

The cerebellum-inspired AI chip could be especially useful for edge computing, where data must be analyzed quickly using minimal power. Edge devices process information locally rather than sending it to a remote cloud server, which can improve speed, privacy and reliability.

“One example could be edge computing, or in cases where access to the cloud is not available or is not desired due to the sensitivity of the data,” Hersam said.

Reducing AI’s dependence on data centers could have significant energy benefits. The International Energy Agency estimates that global data-center electricity demand could reach approximately 945 terawatt-hours by 2030, driven in large part by the growth of AI.

Future memtransistor networks could allow robots, autonomous vehicles and cybersecurity systems to operate continuously while consuming very little power. Instead of analyzing every signal, these systems could process data locally and respond only when they detect an unusual or potentially dangerous event.

“The current AI deployed in cybersecurity and autonomous vehicles uses the massive computing power of data centers or an in-house facility of GPUs and CPUs,” Hersam told Live Science. “In contrast, our approach promises low energy and power consumption by simplifying the number of operations needed. This is not an incremental performance improvement, but a fundamentally different way of detecting anomalies.”

The next phase of research will focus on reproducing another important feature of the cerebellum: its ability to adapt to predictable events. In the human brain, an event that initially seems unusual may eventually become routine after repeated exposure. Teaching AI hardware to make the same adjustment could improve the efficiency and accuracy of future anomaly-detection systems.

Test your knowledge of the most complex organ in the human body with our brain quiz!