权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

A Self-Learning Neural Network LSI Employing Random Noise Generator

采用随机噪声发生器的自学习神经网络LSI

基本信息

批准号：
07505027
负责人：
SHIBATA Tadshi
金额：
$ 6.08万
依托单位：
TOHOKU UNIVERSITY
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (A)
财政年份：
1995
资助国家：
日本
起止时间：
1995 至 1996
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-07505027/
关键词：
Thermal Noise Source Follower Anti-Miller Effect CMOS Random Noise Silicon Retina focal plane processor CMOSアナログ回路カスコードアンプ確率動作ニューラルネットワーク kTCノイズ

项目摘要

Generation of spatially and temporary random noise is essential in implementing intelligent information processing systems like neural networks as integrated circuits hardware. However, this has been a very difficult task. The voltage fluctuation appearing across the two terminals of a resistor due to the random motion of electrons is known as thermal noise, which is known as white noise. If such thermal noise is amplified to an appropriate level for circuit operation, it provides a very good noise source. However, the output is band limited due to the RC circuitry composed of the noise source resistor and the input capacitance of the amplifier, resulting in the available noise power of kT/C.This is only the order of mV for typical amplifiers. We resolved the problem by utilizing the "Anti-Miller Effect" of a source follower which we discovered. In a source follower the input capacitance becomes almost zero due to its almost unity gain. Further more CMOS inverter was utilized as a noise source resistor, yielding better whiteness of the noise. As a result, we developed a random-noise generator composed of only eight transistors and the performance has been experimentally verified. For neural processing, we have developed a retina processing chip which performs noise filtering, edge enhancement by two-dimensional Laplacian processing, and edge detection. The basic operation of the test circuit has been also verified experimentally. By injecting random noise to the pixel processing, we have found that a missing feature can be restored by stochastic resonance. In this manner, the important role of random noise generation in intelligent data processing has been demonstrated.

空间和临时随机噪声的产生是实现智能信息处理系统如神经网络作为集成电路硬件的必要条件。然而，这是一项非常艰巨的任务。由于电子的随机运动而在电阻器两端出现的电压波动称为热噪声，也称为白噪声。如果这种热噪声被放大到电路工作的适当水平，它提供了一个非常好的噪声源。然而，由于由噪声源电阻和放大器输入电容组成的RC电路，导致输出受限，导致可用噪声功率为kT/C。对于典型的放大器来说，这只是mV的数量级。我们利用我们发现的源追随者的“反米勒效应”解决了这个问题。在源从动器中，由于其几乎为单位增益，输入电容几乎为零。进一步利用CMOS逆变器作为噪声源电阻，使噪声的白度更好。因此，我们开发了一个仅由8个晶体管组成的随机噪声发生器，并通过实验验证了其性能。在神经处理方面，我们开发了一种视网膜处理芯片，该芯片可以进行噪声滤波、二维拉普拉斯处理的边缘增强和边缘检测。实验验证了测试电路的基本工作原理。通过在像素处理中注入随机噪声，我们发现可以通过随机共振来恢复缺失的特征。通过这种方式，证明了随机噪声产生在智能数据处理中的重要作用。