TDK and Hokkaido University are to demonstrate real-time learning in a ‘reservoir’ analogue AI chip that mimics the cerebellum at CEATEC in Japan.
TDK reservoir AI chip web
Reservoir computing is not deep-learning, where there is an input layer, multiple hidden layers and an output layer, usually implemented in digital logic that executes many, sometime trillions, of multiply and add operations.
“Reservoir computing consists of an input layer, reservoir layer and output layer,” said TDK. “The reservoir layer does not necessarily require calculations and uses natural phenomena that propagate over time.”
For example, it continued: “In the input layer, the natural phenomenon of multiple water surface waves is used as the input value. The next reservoir layer sends the results of propagation of surface waves and their mutual interference to the output layer. The last output layer properly reads the state of the reservoir layer and deduces the characteristics of how the waves on the surface of the water moved.”
As such, it is not a general purpose AI technique like deep neural networks, but when the task is time-varying time-series data processing, it has the potential to offer lower-latency as well as lower-power consumption.
However, “traditionally, it has been considered difficult to put reservoir computing devices into practical use”, said TDK. “In addition, it was difficult to obtain the benefits of low power when reservoir computing devices were implemented in digital computing, and there were no specific reservoir computing devices that used physical phenomena to consider power consumption and high-speed operations.”
The proof-of-concept device to be demonstrated at CEATEC will play a simple game using analogue circuits, sensors and electronic components, said TDK, and will “show that users can never win in rock-paper-scissors made possible by reservoir computing”.
It works from accelerometer measurements of its opponent’s fingers, determining the approaching hand gesture, and presenting the appropriate winning gesture, before those fingers have stopped moving.
“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,” said the company. “By demonstrating a single use-case, TDK hopes to foster a broader understanding of reservoir computing.”
Commercial applications are expected processing sensor data at the edge, and towards this the collaboration with Hokkaido University is continuing.
CEATEC – Combined Exhibition of Advanced Technologies – is at the Makuhari Messe near Tokyo over 14 to 17 October 2025.
Earlier this year, Japan’s National Institute for Materials Science and the Tokyo University of Science demonstrated triangle to sinewave conversion using analogue reservoir modelling.