Rock-paper-scissors is typically a game of psychology, strategy, and chance. However, researchers from Hokkaido University and TDK Corporation in Japan have developed a chip capable of winning every time by predicting your moves.
The chip doesn’t read your mind. Instead, it uses a wrist-mounted sensor to track your motions, learning what gestures correspond to rock, paper, or scissors. Once trained on your specific movements, it can calculate your next action in the brief moment it takes for you to say, “shoot,” enabling it to counter your play in real time.
This innovation utilizes reservoir computing, a machine-learning approach that leverages complex dynamical systems to decode meaningful features from time-series data. Interest in reservoir computing has surged with advancements in artificial intelligence, thanks to its low power consumption and rapid training capabilities.
The research team focused on minimizing energy use and latency, as noted by Tomoyuki Sasaki, a senior manager at TDK. They developed a CMOS hardware implementation of an analog reservoir computing circuit and showcased their work at the Combined Exhibition of Advanced Technologies conference in Chiba, Japan, and are presenting at the International Conference on Rebooting Computing in San Diego, California.
Reservoir computing differs from traditional neural networks, which consist of layers of artificial neurons connected by adjustable synapses. In contrast, reservoir computing features neurons organized in a web-like structure, allowing for memory and looping within the network. Only the connections leading to the output are adjusted during training, simplifying the process and removing the need for backpropagation.
Despite sounding simplistic, these networks can be effective for specific tasks. They excel in predicting time evolution in chaotic systems, such as weather patterns, due to their operation at the “edge of chaos,” which enables efficient representation of numerous states with a small neural network.
To build a physical reservoir computer, researchers have employed diverse mediums, but the Hokkaido and TDK team aimed for a CMOS-compatible chip for use in edge devices. They created an analog circuit node comprising a non-linear resistor, a memory element, and a buffer amplifier. Their chip features four cores, each containing 121 nodes, connected in a simplified cycle to facilitate reservoir dynamics.
The resulting chip consumes just 20 microwatts of power per core, totaling 80 microwatts—significantly lower than other CMOS-compatible designs.
In addition to its rock-paper-scissors capabilities, the reservoir computing chip can forecast future events in various domains. If present outcomes are influenced by prior data, it can make accurate predictions. The team successfully demonstrated this on a logistic map and real-world weather scenarios, achieving remarkable accuracy in its forecasts.
While predictive precision is important, the chip’s low power requirements and quick responses could pave the way for new applications in real-time learning for wearables and other edge devices. Tomoyuki Sasaki stated that, although the prediction accuracy may match current technology, the power consumption and speed represent a significant improvement.
—
