Updated
Updated · Livescience.com · Aug 11
Scientists Build AI Chip That Cuts Calculations 10,000x as Arrhythmias Are Flagged at 98% Accuracy
Updated
Updated · Livescience.com · Aug 11

Scientists Build AI Chip That Cuts Calculations 10,000x as Arrhythmias Are Flagged at 98% Accuracy

2 articles · Updated · Livescience.com · Aug 11

Summary

  • Simulated tests showed the cerebellum-inspired chip detected ECG arrhythmias within one-fifth of a heartbeat, hitting 98% accuracy and running more than twice as fast as a standard transformer model.
  • The gain comes from a memtransistor that combines memory and processing in one component, easing the von Neumann bottleneck and switching between excitatory and inhibitory modes to ignore routine signals.
  • Researchers modeled the design on the cerebellum rather than the cerebrum, aiming to reserve computing power for unexpected events instead of continuously analyzing every data stream.
  • The approach could support low-power edge AI in health monitors, robots, autonomous vehicles and cybersecurity systems, potentially reducing dependence on energy-hungry data centers whose electricity demand could reach 945 terawatt-hours by 2030.
  • The team reported the results July 10 in Nature Communications and said the next step is teaching the hardware to adapt as repeated events become predictable.

Insights

Could a microscopic AI chip that mimics the human cerebellum finally sever our smart devices' dangerous dependency on the cloud?
If a brain-inspired chip ignores routine data, could it accidentally delete the very clues needed to predict a future catastrophic failure?
When AI hardware decides what is normal without cloud guidance, who is responsible if a fatal anomaly is misclassified as routine?