(Image Credit: BrianPenny/pixabay)
Northwestern University engineers developed an AI chip that behaves like a cerebellum. Rather than continuously analyzing, the chip reacts to unexpected events. During tests, the device detected arrhythmias within one-fifth of a heartbeat and with over 98% accuracy. The team says the chip could lead to low-power, always-on AI systems for wearable health monitors, self-driving vehicles, autonomous robots and cybersecurity systems.
“In the world of brain-like computing, researchers typically try to mimic the cerebrum, which is often viewed as the brain’s ‘thought center,’” said Northwestern’s Mark C. Hersam, who co-led the study. “In our work, we developed a device that mimics the cerebellum, which controls reflex reactions seemingly without even thinking. The cerebellum is excellent at ignoring the expected and reserving its resources for reacting to the unexpected. That approach ultimately translates into lower energy consumption, and that is where we achieve orders of magnitude improvement.”
The team is building on earlier work on improving AI hardware by combining memory and computation into a memtransistor. Modern computers consume a lot of power by shuttling data between separate memory and processing units. In a 2023 study, they demonstrated that two memtransistors performed AI classification requiring over 100 transistors while consuming 100 times less energy.
Their AI chip goes even further. It replicates the cerebellum’s internal circuitry that detects unexpected events and makes quick decisions. Instead of analyzing routine information, it detects unexpected changes, including an irregular heartbeat in a wearable, a human stepping into a robot’s path, or potential cyber threats.
“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.”
The cerebellum balances excitatory and inhibitory signals to determine if there’s unrecognized activity or unexpected events. Typically, those two signals offset each other. However, a sudden change affects that balance, prompting a response. With that in mind, the team replicated that behavior in a memtransistor with two operating modes. In one mode, the device’s response strengthens as stimulation continues. The other mode reacts strongly to an earlier signal before quickly fading.
To switch between the two modes, the applied voltage must be reversed. The device’s behavior occurs due to its asymmetric transistor design. In this case, it uses an extremely thin layer of molybdenum disulfide and an electrode that partially overlaps it via a thin insulating layer. The device uses these features to distinguish routine activity from unexpected changes without constantly processing all the incoming signals.
They used ECG recordings containing normal heart rhythms and arrhythmias to test the AI chip. It didn’t use up power by analyzing every heartbeat, ignoring the normal ones instead. The system also detected an abnormal heartbeat within milliseconds.
The team wants to make the chip learn and adapt over time. For instance, repeated events eventually become familiar, causing the brain to no longer treat them as novel. “We have demonstrated one part of the cerebellum neural circuit, but there is more that we have not yet emulated,” Hersam said. “We intend to continue going down this path to mimic more and more of this complicated system.”
Have a story tip? Message me here at element14.