辅助任务增强的中文跨域NL2SQL算法

自然语言到结构化查询语言(natural language to structured query language,NL2SQL)任务旨在将自然语言询问转化为数据库可执行的结构化查询语言(structured query language,SQL) 语句。本文提出了一种辅助任务增强的中文跨域 NL2SQL 算法,其核心思想是通过在解码阶段添加辅助任务以结合原始模型来进行多任务训练,提升模型的准确率。辅助任务的设计是通过将数据库模式建模成图,预测自然语言询问与数据库模式图中的节点的依赖关系,显式地建模自然语言询问和数据库模式之间的依赖关系。针对特定的自然语言询问,通过辅助任务的提升,模型能够更好地识别数据库模式中哪些表/列对预测目标 SQL 更有效。在中文 NL2SQL 数据集 DuSQL 上的实验结果表明,添加辅助任务后的算法相对于原始模型取得了更好的效果,能够更好地处理跨域 NL2SQL 任务。     

高性能异构加速器MiniGo算子优化方法

根据高性能异构加速器的特性和MiniGo的训练模式提出了一种高效的并行计算方法。对片上计算资源进行合理规划,实现异构设备之间的流水并行优化;根据异构设备间存在共享存储段设计了共享内存编码模式,减少数据传输开销;根据数字信号处理簇内具有多计算资源的特点结合算子计算-访存特性设计了不同的算子并行计算优化策略。同时,面向TensorFlow实现了一个易于使用的高性能计算库。实验结果显示,该方法实现了典型算子的多核并行计算。相对于单核,卷积算子加速比为24.69。相较于裁剪版8核FT2000+CPU,该方法训练和自博弈执行速度加速比分别为3.83和1.5。     

Defense Firms Adapting to Major Changes in the French R&D Funding System

The structural changes inside the French innovation system have impacted the role of defense firms since the late 1980s. Major changes have affected the defense budget and public R&D funding system in particular. The aim of this article is to understand French defense firms’ repositioning within the National Innovation System (NIS) based on an analysis of their R&D behavior over a long period of time (1987–2010). We show that French defense firms remain major players in the NIS and faced up to these major changes by adapting the funding of their R&D and their research priorities and rolling out new innovation capabilities. Additionally, they developed new innovation models to take advantage of new collaborative partnerships developed for civil and military markets.     

基于深度学习的MPSK信号调制识别

为了有效实现信号调制方式的智能识别,提出基于深度学习的多进制相移键控(Multiple Phase Shift Keying, MPSK)信号调制识别方法。分析接收MPSK信号的循环谱,并通过提取MPSK信号循环谱的等高图获得二维特征信息,利用深度学习中的卷积神经网络对二维特征进行训练,使用测试样本对所设计的调制识别方法的有效性进行验证。仿真结果表明,所提方法具有良好的识别性能。     

采用区分性幅相联合字典学习的低截获概率信号分离方法

为解决采用字典学习的信号分离方法存在的相位信息缺失和子字典交叉表示问题,提出一种区分性幅相联合字典学习方法。该方法针对相位信息缺失问题,构建了幅相联合字典模型;针对混合信号在联合字典上投影时存在的交叉表示问题,基于区分性字典学习思想提出在字典学习过程目标函数中加入交叉表示抑制项。仿真结果表明：幅相联合字典能够充分表示典型低截获概率信号的幅相信息,交叉表示抑制项能有效抑制信号间的交叉表示,算法具有良好的分离性能。     

利用Unity3D引擎实现某飞行器攻防对抗仿真

为了实现某飞行器攻防对抗的可视化仿真,利用可视化模型数据库构建技术,构建了全尺寸、高沉浸感、多视角的可视化仿真环境。利用Unity3D引擎的脚本技术,实现了导弹尾焰、大气云层、爆炸冲击波、视点跟随等特效与功能。该平台使用C＃和Unity3D引擎进行开发,实现了攻防双方参数设置、计算数据和图表输出、仿真画面输出等功能,并能在多次仿真后,对不同的参数进行统计分析。该平台具有较好的扩展性,能够逼真再现飞行器的攻防对抗过程。结果表明,对于需要不断修正数学模型以逼近真实的攻防对抗过程,每次修正都能立即通过可视化平台查看、验证模型修改对攻防对抗过程带来的影响,对飞行器的设计、作战使用具有一定的参考价值。     

可移植人体外耳支架的3D打印关键技术

探索使用医用硅胶制作患者个体外耳支架的关键技术,使用计算机断层扫描和逆向建模技术得到对应患者个体的外耳数字模型,利用3D打印技术制造人体外耳支架模具,并使用医用硅胶MED 4735完成人体外耳支架的制作成型。动物实验验证了其生物相容性,证明了方法的可行性。通过该技术制作的硅胶人体外耳支架可以高度模仿患者特定的耳廓形状,支架精度高、制造周期短、无生物排异,可为替代肋软骨雕刻人体外耳支架进行耳廓缺损治疗提供依据。     

Convolutions and generalization of logconcavity: Implications and applications

Additive convolution of unimodal and α‐unimodal random variables are known as an old classic problem which has attracted the attention of many authors in theory and applied fields. Another type of convolution, called multiplicative convolution, is rather younger. In this article, we first focus on this newer concept and obtain several useful results in which the most important ones is that if is logconcave then so are and for some suitable increasing functions ?. This result contains and as two more important special cases. Furthermore, one table including more applied distributions comparing logconcavity of f(x) and and two comprehensive implications charts are provided. Then, these fundamental results are applied to aging properties, existence of moments and several kinds of ordered random variables. Multiplicative strong unimodality in the discrete case is also introduced and its properties are investigated. In the second part of the article, some refinements are made for additive convolutions. A remaining open problem is completed and a conjecture concerning convolution of discrete α‐unimodal distributions is settled. Then, we shall show that an existing result regarding convolution of symmetric discrete unimodal distributions is not correct and an easy alternative proof is presented. © 2016 Wiley Periodicals, Inc. Naval Research Logistics 63: 109–123, 2016     

Lot sizing with learning and forgetting in setups: Analytical results and insights

This article considers the dynamic lot sizing problem when there is learning and forgetting in setups. Learning in setups takes place with repetition when additional setups are made and forgetting takes place when there is a break between two successive setups. We allow the amount forgotten over a break to depend both on the length of the break and the amount of learning at the beginning of the break. The learning and forgetting functions we use are realistic. We present several analytical results and use these in developing computationally efficient algorithms for solving the problem. Some decision/forecast horizon results are also developed, and finally we present managerial insights based on our computational results. © 2016 Wiley Periodicals, Inc. Naval Research Logistics 63: 93–108, 2016     

Strategy in the Age of Artificial Intelligence

ABSTRACT

We argue that Artificial Intelligence (AI) will, in the very near future, have a profound impact on the conduct of strategy and will be disruptive of existing power balances. To do so, we review the psychological foundations of strategy and explore the ways in which AI will impact human decision-making. We then review current and evolving capabilities in ‘narrow’, modular AI that is optimised to perform in a particular environment, and explore its military potential. Lastly, we look ahead to the more distant prospect of a general AI.