TDK Corp. says it has developed a prototype artificial intelligence (AI) chip that mimics the human cerebellum, using analog circuits to process information in ways fundamentally different from conventional AI systems.
The Japanese electronics manufacturer announced Thursday it has created the chip in collaboration with Hokkaido University and plans to demonstrate it at CEATEC next week, where it will power a device that cannot be beaten at rock-paper-scissors.
The technology uses “reservoir computing,” an approach that differs significantly from the deep learning systems that undergird most modern AI applications. While deep learning relies on digital circuits performing trillions of mathematical operations across multiple neural network layers, reservoir computing leverages natural physical phenomena that evolve over time.
“The reservoir layer does not necessarily require calculations and uses natural phenomena that propagate over time,” TDK said in a news release.
TDK illustrated the concept of using water as an example. Input values create waves on a water surface, which propagate and interfere with each other in the reservoir layer, before an output layer reads these patterns to determine outcomes.
The demonstration device combines the analog reservoir chip with TDK’s acceleration sensors to predict an opponent’s hand gestures in rock-paper-scissors before they complete their move. By measuring finger movements through accelerometers, the system learns individual movement patterns in real time and selects the winning gesture while the opponent’s hand is still in motion.
“There are individual differences in finger movement, and to accurately determine what to do next it is necessary to learn those individual differences in real time,” TDK said.
Reservoir computing is not as versatile as general-purpose deep neural networks, but it offers potential advantages for processing time-series data, including lower latency and reduced power consumption.
TDK acknowledged that practical implementation has historically been challenging. Digital implementations fail to deliver the promised power efficiency benefits, while no physical reservoir computing devices had previously addressed both power consumption and high-speed operation requirements.
The company expects commercial applications to focus on processing sensor data at the edge of networks, closer to where data is generated.
The development follows similar research in Japan, including work earlier this year by the National Institute for Materials Science and Tokyo University of Science, which demonstrated waveform conversion using analog reservoir modeling.
