First Realization of Batch Normalization in Flash-Based Binary Neural Networks Using a Single Voltage Shifter

IF 2.1 4区 工程技术 Q3 ENGINEERING, ELECTRICAL & ELECTRONIC IEEE Transactions on Nanotechnology Pub Date : 2024-09-23 DOI:10.1109/TNANO.2024.3466128
Sungmin Hwang;Wangjoo Lee;Jeong Woo Park;Dongwoo Suh
{"title":"First Realization of Batch Normalization in Flash-Based Binary Neural Networks Using a Single Voltage Shifter","authors":"Sungmin Hwang;Wangjoo Lee;Jeong Woo Park;Dongwoo Suh","doi":"10.1109/TNANO.2024.3466128","DOIUrl":null,"url":null,"abstract":"Batch normalization (BN) is a technique used to enhance training speed and generalization performance by mitigating internal covariate shifts. However, implementing BN in hardware presents challenges due to the need for an additional complex circuit to normalize, scale and shift activations. We proposed a hardware binary neural network (BNN) system capable of BN in hardware, which is consist of an AND-type flash memory array as a synapse and a voltage sense amplifier (VSA) as a neuron. In this system, hardware BN was implemented using a voltage shifter by adjusting the threshold of the binary neuron. To validate the effectiveness of the proposed hardware-based BNN system, we fabricated a charge trap flash with a gate stack of SiO\n<sub>2</sub>\n/Si\n<sub>3</sub>\nN\n<sub>4</sub>\n/SiO\n<sub>2</sub>\n. The electrical characteristics were modelled by using BSIM3 model parameters so that the proposed circuit was successfully demonstrated by a SPICE simulation. Moreover, variation effects of the voltage shifter were also analyzed using Monte Carlo simulation. Finally, the performance of the proposed system was proved by incorporating the SPICE results into a high-level simulation of binary \n<italic>LeNet-5</i>\n for MNIST pattern recognition, resulting in the improvement of the proposed system in terms of power and area, compared to the previous studies.","PeriodicalId":449,"journal":{"name":"IEEE Transactions on Nanotechnology","volume":"23 ","pages":"677-683"},"PeriodicalIF":2.1000,"publicationDate":"2024-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Transactions on Nanotechnology","FirstCategoryId":"5","ListUrlMain":"https://ieeexplore.ieee.org/document/10688406/","RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"ENGINEERING, ELECTRICAL & ELECTRONIC","Score":null,"Total":0}
引用次数: 0

Abstract

Batch normalization (BN) is a technique used to enhance training speed and generalization performance by mitigating internal covariate shifts. However, implementing BN in hardware presents challenges due to the need for an additional complex circuit to normalize, scale and shift activations. We proposed a hardware binary neural network (BNN) system capable of BN in hardware, which is consist of an AND-type flash memory array as a synapse and a voltage sense amplifier (VSA) as a neuron. In this system, hardware BN was implemented using a voltage shifter by adjusting the threshold of the binary neuron. To validate the effectiveness of the proposed hardware-based BNN system, we fabricated a charge trap flash with a gate stack of SiO 2 /Si 3 N 4 /SiO 2 . The electrical characteristics were modelled by using BSIM3 model parameters so that the proposed circuit was successfully demonstrated by a SPICE simulation. Moreover, variation effects of the voltage shifter were also analyzed using Monte Carlo simulation. Finally, the performance of the proposed system was proved by incorporating the SPICE results into a high-level simulation of binary LeNet-5 for MNIST pattern recognition, resulting in the improvement of the proposed system in terms of power and area, compared to the previous studies.
查看原文
分享 分享
微信好友 朋友圈 QQ好友 复制链接
本刊更多论文
使用单电压变换器首次实现基于闪存的二进制神经网络批量归一化
批量归一化(BN)是一种通过减轻内部协变量偏移来提高训练速度和泛化性能的技术。然而,由于需要额外的复杂电路来对激活进行归一化、缩放和移位,在硬件中实现 BN 存在挑战。我们提出了一种能够在硬件中实现 BNN 的硬件二元神经网络(BNN)系统,该系统由一个作为突触的 AND 型闪存阵列和一个作为神经元的电压感应放大器(VSA)组成。在该系统中,通过调整二进制神经元的阈值,使用电压移位器实现了硬件 BN。为了验证所提出的基于硬件的 BNN 系统的有效性,我们制作了一个电荷阱闪存,栅极堆叠为 SiO2/Si3N4/SiO2。我们使用 BSIM3 模型参数对其电气特性进行了建模,并通过 SPICE 仿真成功演示了所提出的电路。此外,还利用蒙特卡罗仿真分析了电压变换器的变化效应。最后,通过将 SPICE 结果纳入用于 MNIST 模式识别的二进制 LeNet-5 高级仿真,证明了所提系统的性能,与之前的研究相比,所提系统在功耗和面积方面都有所改进。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 去求助
来源期刊
IEEE Transactions on Nanotechnology
IEEE Transactions on Nanotechnology 工程技术-材料科学:综合
CiteScore
4.80
自引率
8.30%
发文量
74
审稿时长
8.3 months
期刊介绍: The IEEE Transactions on Nanotechnology is devoted to the publication of manuscripts of archival value in the general area of nanotechnology, which is rapidly emerging as one of the fastest growing and most promising new technological developments for the next generation and beyond.
期刊最新文献
Dual Metal Split Gate-Based Emulated Synaptic Device With Redacted Plasticity Utilizing Nanogranular Al2O3 Based Ion Conducting Electrolyte High-Speed and Area-Efficient Serial IMPLY-Based Approximate Subtractor and Comparator for Image Processing and Neural Networks Design of a Graphene Based Terahertz Perfect Metamaterial Absorber With Multiple Sensing Performance Modeling and Simulation of Correlated Cycle-to- Cycle Variability in the Current-Voltage Hysteresis Loops of RRAM Devices Impact of Electron and Hole Trap Profiles in BE-TOX on Retention Characteristics of 3D NAND Flash Memory
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
现在去查看 取消
×
提示
确定
0
微信
客服QQ
Book学术公众号 扫码关注我们
反馈
×
意见反馈
请填写您的意见或建议
请填写您的手机或邮箱
已复制链接
已复制链接
快去分享给好友吧!
我知道了
×
扫码分享
扫码分享
Book学术官方微信
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术
文献互助 智能选刊 最新文献 互助须知 联系我们:info@booksci.cn
Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。
Copyright © 2023 Book学术 All rights reserved.
ghs 京公网安备 11010802042870号 京ICP备2023020795号-1