To realize a convenient and secure information society in which anyone can use information and communication devices, such as smartphones, anywhere, small devices, low power consumption, and efficient information processing capabilities are required. The central processing unit (CPU) of a device now integrates tens of billions of transistors, which are switching elements that control current flow through the device in response to external voltages and currents. The miniaturization and energy saving of transistors using conventional technology are reaching their limits; therefore, information processing technology based on a different principle is required.
The National Institute of Advanced Industrial Science and Technology (AIST) has been developing technologies for manufacturing oxide materials and components, aiming to utilize various physical phenomena that occur in materials as elements. In this research, a group including Researcher Hisashi Inoue and Chief Senior Researcher Isao Inoue from the Correlated Electronics Group, Electronics and Photonics Research Institute, AIST, employed these technologies to develop an artificial device that mimics the behavior of "biological neural tissues," processing slow input signals with high efficiency. Further, they succeeded in developing a transistor that operates at an extremely small power of 500 pW and demonstrated its operation.
This transistor is suitable for building artificial neural networks. In a simulation using the developed element, the researchers successfully detected anomalies in triangles drawn by person A and person B. Meanwhile, simulation using an element 100,000 times faster than the developed transistor failed in the same task. This is because the slower the element operates, the more information it can retain over a longer period of time. In information processing involving human interaction, energy savings need to be achieved by actively using the "slow" nature of biological nervous tissues. In the future, the joint research group plans to build neural networks using the newly developed transistors, thereby establishing information processing infrastructure for edge devices that can operate on low power.
Simulation of handwriting anomaly detection using neural networks
(b) Result of a verification experiment on handwriting anomaly detection in a simulation using the developed element.
The neural network using the slow element, which mimics biological neural tissues, is more successful in detecting abnormalities in handwriting.