Inoue Hisashi, Tamura Hiroto, Kitoh Ai, Chen Xiangyu, Byambadorj Zolboo, Yajima Takeaki, Hotta Yasushi, Iizuka Tetsuya, Tanaka Gouhei, Inoue Isao H
National Institute of Advanced Industrial Science and Technology (AIST), Tsukuba, 305-8565, Japan.
Graduate Schools for Law and Politics, The University of Tokyo, Tokyo, 113-0033, Japan.
Adv Mater. 2025 Jan;37(3):e2407326. doi: 10.1002/adma.202407326. Epub 2024 Nov 27.
Recent advances in neural network-based computing have enabled human-like information processing in areas such as image classification and voice recognition. However, many neural networks run on conventional computers that operate at GHz clock frequency and consume considerable power compared to biological neural networks, such as human brains, which work with a much slower spiking rate. Although many electronic devices aiming to emulate the energy efficiency of biological neural networks have been explored, achieving long timescales while maintaining scalability remains an important challenge. In this study, a field-effect transistor based on the oxide semiconductor strontium titanate (SrTiO) achieves leaky integration on a long timescale by leveraging the drift-diffusion of oxygen vacancies in this material. Experimental analysis and finite-element model simulations reveal the mechanism behind the leaky integration of the SrTiO transistor. With a timescale in the order of one second, which is close to that of biological neuron activity, this transistor is a promising component for biomimicking neuromorphic computing.
基于神经网络的计算技术的最新进展，已在图像分类和语音识别等领域实现了类人信息处理。然而，许多神经网络运行在传统计算机上，这些计算机以吉赫兹时钟频率运行，与生物神经网络（如人类大脑，其尖峰发放速率要慢得多）相比，功耗相当大。尽管已经探索了许多旨在模拟生物神经网络能量效率的电子设备，但在保持可扩展性的同时实现长时间尺度仍然是一个重大挑战。在本研究中，基于氧化物半导体钛酸锶（SrTiO）的场效应晶体管通过利用该材料中氧空位的漂移扩散，在长时间尺度上实现了泄漏积分。实验分析和有限元模型模拟揭示了SrTiO晶体管泄漏积分背后的机制。该晶体管的时间尺度约为一秒，与生物神经元活动的时间尺度相近，是用于仿生神经形态计算的一种很有前景的组件。
