IC neuro chip with analog synapses for a correct reaction neural network

具有模拟突触的 IC 神经芯片，可实现神经网络的正确反应

• 批准号: 05555105
• 负责人: SAWADA Yasuji
• 金额: $ 12.61万
• 依托单位: Tohoku University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Developmental Scientific Research (B)
• 财政年份: 1993
• 资助国家: 日本
• 起止时间: 1993 至 1994
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-05555105/
• 关键词: Neuro Chip; Analog Memory; Boltzmann Machine; Integrated Circuit; Floating Gate; Learning Function; Silicon IC; Hebbian Learning; ボルツマンマシン

We have fabricated a microchip of a neural circuit with pulse representation. The neuron output is a voltage pulse train. The synapse is a constant current source whose output is proportional to the duty ratio of neuron output. Membrane potential is charged by collection of synaptic currents through a RC circuit, providing an analog operation similar to the biological neural systems. We use a 4-bit SRAM as the memory for synaptic weights.Then, we have fabricated a new type of analog memory (Switched Diffusion Analog Memory) with a floating gate witch is used as a key component to store synaptic weights for integrated artificial neural networks. The analog memory comprises a tunnel junction, a connection TFT,two capacitors, and a MOSFET.A small capacitance realized by use of poly-Si oxide film as the tunneling barrier together with the switching of the TFT improves linearity. The experimental result agrees with the theoretical estimation.The last, we designed and fabricated a new analog DBM (Deterministic Boltzmann Machine learning) chip using SDAM.DBM learning circuit designed by Hebbian rule is suitable for integrated analog neuro-chip by use of SDAM which can be operated simultaneously for learning mode and retrieving mode.

我们制作了一个具有脉冲表示的神经电路的微芯片。神经元输出是电压脉冲序列。突触是一个恒流源，其输出与神经元输出的占空比成比例。通过RC电路收集突触电流来对膜电位充电，提供类似于生物神经系统的模拟操作。我们采用4位SRAM作为突触权值的存储器，并研制了一种新型的浮栅模拟存储器（Switched Diffusion Analog Memory），它是集成神经网络中存储突触权值的关键器件。该模拟存储器包括隧道结、连接TFT、两个电容器和MOSFET，通过使用多晶硅氧化物膜作为隧道势垒实现的小电容与TFT的开关一起改善了线性度。最后，我们设计并制作了一种基于SDAM的模拟DBM（DeterministicBoltzmannMachinelearning）芯片，采用Hebbian规则设计的DBM学习电路适用于基于SDAM的集成模拟神经元芯片，可以同时工作于学习模式和检索模式。

项目成果

Y. Mizugaki: "Linearization analysis of threshold characteristics for some applications of mutually coupled SQUIDs" IECE Trans. Electron.E76-C. 1291-1297 (1993)

Y. Mizugaki：“互耦合 SQUID 某些应用的阈值特性的线性化分析”IECE Trans。

Y.Mizugaki: "New Approach Implementation of Neural Circuits Using Superconductive Devices" Extended Abstracts of the 1994 Int.Conf.on SSDM. 364-366 (1994)

Y.Mizugaki：“使用超导器件实现神经电路的新方法”1994 年 Int.Conf.on SSDM 的扩展摘要。

Y.Mizugaki: "Implementation of superconducting synapse into a neuro-based analog-to-digital converter" Appl.Phys.Lett.65. 1712-1713 (1994)

Y.Mizugaki：“将超导突触实现到基于神经的模数转换器”Appl.Phys.Lett.65。

Y. Mizugaki: "Superconducting implementation of neural networks using fluxon pulses" IEEE Trans. on Applied Superconductivity. 3. 2765-2768 (1993)

Y. Mizugaki：“使用磁通量脉冲超导实现神经网络”IEEE Trans。

K.Nakajima: "Hardware Implementation of New Analog Memory for Neural Networks" IEICE Trans.ELECTRON.E78-C. 101-105 (1995)

K.Nakajima：“神经网络新型模拟存储器的硬件​​实现”IEICE Trans.ELECTRON.E78-C。