Tohoku University develops analog circuit that mimics brain processes with ultra-low power consumption

Researchers from the Research Institute of Electrical Communication at Tohoku University, including Specially Appointed Assistant Professor Satoshi Moriya and Professor Shigeo Sato, have successfully implemented a spiking neural network (SNN) that operates like the human brain using analog CMOS circuitry. By utilizing CMOS operating in the subthreshold region, they achieved SNN operation with ultra-low power consumption—approximately one-millionth that of a smartphone. They also demonstrated its application to speech recognition tasks and showed that information processing is possible despite manufacturing variations and temperature changes. Their findings were published in IEEE Transactions on Circuits and Systems I: Regular Papers.

SNNs, conceived from biological neural networks, represent and process information based on spike signals output by neurons. Because of their event-driven nature—where information processing occurs only when spikes are generated—power consumption can be reduced to an absolute minimum. To fully leverage this characteristic, specialized hardware that efficiently implements SNN operation is necessary.

When creating specialized hardware using analog circuits, operating transistors in the subthreshold region can reduce power consumption to less than 1/100 compared to conventional digital circuit operation. However, since analog circuits are significantly affected by transistor manufacturing variations and temperature changes, it was unknown whether information processing that tolerates such variations could be realized with analog circuits operating in the subthreshold region.

The research group constructed an SNN using analog CMOS circuits and successfully produced an ultra-low power SNN chip. The energy consumption per spike in the neuron circuit was 22.7 femtojoules, achieving power consumption 2-3 orders of magnitude lower than conventional SNN implementations using digital circuits.

The researchers built a system capable of connecting up to six of these chips and applied it to speech signal classification tasks using the reservoir computing framework. They achieved over 80% accuracy in classifying spoken digits from "Zero" to "Nine" across ten classes. They also demonstrated that accuracy improves by mixing neurons with different operation modes within the network, similar to biological systems.

Furthermore, they conducted simulations of speech signal classification tasks considering manufacturing variations and temperature changes, which are problematic for analog circuit applications. The results showed that performance is maintained against device size variations and temperature changes when relearning is performed under conditions that account for variations. This suggests high technical compatibility between the proposed system and reservoir computing.

These findings are expected to be applied to ultra-low power information processing devices that can operate in environments where batteries are unnecessary, or battery replacement is limited. Additionally, the research demonstrated the possibility of a new computing approach that tolerates and utilizes variations that significantly affect analog circuit operation.

Journal Information
Publication: IEEE Transactions on Circuits and Systems I: Regular Papers
Title: Analog VLSI Implementation of Subthreshold Spiking Neural Networks and Its Application to Reservoir Computing
DOI: 10.1109/TCSI.2025.3550876