在本文中,我们提出了一个新的实体关系提取任务范例。我们将任务转换为多回合问题回答问题,即,实体和关系的提取被转换为从上下文识别答案跨度的任务。这种多转QA形式化有几个关键优势:首先,问题查询编码我们想要识别的实体/关系类的重要信息;其次,QA提供了一种自然的方式来联合建模实体和关系;第三,它允许我们利用完善的机器阅读理解(MRC)模型。在ACE和CoNLL04公司的实验表明,所提出的范例明显优于以前的最佳模型。我们能够获得所有ACE04,ACE05和CoNLL04数据集的最新结果,增加了三个数据集的SOTA结果49.6(+1.2),60.3(+0.7)和69.2(+1.4) , 分别。此外,我们构建了一个新开发的数据集RESUME,它需要多步推理来构造实体依赖关系,而不是先前数据集中三元组提取中的单步依赖提取。提出的多转QA模型也在RESUME数据集上实现了最佳性能。
translated by 谷歌翻译
将一大块文本分割成单词通常是处理中文文本的第一步,但其必要性很少被探索过。在本文中,我们提出了中国分词(CWS)是否是基于深度学习的中国自然语言处理所必需的基本问题。基于神经词汇的Webenchmark模型依赖于基于神经char的模型的分词,这些模型不涉及四端到端NLP基准测试任务中的分词:语言建模,机器翻译,句子匹配/释义和文本分类。通过这两种模型之间的直接比较,我们发现基于char的模型始终优于基于单词的模型。基于这些观察,我们进行了全面的实验,以研究为什么基于单词的模型在这些基于深度学习的NLP任务中表现不佳。我们表明,因为基于单词的模型更容易受到数据稀疏性和词汇表外(OOV)词的影响,因此更容易过度拟合。我们希望本文能够鼓励社区研究人员重新思考基于深度学习的中国自然语言处理中分词的必要性。 \脚注{Yuxian Meng和Xiaoya Li对本报的贡献相同。}
translated by 谷歌翻译
提出了一种新的双线性判别特征线分析(BDFLA)用于图像特征提取。最近的特征线(NFL)是一个功能强大的分类器。最近引入了一些基于NFL的子空间算法。在大多数基于NFL的经典子空间学习方法中,输入样本是矢量。对于图像分类任务,图像样本应首先转换为矢量。该过程引起高计算复杂性并且还可能导致样品的几何特征的损失。提出的BDFLA是基于矩阵的算法。它旨在最大限度地减少类内散射,并最大化基于二维(2D)NFL的类间散射。双图像数据库的实验结果证实了其有效性。
translated by 谷歌翻译
目的:组织氧饱和度(StO2)的术中测量在缺血检测,监测灌注和鉴别疾病中是重要的。高光谱成像(HSI)测量组织的光学反射光谱,并使用该信息量化其组成,包括StO2。但是,由于捕获率和数据处理时间,实时监控很困难。方法:先前开发了一种基于多光纤探头的内窥镜系统,以稀疏地捕获HSI数据(sHSI)。这些通过深度神经网络与RGB图像组合,以生成高分辨率超立方体并计算StO2。为了提高准确度和处理速度,我们提出了双输入条件生成对抗网络(cGAN)Dual2StO2,通过融合RGB和sHSI的特征来直接估计StO2。结果:在vivoporcine肠道数据中进行了验证实验,其中地面实况StO2是从HSIcamera生成的。在平均StO2预测精度和结构相似性度量方面,还将性能与我们的先前光谱分辨率网络SSRNet进行了比较。 Dual2StO2也使用具有不同光纤数量的模拟探针数据进行测试。结论:在一般结构中,由二重StO2估计的StO2在视觉上更接近于地面实况,与SSRNet相比,实现了更高的预测精度和更快的处理速度。仿真表明,当在探头中使用更多数量的光纤时,结果得到改善。未来的工作将包括网络架构的改进,基于模拟结果的硬件优化,以及超出StO2估计的临床应用技术评估。
translated by 谷歌翻译
我们分析随机梯度算法来优化非凸问题。特别是,我们的目标是找到局部最小值(二阶静止点)而不是仅找到一阶静止点,这可能是某些不稳定的鞍点。我们证明了一个简单的扰动版本的随机递归梯度下降算法(称为SSRGD)可以找到$(\ epsilon,\ delta)$ - 二阶固定点,其中$ \ widetilde {O}(\ sqrt {n} / \ epsilon ^ 2 + \ sqrt {n} / \ delta ^ 4 + n / \ delta ^ 3)非凸有限和问题的$随机梯度复杂度。作为aby-product,SSRGD使用$ O(n + \ sqrt {n} / \ epsilon ^ 2)$随机梯度找到$ \ epsilon $ -first-order固定点。自Fang等人以来,这些结果几乎是最优的。 [2018]提供了一个下限$ \ Omega(\ sqrt {n} / \ epsilon ^ 2)$,用于查找甚至只是$ \ epsilon $ -first-orderstationary点。我们强调用于寻找二阶点的SSRGD算法就像通过有时添加均匀扰动来寻找一阶静止点一样简单,而用于寻找具有相似梯度复杂度的二阶静止点的所有其他算法需要具有负曲率的tocombine搜索子程序(例如,Neon2 [Allen-Zhu andLi,2018])。此外,简单的SSRGD算法得到了更简单的分析。此外,我们还扩展了非凸的有限和问题tononconvex在线(期望)问题的结果,并证明了相应的收敛结果。
translated by 谷歌翻译
多形性胶质母细胞瘤(GBM)是一种头部肿瘤,具有非常复杂的治疗过程。存活期通常为14-16个月,2年存活率约为26%-33%。 GBM的假性进展(PsP)和真实肿瘤进展(TTP)的临床治疗策略是不同的,因此准确区分这两种情况具有特别重要的意义。由于GBM的PsP和TTP在形状和其他特征上相似,因此很难区分这两种形式。精确。为了准确区分它们,本文介绍了一种基于生成对抗网络的特征学习方法:DC-Al GAN。 GAN由两种架构组成:发电机和鉴别器。 Alexnet在这项工作中被用作鉴别者。由于发生器和鉴别器之间的对抗性和竞争性关系,后者在训练期间提取高度协调的特征。在DC-Al GAN中,在最终分类阶段从特征网中提取特征,并且它们的高性质对分类精度有积极贡献。通过添加三个卷积,通过深度卷积生成对抗网络(DCGAN)修改DC-AlGAN中的生成器。层。这有效地生成了更高分辨率的样本图像。特征融合用于将高层特征与低层特征相结合,允许创建和使用更精确的特征进行分类。实验结果证实,DC-Al GAN对用于PsP和TTP图像分类的GBM数据集具有高精度,优于其他最先进的方法。
translated by 谷歌翻译
尽管深度学习取得了巨大成功,但我们对如何训练凸凸神经网络的理解仍然相当有限。大多数现有的理论工作只涉及具有一个隐藏层的神经网络,而且对于多层神经网络知之甚少。递归神经网络(RNN)是在自然语言处理应用中广泛使用的特殊多层网络。与前馈网络相比,它们特别难以分析,因为权重参数在整个时间范围内重复使用。我们可以说是对训练RNN的收敛速度的第一个理论上的理解。具体来说,当神经元的数量足够大时 - 意味着训练数据大小和时间范围内的多项式 - 以及随机初始化权重时,我们表明梯度下降和随机梯度下降都可以最小化线性收敛中的训练损失率,即$ \ varepsilon \ proptoe ^ { - \ Omega(T)} $。
translated by 谷歌翻译
最近,卷积神经网络(CNN)在面部检测方面取得了巨大成功。然而,由于尺度,姿势,遮挡,表现,外观和照明的高度可变性,它对于当前的面部检测方法仍然是一个具有挑战性的问题。在本文中,我们提出了一种名为Dual Shot人脸检测器(DSFD)的人脸检测网络,该网络继承了SSD的体系结构,并引入了一个特征增强模块(FEM),用于传输原始特征图以扩展单镜头检测双镜头检测器。特别地,采用两组锚点计算的渐进锚杆损失(PAL)来有效地促进这些特征。另外,我们提出了一种改进的锚点匹配(IAM)方法,通过在DSFD中集成新的数据增强技术和锚点设计策略,为回归者。流行基准测试的广泛实验:WIDER FACE(简单:$ 0.966 $,中:$ 0.957 $,硬:$ 0.904 $)和FDDB(不连续:$ 0.991 $,连续:$ 0.862 $)展示了DSFD优于最先进的技术面部检测器(例如,PyramidBox和SRN)。代码将在发布时提供。
translated by 谷歌翻译
Residential transformer population is a critical type of asset that many electric utility companies have been attempting to manage proactively and effectively to reduce unexpected transformer failures and life loss that are often caused by overloading. Within the typical power asset portfolio, the residential transformer asset is often large in population, has the lowest reliability design, lacks transformer loading data and is susceptible to customer loading behaviors, such as adoption of distributed energy resources and electric vehicles. On the bright side, the availability of more residential service operation data along with the advancement of data analytics techniques has provided a new path to further our understanding of residential transformer overloading risk statistically. This research developed a new data-driven method that combines a transformer temperature rise and insulation life loss simulation model with clustering analysis technique. It quantitatively and statistically assesses the overloading risk of residential transformer population in one area and suggests proper risk management measures according to the assessment results. Multiple application examples for a Canadian utility company have been presented and discussed in detail to demonstrate the applicability and usefulness of the proposed method. Index Terms-power system reliability, clustering methods, transformers, life estimation, unsupervised learning. Ming Dong (S ' 08 , M ' 13, SM'18) received his doctoral degree from in 2013. Since graduation, he has been working in various roles in two major electric utility companies in West Canada as a Professional Engineer (P.Eng.) and Senior Engineer for more than 5 years. In 2017, he received the Certificate of Data Science and Big Data Analytics from Massachusetts Institute of Technology. He is also a regional officer of Alberta Artificial Intelligence Association. His research interests include applications of artificial intelligence and big data technologies in power system planning and operation, power quality data analytics, power equipment testing and system grounding. Alexandre Nassif (S'05, M'09, SM'13) is a specialist engineer in ATCO Electric. He published more than 50 technical papers in international journals and conferences in the areas of power quality, DER, microgrids and power system protection and stability. Before joining ATCO, he simultaneously worked for Hydro One as a protection planning engineer and Ryerson University as a post-doctoral research fellow. He holds a doctoral degree from the University of Alberta and is a Professional Engineer in Alberta. Benzhe Li (M'18) received his Master's degree from Department of Electrical and Computer Engineering, University of Alberta, Canada in 2015. He is currently an electrical engineer with Energy Ottawa. His research interests include advanced power quality data analytics and equipment condition monitoring.
translated by 谷歌翻译
Heterogeneous network embedding (HNE) is a challenging task due to the diverse node types and/or diverse relationships between nodes. Existing HNE methods are typically unsupervised. To maximize the profit of utilizing the rare and valuable supervised information in HNEs, we develop a novel Active Heterogeneous Network Embedding (Ac-tiveHNE) framework, which includes two components: Discriminative Heterogeneous Network Embedding (DHNE) and Active Query in Heterogeneous Networks (AQHN). In DHNE, we introduce a novel semi-supervised heterogeneous network embedding method based on graph convolutional neu-ral network. In AQHN, we first introduce three active selection strategies based on uncertainty and representativeness, and then derive a batch selection method that assembles these strategies using a multi-armed bandit mechanism. ActiveHNE aims at improving the performance of HNE by feeding the most valuable supervision obtained by AQHN into DHNE. Experiments on public datasets demonstrate the effectiveness of ActiveHNE and its advantage on reducing the query cost.
translated by 谷歌翻译