Self-Learning Functions in Electron Devices and Their Applications to Neural Networks

电子器件中的自学习功能及其在神经网络中的应用

基本信息

批准号：
04402035
负责人：
ISHIWARA Hiroshi
金额：
$ 10.18万
依托单位：
Tokyo Institute of Technology
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (A)
财政年份：
1992
资助国家：
日本
起止时间：
1992 至 1994
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-04402035/
关键词：
Neural-Networks Ferroelectrics Si Self-Learning Adaptive-Learning MISFET 自己学習機能 MOSFET ニューロ素子

项目摘要

In this research project, in order to extract self-learning functions in electron devices we proposed a novel adaptive-learning neurodevice and a neuron circuit using ferroelectric materials. In the proposed neuron circuit, the adaptive-learning MISFETs are employed to realize synaptic operations and an UJT (uni-junction transistor) is used as a switching device to obtain the output signal.First, we studied partial switching characteristics of PbZrTiO_3(PZT) films, which is a well-known ferroelectric material, when short input pulses are applied. It was found that ferroelectric PZT has adaptive-learning functions and it is applicable to neurodevices. Next, to fabricate the adaptive-learning neurodevices, we formed PZT films on Si substrates. However, it was shown tha the interdiffusion between PZT and Si was extremely serious and that PZT films with good-crystalline quality could not obtained. To overcome the difficulty, we proposed here to use a SrTiO_3(STO) as a buffer layr. The STO buffer layr suppressed the interdiffusion and PZT films with good crystalline quality were successfully grown on Si substrates. Furthermore, we studied on a new ferroelectric material, BaMgF_4(BMF) which can be grown on Si or GaAs substrates. BMF films were prepared on Si substrates by molecular beam epitaxy(MBE) and a remanent polarization of 1muC/cm^2 was obtained.In addition, we experimentally and theoretically studied on UJTs, which are used in the neuron circuits. We fabricated various size of UJTs on SOI(silicon on insulator) substrates and demonstrated that the output frequency can be controlled by a CR time constant in the neuroncircuit. Furthermore, we theoretically calculated the electrical characteristics of UJTs and showed that the characteristics of UJTs were strongly affected by surface recombinations at the Si/SiO_2 interface.

在这个研究项目中，为了提取电子设备中的自学习功能，我们提出了一种新型的自适应学习神经设备和使用铁电材料的神经元电路。在所提出的神经元电路中，采用自适应学习MISFET来实现突触操作，并使用UJT（单结晶体管）作为开关器件来获得输出信号。首先，我们研究了PbZrTiO_3（PZT）的部分开关特性当施加短输入脉冲时，薄膜是一种众所周知的铁电材料。研究发现铁电PZT具有自适应学习功能，适用于神经设备。接下来，为了制造自适应学习神经设备，我们在硅基板上形成了 PZT 薄膜。但结果表明PZT与Si之间的相互扩散极其严重，无法获得结晶质量良好的PZT薄膜。为了克服这个困难，我们在这里建议使用 SrTiO_3(STO) 作为缓冲层。 STO缓冲层抑制了相互扩散，在Si衬底上成功生长了具有良好结晶质量的PZT薄膜。此外，我们研究了一种新型铁电材料BaMgF_4(BMF)，它可以在Si或GaAs衬底上生长。采用分子束外延（MBE）技术在Si衬底上制备了BMF薄膜，获得了1μC/cm^2的剩余极化强度。此外，我们对神经元电路中使用的UJT进行了实验和理论研究。我们在 SOI（绝缘体上硅）基板上制造了各种尺寸的 UJT，并证明输出频率可以通过神经电路中的 CR 时间常数来控制。此外，我们从理论上计算了UJT的电特性，结果表明UJT的特性受到Si/SiO_2界面处表面复合的强烈影响。