基于BP算法的片上学习CNN硬件加速器

为了适应便携式应用场合卷积神经网络(convolutional neural network, CNN)硬件加速器片上学习功能的需要,文章设计了一种多核并行运算的CNN硬件加速器,利用运算器内嵌缓存结构与运算过程分割和数据复用,减少运算器和存储器之间的数据交互,提高CNN运算的并行度,提升训练和推理过程的效率。该架构包含1组二维运算阵列和激活函数运算模块,以及相应的数据分配器和指令存储器;以1个16单元的CNN加速器设计为例,验证了所设计CNN加速器架构运行多种CNN模型时的性能和运算准确性。实验结果表明,文中提出的加速器架构与Intel9400F CPU相比,最大误差为8.043 7×10~(-6),识别精度下降0.63%,运行速度提高7.67倍。     

基于流水线架构的卷积神经网络FPGA实现

卷积神经网络（CNN）已被广泛用于图像处理领域,且通常在CPU和GPU平台上进行计算,然而在CNN推理阶段存在CPU计算速度慢和GPU功耗高的问题。鉴于现场可编程门阵列（field programmable gate array,FPGA）能够实现计算速度和功耗的平衡,针对当前在卷积结构设计、流水线设计、存储优化方面存在的问题,设计了基于FPGA的卷积神经网络并行加速结构。首先将图像数据和权值数据定点化为16 bit定点数,一定程度上减少了乘加运算的复杂性;然后根据卷积计算的并行特性,设计了一种高并行流水线卷积运算电路,提高了卷积运算性能,同时也对与片外存储进行数据交互的流水线存储结构进行了优化,以减少数据传输的时间消耗。实验结果表明,整体加速器在ImageNet数据集上的识别率达到94.6%,与近年来相关领域的报道结果相比,本文在计算性能方面有一定的优势。     

基于卷积神经网络的无人机识别方法仿真研究

为了提高卷积神经网络(CNN)的泛化性和鲁棒性,改善无人机航行时识别目标图像的精度,提出了一种CNN与概率神经网络(PNN)相结合的混合模型。利用CNN提取多层图像表示,使用PNN提取特征对图像进行分类以替代CNN内部的BP神经网络,采用均方差和降梯度法训练模型,通过将预处理的图像传输到CNN-PNN模型,对图像纹理和轮廓进行分类识别,并将此模型的仿真结果与卷积神经网络模型、卷积神经网络-支持向量机模型的结果进行对比。仿真结果表明,与其他两种模型相比,CNN-PNN模型具有更好的精准度,识别率高达96.30%。因此,CNN-PNN模型能够快速有效地识别图像,准确度和实时性较高,在图像识别等方面具有很好的应用前景。     

基于Dropout卷积神经网络的行为识别

近年来,卷积神经网络(CNN)已经成为很多科学领域的研究热点之一.卷积神经网络作为一种深度模型可以直接作用于原始输入,不需要手动设计特征描述子.与传统神经网络相比识别效果有很大的提高.它已经建立了一类强大的模型来处理图像识别,并对其扩展到三维卷积神经网络(3D CNN)来处理视频识别问题.在此基础上,笔者对三维卷积神经网络做了如下改进:用Gabor小波核来初始化卷积操作,以达到模拟人类视觉系统对视觉刺激的响应;在网络训练的过程中加入Dropout技术,随机选择删除部分神经元,以此来提高网络的泛化能力,有效防止过拟合.提出的方法在KTH和UCF-YouTube数据集上进行验证,取得了很好地识别效果.     

基于卷积神经网络的数字识别

为了解决银行、邮局等场合的实时数字识别问题,提出了一种优化的卷积神经网络(Convolutionnal Neural Network,CNN)数字识别方法。以Lenet-5模型为基础改进了卷积神经网络结构并推导了改进后的前向和反向传播算法,将改进的卷积神经网络在手写、印刷数字组合数据库上进行测试,分析了不同样本数量、训练迭代次数等参数对识别准确率的影响,并与传统算法进行比较分析。结果表明改进后的CNN结构简单,处理速度快,识别准确率高,具有良好的鲁棒性和泛化性,识别性能明显高于传统网络结构。     

基于FPGA的YOLOv2加速器设计

卷积神经网络(CNN:Convolutional Neural Network)计算量较大,为达到快速处理数据的目的,需借助硬件手段进行加速.因此,利用现场可编程门阵列(FPGA:Field Programmable Gate Array)并行计算的架构特性,提出了基于FPGA的并行计算加速策略.该策略采用的具体方法包括:合理分布片上内存与片下存储,降低数据读取延迟;采用多通道并行流水结构加速卷积操作;通过卷积层数据共享减少访存延迟.利用PYNQ-z2开发平台加速卷积神经网络YOLOv2,最终实现目标物体的检测识别,该设计的处理能力为27.03 GOP/s(Giga Operations Per Second,10亿次运算/s),与CPU(E5-2620V4)相比,处理能力是CPU的6.57倍,功耗是CPU的3％.     

基于CNN和随机弹性形变的相似手写汉字识别

针对手写汉字中相似汉字的识别问题,构建了一种卷积神经网络(CNN)模型,并给出了其网络拓扑结构,通过随机弹性形变对样本集进行扩展,以提高模型的泛化性能.相似手写汉字的识别实验结果表明:相对于常规的CNN模型,文中CNN模型的手写汉字识别正确率提高1.66%,特别是对于变形的手写汉字,识别正确率提高12.85%;相对于传统的手写汉字识别方法,文中方法的识别错误率降低36.47%,从而验证了文中识别方法的有效性.     

一种基于双流卷积神经网络跌倒识别方法

针对跌倒行为的视觉特征难以提取的问题,提出一种由两路卷积神经网络和模型融合部分组成的双流卷积神经网络(Two-Stream CNN)的跌倒识别方法.该方法的一路对视频帧的运动人加框标记后,送三维卷积神经网络(3D-CNN)处理来消除视频背景的干扰;另一路从相邻视频帧获取光流图后,送VGGNet-16卷积神经网络处理;最后将3D-CNN和VGGNet-16的Softmax输出识别概率加权融合作为Two-Stream CNN输出结果.实验结果表明:标记运动人并经3D-CNN处理有效地消除了视频背景的干扰;Two-Stream CNN跌倒识别率为96%,比3D-CNN提高了4%,比VGGNet-16网络提高了3%.     

卷积神经网络声学模型的结构优化和加速计算

将卷积神经网络(convolutional neural networks,CNN)声学模型应用于中文大词表连续电话语音识别任务中,分析了卷积层数、滤波器参数等变量对CNN模型性能的影响,最终在中文电话语音识别测试中,CNN模型相比传统的全连接神经网络模型取得了识别字错误率1.2％的下降.由于卷积结构的复杂性,常规的神经网络加速方法如定点量化和SSE指令加速等方法对卷积运算的加速效率较低.针对这种情况,对卷积结构进行了优化,提出了2种卷积矢量化方法:权值矩阵矢量化和输入矩阵矢量化对卷积运算进行改善.结果表明,输入矩阵矢量化方法的加速效率更高,结合激活函数后移的策略,使得卷积运算速度提升了8.9倍.     

基于改进的卷积神经网络图像识别方法

当前的图像特征识别大多采用的是传统的机器学习方法与卷积神经网络方法。传统的机器学习对图像识别的研究,特征提取多是通过人工完成,泛化能力不够强。最早的卷积神经网络也存在诸多缺陷,如硬件要求高,需要的训练样本量大,训练时间长。针对以上问题,提出了一种改进的神经网络模型,在LeNet-5模型的基础上并在保证识别率的情况下,简化网络结构,提高训练速度。将改进的网络结构在MINIST字符库上进行识别实验,分析网络结构在不同参量中的识别能力,并与传统算法进行对比分析。结果表明提出的改进结构在当前识别正确率上,明显高于传统的识别算法,为当前的图像识别提供新的参考。     

Characterization of Ordovician carbonate reservoirs, southeastern Saskatchewan, Canada

The discovery of the prolific Ordovician Red River reservoirs in 1995 in southeastern Saskatchewan was the catalyst for extensive exploration activity which resulted in the discovery of more than 15 new Red River pools. The best yields of Red River production to date have been from dolomite reservoirs. Understanding the processes of dolomitization is, therefore, crucial for the prediction of the connectivity, spatial distribution and heterogeneity of dolomite reservoirs.The Red River reservoirs in the Midale area consist of 3～4 thin dolomitized zones, with a total thickness of about 20 m, which occur at the top of the Yeoman Formation. Two types of replacement dolomite were recognized in the Red River reservoir: dolomitized burrow infills and dolomitized host matrix. The spatial distribution of dolomite suggests that burrowing organisms played an important role in facilitating the fluid flow in the backfilled sediments. This resulted in penecontemporaneous dolomitization of burrow infills by normal seawater. The dolomite in the host matrix is interpreted as having occurred at shallow burial by evaporitic seawater during precipitation of Lake Almar anhydrite that immediately overlies the Yeoman Formation. However, the low δ18O values of dolomited burrow infills (-5.9‰～ -7.8‰, PDB) and matrix dolomites (-6.6‰～ -8.1‰, avg. -7.4‰ PDB) compared to the estimated values for the late Ordovician marine dolomite could be attributed to modification and alteration of dolomite at higher temperatures during deeper burial, which could also be responsible for its 87Sr/86Sr ratios (0.7084～0.7088) that are higher than suggested for the late Ordovician seawaters (0.7078～0.7080). The trace amounts of saddle dolomite cement in the Red River carbonates are probably related to "cannibalization" of earlier replacement dolomite during the chemical compaction.     

A computer generator for randomly layered structures

AcomputergeneratorforrandomlylayeredstructuresYUJia shun1,2,HEZhen hua2(1.TheInstituteofGeologicalandNuclearSciences,NewZealand;2.StateKeyLaboratoryofOilandGasReservoirGeologyandExploitation,ChengduUniversityofTechnology,China)Abstract:Analgorithmisintrod…     

海南岛地体及其毗邻陆缘晚中生代—新生代古地磁研究和构造演化

本文叙述了对海南岛及其毗邻大陆边缘白垩纪到第四纪地层岩石进行古地磁研究的全部工作过程。通过分析岩石中剩余磁矢量的磁偏角及磁倾角的变化,提出海南岛白垩纪以来经历的构造演化模式如下:早期伴随顺时针旋转而向南迁移,后期伴随逆时针转动并向北运移。联系该地区及邻区的地质、地球物理资料,对海南岛上述的构造地体运动提出以下认识:北部湾内早期有一拉张作用,主要是该作用使湾内地壳显著伸长减薄,形成北部湾盆地。从而导致了海南岛的早期构造运动,而海南岛后期的构造运动则主要是受南海海底扩张的影响。海南地体运动规律的阐明对于了解北部湾油气盆地的形成演化有重要的理论和实际意义。     

Exciton Seebeck in Molecular Systems

Various applications relevant to the exciton dynamics,such as the organic solar cell,the large-area organic light-emitting diodes and the thermoelectricity,are operating under temperature gradient.The potential abnormal behavior of the exicton dynamics driven by the temperature difference may affect the efficiency and performance of the corresponding devices.In the above situations,the exciton dynamics under temperature difference is mixed with     

Achieving Unusual States of Matter Under High Pressure

The elongation method,originally proposed by Imamura was further developed for many years in our group.As a method towards O(N)with high efficiency and high accuracy for any dimensional systems.This treatment designed for one-dimensional(ID)polymers is now available for three-dimensional(3D)systems,but geometry optimization is now possible only for 1D-systems.As an approach toward post-Hartree-Fock,it was also extended to     

《Chinese Science Bulletin》SUBJECT INDEX TO VOLUME 59(2014)

正~~     

An on-line scaling method for improving scalability of a database cluster

The explosive growth of the Internet and database applications has driven database to be more scalable and available, and able to support on-line scaling without interrupting service. To support more client's queries without downtime and degrading the response time, more nodes have to be scaled up while the database is running. This paper presents the overview of scalable and available database that satisfies the above characteristics. And we propose a novel on-line scaling method. Our method improves the existing on-line scaling method for fast response time and higher throughputs. Our proposed method reduces unnecessary network use, i.e. , we decrease the number of data copy by reusing the backup data. Also, our on-line scaling operation can be processed parallel by selecting adequate nodes as new node. Our performance study shows that our method results in significant reduction in data copy time.     

An improved R-tree based on childnode''s probability

R-Tree is a good structure for spatial searching. But in this indexing structure,either the sequence of nodes in the same level or sequence of traveling these nodes when queries are made is random. Since the possibility that the object appears in different MBR which have the same parents node is different, if we make the subnode who has the most possibility be traveled first, the time cost will be decreased in most of the cases. In some case, the possibility of a point belong to a rectangle will shows direct proportion with the size of the rectangle. But this conclusion is based on an assumption that the objects are symmetrically distributing in the area and this assumption is not always coming into existence. Now we found a more direct parameter to scale the possibility and made a little change on the structure of R-tree, to increase the possibility of founding the satisfying answer in the front sub trees. We names this structure probability based arranged R-tree (PBAR-tree).     

An Index mechanism for multi-scale view in spatial databases

There are numerous geometric objects stored in the spatial databases. An importance function in a spatial database is that users can browse the geometric objects as a map efficiently. Thus the spatial database should display the geometric objects users concern about swiftly onto the display window. This process includes two operations:retrieve data from database and then draw them onto screen. Accordingly, to improve the efficiency, we should try to reduce time of both retrieving object and displaying them. The former can be achieved with the aid of spatial index such as R-tree, the latter require to simplify the objects. Simplification means that objects are shown with sufficient but not with unnecessary detail which depend on the scale of browse. So the major problem is how to retrieve data at different detail level efficiently. This paper introduces the implementation of a multi-scale index in the spatial database SISP (Spatial Information Shared Platform) which is generalized from R-tree. The difference between the generalization and the R-tree lies on two facets: One is that every node and geometric object in the generalization is assigned with a importance value which denote the importance of them, and every vertex in the objects are assigned with a importance value,too. The importance value can be use to decide which data should be retrieve from disk in a query. The other difference is that geometric objects in the generalization are divided into one or more sub-blocks, and vertexes are total ordered by their importance value. With the help of the generalized R-tree, one can easily retrieve data at different detail levels.Some experiments are performed on real-life data to evaluate the performance of solutions that separately use normal spatial index and multi-scale spatial index. The results show that the solution using multi-scale index in SISP is satisfying.     

Integrating GIS Web services based on mediating architecture

The geographic information service is enabled by the advancements in general Web service technology and the focused efforts of the OGC in defining XML-based Web GIS service. Based on these models, this paper addresses the issue of services chaining,the process of combining or pipelining results from several interoperable GIS Web Services to create a customized solution. This paper presents a mediated chaining architecture in which a specific service takes responsibility for performing the process that describes a service chain. We designed the Spatial Information Process Language (SIPL) for dynamic modeling and describing the service chain, also a prototype of the Spatial Information Process Execution Engine (SIPEE) is implemented for executing processes written in SIPL. Discussion of measures to improve the functionality and performance of such system will be included.