模型剪枝与低精度量化融合的DNN模型压缩算法

拥有庞大参数量的网络模型很难部署在智能手机、可穿戴智能设备等资源受限的移动设备上。从深度神经网络模型的基本原理出发，在现有压缩算法的基础上，采用优化剪枝策略与参数量化的方法相融合，提出了一种结果导向的数据驱动剪枝算法，利用低精度的量化算法来进一步压缩模型。使用VGGNet作为原始模型，在Kaggle猫狗图像和Oxford102植物样本集上进行微调。实验数据表明，使用本实验改进的方法，模型压缩的存储容量下降到113.1 MB，识别率提高到86.74%。     

基于多深度置信网络融合的航空电子设备故障预测

由于航空电子设备性能退化趋势与工作环境（温度、振动、负载等）存在强耦合关系,历史数据和实时数据分布存在难以量化的差异,因此航空电子设备的故障预测一直是业内难题。针对工程应用中的故障预测需求,提出了一种基于多深度置信网络（Multi-deep Belief Network,DBN）模型融合的故障预测方法,基于历史数据和实时数据对多个DBN模型进行含Dropout的迁移训练,有效解决历史域和目标域数据分布差异带来的预测偏差;采用改进遗传算法对DBN模型组进行融合,在迁移学习的基础上进一步提升故障预测的精准度。实验显示,所提方法预测的均方根误差为0.008,相对误差均值为0.9〖WT《Times New Roman》〗%〖WTBZ〗,相关度为0.964 7,预测精度高于单一DBN模型和支持向量机,在航空电子设备的故障预测领域有一定的应用价值。     

基于频域奇异值分解的LPI雷达信号降噪

随着低截获技术在战场上的大量使用,侦察接收机接收到的低截获（LPI）信号大多都湮没在噪声中。为了准确地检测威胁目标,在研究时域奇异值分解（SVD）降噪的基础上,提出了基于频域SVD的LPI雷达信号降噪方法。仿真实验表明,该方法可以将LPI雷达常用的线性调频连续波（LFMCW）信号和二相编码（BPSK）信号的信噪比从0 dB提升到6 dB以上。频域SVD降噪使侦察机发现LPI雷达的概率大大提高。     

新品推荐

简评丽台公司出品的这款Modem,在技术上比较全面,其中包括可作为数位式电话答录系统,可使用电话线做视讯。在硬体资料压缩后,SVD的传输率可达115.2Kbps。对于家庭用户,除一般的声音功能外,SVD能支持用户在收到邮件后自动呼叫使用者的呼叫器,当使用者拥有多个邮件信箱时,还可以设定不同的密码。     

基于Stacking集成学习的Lasso-GBDT组合房价预测模型研究

房地产行业是我国国民经济的重要组成部分，关乎国计民生。房价预测的准确性与稳定性，对政府、开发商和广大市民均具有重要的现实意义。本文利用网络爬虫方法获取银川市2015年4月至2021年3月安居客房产信息服务平台样本住宅价格数据，分别利用Lasso模型和梯度提升决策树（GBDT）模型对银川市房价进行预测；综合两种预测方法的优点，建构基于Stacking集成学习的Lasso-GBDT组合回归预测模型。通过实例预测结果比较，组合预测模型预测精度均在0.98以上，能有效避免病态数据对拟合程度的影响，较单项预测模型有更高的准确性和稳定性。     

使用变换域编码技术压缩SAR原始数据

首先分析了在合成孔径雷达（SAR）原始数据中通常使用的块自适应量化（BAQ）算法，然后在此基础上详细讨论了两种基于块自适应量化的变换域编码算法，即基于快速傅里叶变换块自适应量化（FFT-BAQ）和基于小波变换块自适应量化（WT-BAQ），并对这两种算法压缩得到数据解压缩获得图像与块自适应量化得到的图像进行分析比较，结果显示变换域编码技术能改善SAR原始数据压缩性能。     

金融市场的风险值测量及动态分析

介绍了基于高斯核估计历史模拟法的金融市场风险值（VAR）测量的算法理论过程，再建立时变系数的单因素资本资产定价模型（CAPM），并用带时变参数系统的Kalman滤波法对该模型下时变系数进行估计．最后对上海证券市场商业版块类做实证分析。     

面向SAR图像处理的生成式对抗网络应用综述

针对缺少合成孔径雷达（Synthetic Aperture Radar，SAR）目标图像数据导致的识别网络难以训练的问题，总结了现有的基于深度学习方法的解决方案。归纳了现阶段生成式对抗网络（Generative Adversarial Network，GAN）的发展情况，以及主要的衍生模型及其特点与优势。综述了GAN在SAR图像生成与风格迁移两方面的应用情况，并合理分析了应用中的技术难点和问题。最后结合深度学习的发展趋势，展望了GAN在SAR智能解译方面的应用。     

电视与网络视频广告价值融合评估——CMMG跨媒体评估指标体系评析之一

网络视频的兴起为广告从传统电视媒体向视频网站迁移提供了通道，同时，国内视频网站饱受运营成本高昂之苦，急需从传统电视媒体中分得更多的广告收益，因此，对于融合两种不同媒体形式的广告价值评估体系就显得尤为迫切。2010年4月，互联网数据中心（DCCI）会同产业界成立了“跨媒体任务小组”，     

迁移学习中的知识获取技术及在文本分类中的应用

迁移学习的主要目标是将一个已知环境中学习的知识，迁移到新环境或变化的环境中，以帮助新的学习。本文提出了一种新的迁移知识获取策略，井将其应用于自动文本分类中进行验证，以打破传统机器学习任务对于训练集和测试集要求同一概率分布和特征空间的假设。迁移知识将通过学习一些特征词的语言学信息及统计信息，构成一种统一的表示形式。相关实验在相同的特征空间及不同的特征空间进行，实验表明，迁移学习方法是有效的，在不同的测试环境下均取得了较理想的结果。     

Variable selection in international diffusion models

Prior research comes to different conclusions as to what country characteristics drive diffusion patterns. One prime difficulty that may partially explain this divergence between studies is the sparseness of the data, in terms of the periodicity as well as the number of products and countries, in combination with the large number of potentially influential country characteristics. In face of such sparse data, scholars have used nested models, bivariate models and factor models to explore the role of country covariates. This paper uses Bayesian Lasso and Bayesian Elastic Net variable selection procedures as powerful approaches to identify the most important drivers of differences in Bass diffusion parameters across countries. We find that socio-economic and demographic country covariates (most pronouncedly so, economic wealth and education) have the strongest effect on all diffusion metrics we study. Our findings are a call for marketing scientists to devote greater attention to country covariate selection in international diffusion models, as well as to variable selection in marketing models at large.     

基于子孔径方位特性的穿墙雷达虚拟多视角成像方法

针对穿墙雷达(TWR)室内不规则场景中信号传播的多径效应所导致成像区域“幻影”存在的问题，提出了一种基于脉冲雷达信号平台频域压缩感知(CS)框架的多径利用方法。在建立室内多径回波脉冲信号矢量模型的基础上，将每条多径分量作为一个特殊的观测通道，并利用目标、多径“幻影”依赖于阵列子孔径方位(AD)的特性，由此来分析目标、“幻影”与子孔径位置的关联性，最后通过调和平均(HMAM)子孔径图像融合方法实现 “幻影”抑制和目标的虚拟多视角成像。该方法克服了脉冲信号时域压缩感知计算复杂度高以及需要事先已知室内几何反射位置的不足。不同场景的仿真结果验证了此方法的可行性和优越性。     

TD-LTE系统中基于奇异值分解的高效波束赋形方法

为了克服TD-LTE系统中传统的基于矩阵信道的迭代波束赋形方法复杂度较大且可能 不收敛的问题,提出一种基于奇异值分解的高效波束赋形方法。该方法利用系统终端侧 仅有2根天线的系统特性,采用基于2×2矩阵特征值分解公式的矩阵信道奇异值分解获得波 束赋形天线加权向量;不仅可以直接获得精确的加权向量,而且相对于传统迭代方法大幅降 低了复杂度。理论分析和仿真结果验证了所提方法的有效性。     

短码、周期长码直扩信号伪码序列盲估计

针对短码、周期长码直扩信号在不同的时延下伪码序列估计问题，提出了一种基于奇异值分解的盲解扩算法。在已知信息码元速率和伪码周期条件的前提下，算法首先把接收到的直扩信号按照一定长度进行分段构成相关矩阵并对此矩阵进行奇异值分解得出信号子空间，然后根据信号子空间和伪码序列的模糊关系，利用求解的模糊酉矩阵和特定约束条件（如m序列）去其模糊性，最终估计出伪码序列。仿真结果表明，该算法不仅解决了在不同的时延下估计伪码序列带来的问题，而且具有稳定性高、在低信噪比条件下有良好的估计性能等优点。     

Labour market institutions,learning and self-employment

We develop a dynamic partial-equilibrium model to analyse how labour market institutions (wage compression, minimum wages, unemployment benefits, mobility costs and fixed-costs of self-employment) and learning affect who and when people become self-employed. We find that certain ability groups of workers become self-employed for both “carrot” and “stick” reasons: Some prefer self-employment to the low institutionalised wage, while others are not productive enough to qualify for a job at the institutionalised wage. Furthermore, wage compression and learning may give rise to a class of switchers who start in wage employment and later switch to self-employment. Several predictions of the model are consistent with observed empirical regularities, such as the existence of a group of low-skilled self-employed workers, the increasing propensity for self-employment over age groups and the larger spread in earnings among self-employed.     

A hybrid data mining model of feature selection algorithms and ensemble learning classifiers for credit scoring

Data mining techniques have numerous applications in credit scoring of customers in the banking field. One of the most popular data mining techniques is the classification method. Previous researches have demonstrated that using the feature selection (FS) algorithms and ensemble classifiers can improve the banks' performance in credit scoring problems. In this domain, the main issue is the simultaneous and the hybrid utilization of several FS and ensemble learning classification algorithms with respect to their parameters setting, in order to achieve a higher performance in the proposed model. As a result, the present paper has developed a hybrid data mining model of feature selection and ensemble learning classification algorithms on the basis of three stages. The first stage, as expected, deals with the data gathering and pre-processing. In the second stage, four FS algorithms are employed, including principal component analysis (PCA), genetic algorithm (GA), information gain ratio, and relief attribute evaluation function. In here, parameters setting of FS methods is based on the classification accuracy resulted from the implementation of the support vector machine (SVM) classification algorithm. After choosing the appropriate model for each selected feature, they are applied to the base and ensemble classification algorithms. In this stage, the best FS algorithm with its parameters setting is indicated for the modeling stage of the proposed model. In the third stage, the classification algorithms are employed for the dataset prepared from each FS algorithm. The results exhibited that in the second stage, PCA algorithm is the best FS algorithm. In the third stage, the classification results showed that the artificial neural network (ANN) adaptive boosting (AdaBoost) method has higher classification accuracy. Ultimately, the paper verified and proposed the hybrid model as an operative and strong model for performing credit scoring.     

基于字典学习的毫米波大规模MIMO系统混合预编码

针对毫米波大规模多输入多输出（MIMO）系统混合预编码方案设计的难题，采用字典学习的思想，提出了一种高精度混合预编码方案。该方案首先对全数字预编码矩阵的各列采用稀疏表示；进而按列将字典原子从稀疏表示中分离出来，通过对误差矩阵采用奇异值分解（SVD）来更新对应的字典原子，直到所有字典原子更新，以形成新的字典矩阵;最后，利用更新后的字典矩阵稀疏重构全数字预编码矩阵，以得到模拟预编码矩阵和数字预编码矩阵。仿真结果表明，相较于基于正交匹配追踪（OMP）的混合预编码方案，所提方案在提升系统频谱效率和降低误码率方面具有明显的优势。     

Myopic Selection

The severity of selection mechanisms and the myopia of selection are explored through a duopoly model where one firm tries to move down a learning curve in which costs are initially higher than its rival's but ultimately much lower. A trade‐off is found between catch‐up time and asymptotic market share: the more severe are selection pressures, the less likely is it that the learning technology will survive; however, if it does survive, the learning technology will in the limit be more competitive the more severe are selection pressures. We explore the dynamics of the model under unit cost and strategic pricing and find that the optimal pricing rule depends on the parameters governing firm learning and market selection.     

外弹道测量多测速元数据融合同步修正方法

采用数据融合算法完成高精度外弹道测量对各测元时间同步有着严格的技术要求。在分析时间不同步对融合弹道测速精度的影响基础上,分析了引起无线电测速时间不同步的因素,并构建了有效的数学修正模型。针对融合弹道速度异常超差现象,依据对测速误差理论模型公式和无线电测速原理分析结果,提出基于各测量设备测速数据解算模型的积分点移位和传播延迟修正模型算法,并应用于潜射弹的数据融合同步修正。工程应用结果表明,融合弹道偏差优于0.04 m/s,较修正前精度提高了4倍以上。由于陆基无线电测量系统的测距、测角具有类似的时间不同步特征,因此该模型也可作为多测元融合弹道解算过程中的一般方法推广使用。     

Executive migration and international mergers and acquisitions

This study applies the microfoundations approach to examine the impact of migrating executives on firms’ selection of host country in their international acquisition decisions. Viewing executive migration as a conduit for inter-organizational learning, this study offers the first empirical findings that a migrating executive’s inter-organizational learning associated with two specific types of international acquisition experience, i.e., general and country-specific, accumulated at a prior affiliated firm positively impacts host country selection in the executive’s current affiliated firm’s acquisition decisions. The findings further suggest that the migrating executive’s prior country-specific international acquisition experience has a greater influence compared with the migrating executive’s prior general international acquisition experience. Last, the findings show that industry similarity between a migrating executive’s prior and current affiliated firms enhances the effect of the migrating executive’s specific host country acquisition experience. In sum, this study contributes by introducing migrating executives as a new learning conduit in international acquisitions.