Beyond classification
This new device represents the latest advancement in the Hersam Lab’s broader effort to fundamentally rethink AI hardware. Traditional computers constantly shuttle data between physically separated memory and processors, a process that consumes large amounts of energy. Hersam’s group instead integrates memory and computation into a single device called a memory transistor.
In a 2023 study published in nature electronicsthe team demonstrated that AI classification tasks that traditionally required more than 100 transistors can be performed with just two memory transistors. This approach reduced energy consumption by approximately 100 times.
New research pushes the concept beyond low-energy classification. Rather than simply making their AI hardware more efficient, the Northwestern University team redesigned the device to mimic specific circuits in the cerebellum that are better at detecting novelty and making split-second decisions.
This approach allows AI to immediately flag unexpected events while ignoring routine information. For wearable heart monitors, that might mean detecting the first signs of irregular heartbeats. For robots, that might mean recognizing when a person suddenly steps in front of them. And for cybersecurity systems, it can mean spotting suspicious network activity before it escalates into a full-blown attack.
“Today’s AI is very good at recognizing patterns, but it often spends huge amounts of computing power continuously analyzing data streams, even when nothing is changing,” says Hersam. “Therefore, you end up wasting energy on unnecessary analysis.”
Reproduction of excitatory and inhibitory responses
In the cerebellum, neural circuits contain two competing signals, one excitatory and one inhibitory, that constantly balance each other. During normal activity, the signal remains in equilibrium. But when something surprising happens, that balance shifts temporarily and your brain alerts you to the need to react.
Northwestern’s team recreated this same dynamic in hardware. Engineers developed this device to serve two different roles. In one mode, the response gradually strengthens as the signal continues, acting like an excitatory synapse. In the other mode, they act like inhibitory synapses, responding strongly at first and then fading quickly. These complementary actions allow the device, like the cerebellum, to distinguish between normal and truly novel events.
To build the device, the researchers used molybdenum disulfide, an atomically thin semiconductor known for its electrical properties. Next, they designed an asymmetric transistor architecture in which one electrode partially overlaps the semiconductor through a thin insulating layer. A seemingly small design change fundamentally changed the way electricity flows within the device. Simply reversing the direction of the applied voltage switches the memory transistor between excitatory and inhibitory modes.
Test your device
To test the system, the researchers fed the device a series of electrocardiogram (ECG) recordings that included both normal heart rhythms and arrhythmias. Instead of wasting energy by completely analyzing each heartbeat, the device managed to ignore normal heartbeats. But then they quickly identified the abnormal heartbeat within just a few milliseconds.
“Our cerebellum-inspired memory transistor detected irregular heartbeats within seconds, before the heartbeat had finished,” Hartham said. “This is more than twice as fast as traditional AI.”
Next, Hersam plans to explore ways to mimic the cerebellum’s ability to learn and adapt over time. For example, if a once unexpected event occurs repeatedly, the brain gradually learns and stops treating the repeated event as unusual.
“We have demonstrated some parts of the cerebellar neural circuitry, but there are still many parts that we have yet to mimic,” Hersam said. “We intend to continue down this path to further mimic this complex system.”
