长向量处理器高效RNN推理方法

模型深度的不断增加和处理序列长度的不一致对循环神经网络在不同处理器上的性能优化提出巨大挑战。针对自主研制的长向量处理器FT-M7032,实现了一个高效的循环神经网络加速引擎。该引擎采用行优先矩阵向量乘算法和数据感知的多核并行方式,提高矩阵向量乘的计算效率;采用两级内核融合优化方法降低临时数据传输的开销;采用手写汇编优化多种算子,进一步挖掘长向量处理器的性能潜力。实验表明,长向量处理器循环神经网络推理引擎可获得较高性能,相较于多核ARM CPU以及Intel Golden CPU,类循环神经网络模型长短记忆网络可获得最高62.68倍和3.12倍的性能加速。     

基于转置卷积神经网络的翼型结冰冰形图像化预测方法

结冰问题严重影响飞机飞行安全,结冰智能预测是飞机智能防除冰系统设计和安全设计的重要依据和支撑.为解决复杂冰形在翼面同一位置的法线方向冰形厚度存在多值的问题,提出基于转置卷积神经网络的翼型结冰冰形图像化预测方法.设计预测模型的神经网络结构、损失函数、数据规范等,直接将影响飞机结冰的飞行和大气条件作为输入,以灰度化的冰形图...     

一种复合模式的人脸识别系统设计

通过综合考察红外 可见光图像人脸识别的优缺点,利用红外温谱图进行人脸检测与识别的优势,在研究可见光和红外人脸温谱图识别技术以及两者融合技术的基础上,提出和设计了一种新的红外 可见光复合模式的人脸识别系统。初步实验结果表明,这种识别方法不论从理论上还是从实验上都是可行的,系统具有良好的识别能力。     

Too little,too late: protecting American soft networks in COIN/CT

ABSTRACT

Assistance of local-national partners is necessary to the USA in order to protect its national security interests throughout the world. These partners, typically individuals who support USA diplomats, service members, and non-governmental organizations (NGOs) in non-combatant roles, form USA soft networks. Due to the nature of their work, many of these individuals associated with the policy and actions of the USA become vulnerable to violent threats from adversaries. In fact, adversaries have grasped that attacking USA soft networks is a logical approach for enemies fighting from positions of weakness. As a result, examination of other domains may yield best practices that build resiliency in USA soft networks, thereby cementing national security interests. Further, agency theory illuminates critical principles in the relationships between local-national partners and field practitioners, and informs policy development efforts as a result.