We're bombarded with information at all times, so much that if we had to try and process it all times, we'd be paralyzed.
Our cerebellum does information damage control by filtering out what changes. Evolution made it possible for us to conserve energy while still processing anomalies quickly. AI needs a similar helping hand, especially as LLMs have begun generating their own content based on real content. Now, researchers have created a cerebellum-type proof-of-concept that identified abnormal heart rhythms with 98 percent accuracy and within afraction of a heartbeat - all while using 10,000 times fewer comoperations than existing approaches, which will also save energy.
The obvious goal is low-power, always-on, AI-optimized systems for improvements in wearable health monitors and even life-saving autonomous cars.
It could work because, unlike conventional computers, which migrate between memory and processors at relatively high cost, their molybdenum disulfide memristor can replace the tasks of 50 conventional transistors, which reduces equivalent energy consumption, by mimicking a specific circuit in the cerebellum, one that acts almost like reflex for new observations. That is why it might be valuable for heartbeats, or cars if someone steps into the road. Without the usual energy cost of AI analyzing every piece of data. One electrode partially overlapped the semiconductor using a dielectric that changed how electricity flows. Reversing the direction of the applied voltage switches the memtransistor between excitatory and inhibitory modes.
Molybdenum disulfide semiconductor with one electrode partially overlapping the semiconductor through a thin layer of insulation. Reversing the applied voltage direction changed the memtransistor between excitatory and inhibitory modes. Credit: Mark C. Hersam/Northwestern University
They say that causes their tool to improve speed by 100 percent.
Next, the team wants to try and mimic the cerebellum’s ability to learn and adapt over time. In our brains a surprise event that happens a few times is no longer novel so they off to discover more to optimize the cerebellum neural circuit.
Citation: Min-A Kang, Spencer T. Brown, Nethmi Jayasinghe, Meghana R. Holla, Thang T. Pham, Thomas T. Zeng, Ruiqin Wu, Zachary J. Trdinich, Xudong Zhuang, Vinayak P. Dravid, Indira M. Raman, Amit R. Trivedi, Vinod K. Sangwan & Mark C. Hersam. 'Cerebellum-inspired memtransistors enable emergent differentiation for hardware-efficient novelty detection'. Nat Commun (2026). DOI:10.1038/s41467-026-75212-4
