基于深度强化学习的维吾尔语人称代词指代消解

针对深度神经网络模型仅学习当前指代链语义信息忽略了单个指代链识别结果的长期影响问题,提出一种结合深度强化学习（deep reinforcement learning）的维吾尔语人称代词指代消解方法.该方法将指代消解任务定义为强化学习环境下顺序决策过程,有效利用之前状态中先行语信息判定当前指代链指代关系.同时,采用基于整体奖励信号优化策略,相比于使用损失函数启发式优化特定的单个决策,该方法直接优化整体评估指标更加高效.最后在维吾尔语数据集进行实验,实验结果显示,该方法在维吾尔语人称代词指代消解任务中的F值为85.80%.实验结果表明,深度强化学习模型能显著提升维吾尔语人称代词指代消解性能.     

跨模态医学图像预测综述

医学影像技术与设备的进步在生物医学领域的各项研究中发挥着重要作用.跨模态医学图像预测旨在由一种模态图像预测另一种模态图像.本文详细综述了由MRI预测CT图像、7T-Like图像重构、PET预测及其他医学模态预测研究,阐述了各类模态预测的必要性及存在的挑战,说明各类预测方法的特点并进行性能比较,最终得出结论：基于深度学习的跨模态预测在预测精度和预测时间两方面更具优势.     

基于迁移学习的唐诗宋词情感分析

随着计算社会学的兴起,利用数据挖掘分析社会情感是近期的研究重点.当前的研究主要针对现代文本,对于古代诗歌这类短文本的情感分析相对较少.本文提出了一个基于短文本特征扩展的迁移学习模型CATL-PCO,通过分析诗歌情感对当时社会及文化进行进一步了解.该模型首先基于频繁词对对古文特征向量进行扩展,再通过迁移学习方式,建立三个分类器并投票得出最后的情感分析结果.CATL-PCO模型首先能够解决古文短文本特征稀疏的问题,在此基础上进一步解决由于现代译文信息匮乏所导致的古代诗歌情感分析困难问题,从而准确的分析古诗词情感倾向,从计算社会学的角度,增进对中国历史的认识.实验表明,当训练集为中国唐诗时,本文提出方法能够准确的对唐代诗歌进行情感分类,并能应用于唐代和宋代各个时期情感分析及代表流派分析.     

基于微信公众平台的外语微学习资源设计与应用研究

针对微信使用群体的普及,本文提出在构建外语学习资源信息管理系统的基础上,搭建一个用于学生校园学习的微信服务账号,从而通过该微信服务账号实现学生对学习资源的利用,更好的促进语言能力的提升。本文借助Web服务器、Java技术、数据库、XML解析技术等系统进行开发,并对其实现进行详细的阐述,从而实现信息化技术与教学方式的创新,具有很大的推广和应用价值。     

基于MG-LTP与ELM的微表情识别

特征提取和表情分类是表情识别的关键技术。针对传统方法识别率低的缺点,首先,提出了一种基于平均灰度的局部三值模式(MG-LTP)新算法,用于提取表情特征;其次,使用极限学习机(ELM)作为分类器,用于特征分类;最后,将二者结合用于表情识别,并进一步应用于人脸微表情识别中。在JAFFE数据库及CASME人脸微表情数据库进行试验,与传统方法对比,取得了较好的效果。     

Cooperative Q‐learning techniques for distributed online power allocation in femtocell networks

In this paper, we address the problem of distributed interference management of femtocells that share the same frequency band with macrocells using distributed multi‐agent Q‐learning. We formulate and solve two problems representing two different Q‐learning algorithms, namely, femto‐based distributed and sub‐carrier‐based distributed power controls using Q‐learning (FBDPC‐Q and SBDPC‐Q). FBDPC‐Q is a multi‐agent algorithm that works on a global basis, for example, deals with the aggregate macrocell and femtocell capacities. Its complexity increases exponentially with the number of sub‐carriers in the system. Also, it does not take into consideration the sub‐carrier macrocell capacity as a constraint. To overcome these problems, SBDPC‐Q is proposed, which is a multi‐agent algorithm that works on a sub‐carrier basis, for example, sub‐carrier macrocell and femtocell capacities. Each of FBDPC‐Q and SBDPC‐Q works in three different learning paradigms: independent (IL), cooperative (CL), and weighted cooperative (WCL). IL is considered the simplest form for applying Q‐learning in multi‐agent scenarios, where all the femtocells learn independently. CL and WCL are the proposed schemes in which femtocells share partial information during the learning process in order to strike a balance between practical relevance and performance. We prove the convergence of the CL paradigm when used in the FBDPC‐Q algorithm. We show via simulations that the CL paradigm outperforms the IL paradigm in terms of the aggregate femtocell capacity, especially in networks with large number of femtocells and large number of power levels. In addition, we propose WCL to address the CL limitations. Finally, we evaluate the robustness and scalability of both FBDPC‐Q and SBDPC‐Q, against several typical dynamics of plausible wireless scenarios (fading, path loss, random activity of femtocells, etc.). We show that the CL paradigm is the most scalable to large number of femtocells and robust to the network dynamics compared with the IL and WCL paradigms. Copyright © 2014 John Wiley & Sons, Ltd.     

Learning curve analysis of concentrated photovoltaic systems

Price declines and volume growth of concentrated photovoltaic (CPV) systems are analysed using the learning curve methodology and compared with other forms of solar electricity generation. Logarithmic regression analysis determines a learning rate of 18% for CPV systems with 90% confidence of that rate being between 14 and 22%, which is higher than the learning rates of other solar generation systems (11% for CSP and 12 to 14% for PV). Current CPV system prices are competitive with PV and CSP, which, when combined with the higher learning rate, indicates that CPV is likely to further improve its marketability. A target price of 1 $/W in 2020 could be achieved with a compound growth rate of 67% for the total deployed volume between 2014 and 2020, which would realize a cumulative deployed volume of 7900 MW. Other projections of deployment volumes from commercial sources are converted using the learning rate into future price scenarios, resulting in predicted prices in the range of 1.1 to 1.3 $/W in 2020. © 2014 The Authors. Progress in Photovoltaics: Research and Applications published by John Wiley & Sons Ltd.     

基于快速神经网络架构搜索的鲁棒图像水印网络算法

为解决深度学习在图像水印算法中计算量大且模型冗余的问题,提高图像水印算法在抵抗噪声、旋转和剪裁等攻击时的鲁棒性,提出基于快速神经网络架构搜索（neural architecture search,NAS）的鲁棒图像水印网络算法。通过多项式分布学习快速神经网络架构搜索算法,在预设的搜索空间中搜索最优网络结构,进行图像水印的高效嵌入与鲁棒提取。首先,将子网络中线性连接的全卷积层设置为独立的神经单元结构,并参数化表示结构单元内节点的连接,预先设定结构单元内每个神经元操作的搜索空间;其次,在完成一个批次的数据集训练后,依据神经元操作中的被采样次数和平均损失函数值动态更新概率;最后,重新训练搜索完成的网络。水印网络模型的参数量较原始网络模型缩减了92%以上,大大缩短了模型训练时间。由于搜索得到的网络结构更为紧凑,本文算法具有较高的时间性能和较好的实验效果,在隐藏图像时,对空域信息的依赖比原始网络更少。对改进前后的2个网络进行了大量鲁棒性实验,对比发现,本文算法在CIFAR-10数据集上对抵抗椒盐噪声和旋转、移除像素行（列）等攻击优势显著;在ImageNet数据集上对抵抗椒盐高斯噪声、旋转、中值滤波、高斯滤波、JPEG压缩、裁剪等攻击优势显著,特别是对随机移除行（列）和椒盐噪声有较强的鲁棒性。