神经形态工程学
尖峰神经网络
计算机科学
人工神经网络
调制(音乐)
晶体管
电子工程
材料科学
人工智能
光电子学
电压
电气工程
物理
工程类
声学
作者
Min‐Kyu Kim,Ik‐Jyae Kim,Jang‐Sik Lee
摘要
Neuromorphic computing that mimics the biological brain has been demonstrated as a next-generation computing method due to its low power consumption and parallel data processing characteristics. To realize neuromorphic computing, diverse neural networks such as deep neural networks (DNNs) and spiking neural networks (SNNs) have been introduced. DNNs require artificial synapses that have analog conductance modulation characteristics, whereas SNNs require artificial synapses that have conductance modulation characteristics controlled by temporal relationships between signals, so the development of a multifunctional artificial synapse is required. In this work, we report a ferroelectric thin-film transistor (FeTFT) that uses zirconium-doped hafnia (HfZrOx) and indium zinc tin oxide (IZTO) for neuromorphic applications. With reliable conductance modulation characteristics, we suggest that the FeTFT with HfZrOx and IZTO can be used as an artificial synapse for both DNNs and SNNs. The linear and symmetric conductance modulation characteristics in FeTFTs result in high recognition accuracy (93.1%) of hand-written images, which is close to the accuracy (94.1%) of an ideal neural network. Also, we show that the FeTFTs can emulate diverse forms of spike-time-dependent plasticity, which is an important learning rule for SNNs. These results suggest that FeTFT is a promising candidate to realize neuromorphic computing hardware.
科研通智能强力驱动
Strongly Powered by AbleSci AI