基于软件定义的可重构卷积神经网络架构设计

为满足卷积神经网络业务处理的灵活性和高性能需求，提出一种基于软件定义的可重构卷积神经网络架构。该架构采用归一化处理流程实现卷积层网络的动态重构与运算模式的加速。采用AHB和AXI的双总线架构，实现卷积神经网络的流水计算。通过软件定义在FPGA上实现了不同网络结构下的数据集实时处理。实验结果表明，所设计的FPGA电路能够实现两种网络模型的软件定义，网络模型与输入数据集相同的条件下，该架构的运算处理能力为CPU的10倍，运算能耗比为GPU的2倍。     

面向卷积神经网络的FPGA加速器架构设计

随着人工智能的快速发展,卷积神经网络(CNN)在很多领域发挥着越来越重要的作用。分析研究了现有卷积神经网络模型,设计了一种基于现场可编程门阵列(FPGA)的卷积神经网络加速器。在卷积运算中四个维度方向实现了并行化计算;提出了参数化架构设计,在三种参数条件下,单个时钟周期分别能够完成512、1024、2048次乘累加;设计了片内双缓存结构,减少片外存储访问的同时实现了有效的数据复用;使用流水线实现了完整的神经网络单层运算过程,提升了运算效率。与CPU、GPU以及相关FPGA加速方案进行了对比实验,实验结果表明,所提出的设计的计算速度达到了560.2 GOP/s,为i7-6850K CPU的8.9倍。同时,其计算的性能功耗比达到了NVDIA GTX 1080Ti GPU的3.0倍,与相关研究相比,所设计的加速器在主流CNN网络的计算上实现了较高的性能功耗比,同时不乏通用性。     

基于FPGA的卷积神经网络加速器

现有软件实现方案难以满足卷积神经网络对运算性能与功耗的要求。为此,设计一种基于现场可编程门阵列(FPGA)的卷积神经网络加速器。在粗粒度并行层面对卷积运算单元进行并行化加速,并使用流水线实现完整单层运算过程,使单个时钟周期能够完成20次乘累加,从而提升运算效率。针对MNIST手写数字字符识别的实验结果表明,在75 MHz的工作频率下,该加速器可使FPGA峰值运算速度达到0.676 GMAC/s,相较通用CPU平台实现4倍加速,而功耗仅为其2.68%。     

基于硬件加速的轻量级网络心音分类器

近年来,卷积神经网络被广泛应用于心音信号分类。为满足先心病机器辅助诊断系统低功耗、可移动等方面需求,基于轻量级神经网络MobileNet,实现了一种适用于FPGA硬件平台的心音分类器。心音分类器的深度卷积、逐点卷积与最大池化等模块通过高层次综合进行设计。该心音分类器在利用深度可分离卷积减少网络参数与运算量的同时,通过多像素多通道并行及定点量化等方式,提升了分类器运行速度。经心音数据集实验结果表明,在计算效率方面,该心音分类器在FPGA上相较于在通用CPU上实现约14倍加速。     

基于FPGA的通用卷积神经网络识别系统研究

针对目前在中央处理器(CPU)中部署卷积神经网络速度慢、在图形处理器(GPU)中功耗高等问题,采用基于现场可编程门阵列(FPGA)平台开发的卷积神经网络识别系统,对卷积神经网络的各个环节进行算法加速。考虑到算法的计算量和逻辑资源的消耗主要集中在卷积层,提出了在特征图的通道方向进行双卷积并行模块设计。在卷积神经网络的池化层和激活函数Softmax中,设计了流式池化,并提出改进的分段查表计算Softmax函数的方法。另外,在归一化和预处理阶段也分别进行了优化。卷积神经网络识别系统选用XILINX公司的ZCU104开发平台。该平台内部包含片上系统与可编程逻辑控制器。通过自制水果数据集,分别在ZCU104、CPU和GPU上进行试验。试验结果显示,ZCU104分类的准确率达到了95.8%,识别速度约为计算机端上CPU实现同种网络模型的3倍,并且高于GPU。此外,该系统通用性高、资源占用率低,可应用在其他神经网络模型中。     

基于GPU加速的图像双向相似性计算

针对双向相似性计算在CPU下串行计算效率低下,无法满足实际需求的问题,利用该计算中数据独立性的特点,应用CUDA编程模型实现基于GPU加速的图像双向相似性计算。与CPU相比,在392x300的分辨率实验下,该算法在GPU上可获得超过1200倍的加速比。     

基于FPGA的卷积神经网络并行加速设计

为提升在资源、功耗受限的嵌入式平台上运行的深度卷积网络算法的速度和能效,提出一种基于现场可编程门阵列(FPGA)的卷积并行加速方案。利用卷积层与批归一化(batch normalization,BN)层融合减少计算复杂度;利用数据分片减少片上存储消耗;利用数据复用、并行计算提升运算速度,减少系统硬件开销;利用设计空间探索找到最符合硬件资源约束的计算并行度。实验结果表明,在100MHz的工作频率下,加速器的峰值计算性能可以达到52.56GFLOPS,性能是CPU的4.1倍,能耗仅为GPU的9.9%,与其它FPGA方案相比综合性能有一定的提升。     

深度学习中的卷积神经网络系统设计及硬件实现

针对目前深度学习中的卷积神经网络(CNN)在CPU平台下训练速度慢、耗时长的问题,采用现场可编程门阵列(FPGA)硬件平台设计并实现了一种深度卷积神经网络系统。该系统采用修正线性单元(Re LU)作为特征输出的激活函数并使用Softmax函数作为输出分类器。利用流水线技术并针对每一层的特征运算进行了并行处理,从而能够在1个系统时钟周期内完成整个CNN中的295次卷积运算。系统最后采用MNIST数据集作为实验样本,实验结果表明,在50 MHz的工作频率下,FPGA的训练用时相较于通用CPU的训练用时提升了8.7倍,经过2 000次迭代后系统识别的准确率为92.42%。     

基于GPU的卷积检测模型加速

近年来,形变部件模型和卷积神经网络等卷积检测模型在计算机视觉领域取得了极大的成功。这类模型能够进行大规模的机器学习训练,实现较高的鲁棒性和识别性能。然而训练和评估过程中卷积运算巨大的计算开销,也限制了其在诸多实际场景中进一步的应用。利用数学理论和并行技术对卷积检测模型进行算法和硬件的双重加速。在算法层面,通过将空间域中的卷积运算转换为频率域中的点乘运算来降低计算复杂度;而在硬件层面,利用GPU并行技术可以进一步减少计算时间。在PASCAL VOC数据集上的实验结果表明,相对于多核CPU,该算法能够实现在单个商用GPU上加速卷积过程2.13~4.31倍。     

基于ZYNQ的卷积神经网络加速器设计

针对卷积神经网络在嵌入式系统需要耗费大量计算资源、计算复杂度高等问题,提出一种基于ZYNQ系列FPGA的加速方法。通过HLS工具对卷积神经网络加速器进行设计,提出相邻层位宽合并和权重参数重排序的策略实现数据传输的优化,利用卷积分解、并行展开充分发挥FPGA并行计算的优势。为验证卷积神经网络加速器的加速效果,将YOLO目标检测模型进行部署。实验结果表明,在PYNQ-Z2上达到了39.39GOP/s的计算性能,是intel i5-2400 CPU的3.4倍,是ARM-Cortex A9 CPU的147.5倍。在相同FPGA平台上与之前的工作相较也有更高的性能。     

煤矿井下安全监控分站的设计及其仿真实现

为了设计一种具有低成本、低功耗、易操作、功能强且可靠性高的煤矿井下安全分站,针对煤矿安全生产实际,文章提出了采用MCS-51系列单片机为核心、具有CAN总线通信接口的煤矿井下安全监控分站的设计方案;首先给出煤矿井下安全监控分站的整体构架设计,然后着重阐述模拟量输入信号处理系统的设计过程,最后说明单片机最小系统及其键盘、显示、报警、通信等各个组成部分的设计;为验证设计方案的可行性与有效性,使用Proteus软件对设计内容进行仿真验证,设计的煤矿井下安全监控分站具有瓦斯、温度等模拟量参数超标报警功能和电机开停、风门开闭等开关量指示功能;仿真结果表明:设计的煤矿井下安全监控分站具有一定的实际应用价值.     

View-based model-driven architecture for enhancing maintainability of data access services

In modern service-oriented architectures, database access is done by a special type of services, the so-called data access services (DAS). Though, particularly in data-intensive applications, using and developing DAS are very common today, the link between the DAS and their implementation, e.g. a layer of data access objects (DAOs) encapsulating the database queries, still is not sufficiently elaborated, yet. As a result, as the number of DAS grows, finding the desired DAS for reuse and/or associated documentation can become an impossible task. In this paper we focus on bridging this gap between the DAS and their implementation by presenting a view-based, model-driven data access architecture (VMDA) managing models of the DAS, DAOs and database queries in a queryable manner. Our models support tailored views of different stakeholders and are scalable with all types of DAS implementations. In this paper we show that our view-based and model driven architecture approach can enhance software development productivity and maintainability by improving DAS documentation. Moreover, our VMDA opens a wide range of applications such as evaluating DAS usage for DAS performance optimization. Furthermore, we provide tool support and illustrate the applicability of our VMDA in a large-scale case study. Finally, we quantitatively prove that our approach performs with acceptable response times.     

A pragmatic multi-method investigation of discrepant technological events: Coping,attributions, and ‘accidental’ learning

Discrepant technological events or situations that entail a problem, a misunderstanding or a difficulty with the Information Technology (IT) being employed, are common in the workplace, and can lead to frustration and avoidance behaviors. Little is known, however, about how individuals cope with these events. This paper examines these events by using a multi-method pragmatic approach informed by coping theory. The results of two studies – a critical incident study and an experiment – serve to build and test, respectively, a theoretical model that posits that individuals use a variety of strategies when dealing with these events: they experience negative emotions, make external attributions, and adopt engagement coping strategies directed at solving the event, eventually switching to a disengagement coping strategy when they feel they have no control over the situation. Furthermore, users’ efforts may result in ‘accidental’ learning as they try to overcome the discrepant IT events through engagement coping. The paper ends with a discussion of the results in light of existing literature, future opportunities for research, and implications for practice.     

The Distributed Time-Triggered Simulation Scheme: Core Principles and Supporting Execution Engine

Distributed real-time simulation is a young technology field but its practice is under increasing demands. In recent years the author and his collaborators have been establishing a new approach called the distributed time-triggered simulation (DTS) scheme which is conceptually simple and easy to use but widely applicable. The concept was initiated in the course of developing a new-generation object-oriented real-time programming scheme called the time-triggered message-triggered object (TMO) programming scheme. Some fundamental issues inherent in distributed real-time simulation that were learned during recent experimental studies are discussed along with some approaches for resolving the issues. An execution engine developed to support both the TMOs engaged in control computation and the TMOs engaged in DTS is also discussed along with its possible extensions that will enable significantly larger-scale DTSs.     

Software-Konfigurationsmanagement in großen Softwareprojekten

Zusammenfassung Mit zunehmender Größe der Softwaresysteme verschärfen sich die für die Software-Herstellung typischen Probleme: Beherrschen großer Objektmengen, Erhalten der Systemkonsistenz, Kontrolle der ständigen Änderungseinflüsse und Gewährleisten einer langen Lebensdauer. Die Disziplin Konfigurationsmanagement bildet den methodischen Ansatz, diese Probleme besser zu beherrschen. Software-Konfigurationsmanagement faßt die Herstellung von Softwaresystemen als eine Abfolge von kontrollierten Änderungen an gesicherten Zwischen- und Endergebnissen auf. Dargestellt werden die Objekte und Funktionen des Software-Konfigurationsmanagements sowie die hierfür in großen Software-Projekten benötigten Methoden, Instanzen und Hilfsmittel.     

Information for Authors

正http://www.zju.edu.cn/jzus http://www.springerlink.com Aim The Journals of Zhejiang University-SCIENCE(A/B/C)are edited by the international board of distinguished Chinese and foreign scientists,and are aimed to present the latest developments and achievements in scientific research in China and     

Editorial:Brain-machine interface(BMI) and cyborg intelligence

正Brain-machine interfaces(BMIs)aim at building a direct communication pathway between the brain and an external device,and represent an area of research where significant progress has been made during the past decade.Based on BMIs,mind information can be read out by neural signals to control     

Information for Authors

正http://www.zju.edu.cn/jzus http://www.springerlink.com Aim The Journals of Zhejiang University-SCIENCE(A/B/C)are edited by the international board of distinguished Chinese and foreign scientists,and are aimed to present the latest developments and achievements in scientific research in China and overseas to the world’s scientific circles,especially to stimulate     

SCIENCE CHINA Information Sciences SUBJECT INDEX TO VOLUME 57(2014)

正~~     

Information for Authors

正http://www.zju.edu.cn/jzus http://www.springerlink.com Aim The Journals of Zhejiang University-SCIENCE(A/B/C)are edited by the international board of distinguished Chinese and foreign scientists,and are aimed to present the latest developments and achievements in scientific research in China and overseas to the world’s scientific circles,especially to stimulate