基于生成对抗网络的音乐标签自动标注

针对如何快速有效地对音乐信息进行查询、检索和组织的问题,提出了一种基于生成对抗网络模型的多标签音乐自动标注系统.通过音乐自动语义标注技术,可以提高音乐检索系统的性能.利用LDA方法对音乐标签进行聚类以获取主题类别,再通过生成对抗网络,找到音乐的音频特征与语义特征之间的映射关系.应用于CAL500数据集的5次交叉验证实验结果表明,该方法的综合性能指标与现有方法相比有较大的提升.     

系统辨识(4)：辅助模型辨识思想与方法

辅助模型辨识思想、多新息辨识理论、递阶辨识原理、耦合辨识概念是该文作者提出的研究辨识问题的原创性新方法,已经被用在很多辨识问题的研究中,形成了不同的辨识方法族,可以用于解决许多线性或非线性模型的自适应信号处理、自适应参数估计、自适应滤波和预测、自适应控制等问题.由于客观事物具有双重属性:一些特征变量是可观测的;一些是不可测的.如果表征系统特征的观测变量都是可测的,就容易建立描述其运动规律的数学模型.客观事物的不可测属性给建立系统数学模型带来特别的困难.在这种情况下,如何利用系统的可测信息,实现对系统未知变量的估算,来建立系统的数学模型,是辨识领域极具挑战性的研究课题.辅助模型辨识思想就是在这样的背景下发展起来的.该文介绍辅助模型辨识思想和一些基于辅助模型的辨识方法.     

小样本目标检测的研究现状

近年来,随着深度学习技术的进步与推广,目标检测领域得到快速发展.但目前基于深度学习的方法大多对大规模标注数据有着极高的需求,而现实场景中大量标注数据往往不可能.因此,基于少量标注样本的目标检测领域逐渐得到大家关注.本文系统地总结与分析了目前有关小样本目标检测的方法,指出了目前方法的缺陷,并提出了一些可能的发展方向.     

基于S-LSTM模型利用‘槽值门’机制的说话人意图识别

对说话人意图的识别极大地推进了自然语言理解任务的发展.之前的工作大多采用Bi-LSTM即双向LSTM模型进行词汇特征与词汇之间语义关系的提取,但这并不能很好地使句子整体和构成句子的词汇个体之间的信息进行交流.而S-LSTM（Sentence-state LSTM）模型,即句子状态LSTM模型可以很好地将自然语言中句子整体与词汇个体的信息相结合,以便于我们挖掘与利用意图检测与槽值填充之间的关系成立联合模型来更好地理解应答系统中蕴含的语义.因此,本文引入了‘槽值门’机制解决S-LSTM应用于意图检测与槽填充的联合任务时最新迭代时刻的句子状态信息没有得到充分利用的问题.最终的实验结果在ATIS数据集和Snips数据集上均取得了优于目前最先进算法的结果.     

基于实例的词性标注方法研究

汉语的词性自动标注工作是自然语言处理的基础性课题，在文本自动分类、全文自动检索和自动文摘等领域都有重要的实际应用价值，本文首先对传统的词性标注方法进行了介绍，然后提出了一种基于实例的词性标注方法，并对其进行了一些尝试性的研究。     

基于用户点击数据的细粒度图像识别方法概述

近年来,细粒度图像识别逐渐成为计算机视觉领域的研究热点.由于不同类别图像间的视觉差异小、语义鸿沟问题严重,传统的基于视觉特征的细粒度图像识别性能往往不尽人意.针对这些挑战,目前许多学者都在研究基于用户点击数据的图像识别.本文围绕点击数据在图像识别中数据预处理、特征提取和模型构建3大模块中的应用,总结了已有的基于点击数据的识别算法及最新的研究进展.     

基于加密降水资料的贵州地质灾害概率预报模型

运用信息量法进行了地质灾害易发性评价,将全省划分为3个易发等级。利用贵州省2009—2014年的地质灾害和降水数据,分析了每个等级中地质灾害概率与有效降水量之间的关系。在此基础上,利用概率论方法将动态临界雨量引入预报模型,建立了基于动态临界雨量的地质灾害概率预报模型。结果表明:对于任意的易发区划等级,诱发地质灾害的临界雨量并不是静态的,而是分布在一个非常大的区间上;当有效降水量为该区间上一个特定的值时,用相应的概率来描述地质灾害发生可能性的大小,而不是地质灾害一定发生更为恰当;在2015年的检验中,模型预报准确率超过80%,说明该概率预报方法具有实际应用价值,能起到良好的预报预警作用。     

基于冗余属性的OOX文本数字水印算法

针对流行最广的Office Open XML 格式文档(即MS Office 2007—2013),提出一种基于冗余属性的文本数字水印算法.利用OOX(Office Open XML)文档包中部件的冗余信息来嵌入水印,并使水印与文档的格式信息绑定,可有效抵抗针对文本内容的攻击.该方法具有较强的鲁棒性和较大的嵌入容量,能够抵抗"另存为"、"清除格式"等攻击,且没有改变原始文档显示字符的任何格式信息,这是以往针对格式文档提出的水印方法做不到的,同时,也没有改变原始文档的任何语义信息,这也是以往的基于语义(自然语言)的水印方法做不到的.实验结果说明了该算法的可行性.     

天气预报的模糊推理

目前应用的模糊推理,其前提是根据自身有关资料建立模糊集合,结论的模糊集合则是人为规定。如有雨规定为1,可能有雨规定为0.5,无雨规定为0。人为规定不能正确描述结论中各级天气模糊数量关系,而且也没有针对天气预报特点将前提、结论的模糊属性同前提、结论之间的概率属性联系起来,基于以上考虑,本文针对天气预报模糊、概率双重属性的特点,设计了一个模糊推理模型,并进行了实际应用,现介绍如下:     

知识图谱发展与构建的研究进展

近年来知识图谱技术引起了广泛的关注和研究,本文介绍了近期知识图谱的发展、构建方法、详细的构建过程,并对知识图谱在交叉学科领域的应用和未来的研究方向做了总结.本文详细介绍了构建文本知识图谱、视觉知识图谱、多模态知识图谱的关键技术,比如信息提取、知识融合、知识表示等.作为知识工程的重要组成部分,知识图谱,尤其是多模态知识图谱的发展对大数据时代的高效知识管理、知识获取、知识共享有着重要的意义.     

RGB-T目标跟踪综述

RGB-T目标跟踪是基于RGB目标跟踪问题发展而来的.为了提高复杂环境下的目标跟踪性能,学者们提出结合可见光和热红外的信息来克服单一成像受限的问题.本文首先介绍了RGB-T目标跟踪的研究背景,并指出该任务所面临的挑战,然后归纳并介绍了目前已有的RGB-T目标跟踪的几类方法,包括传统方法和深度学习方法.最后,本文对现有的RGB-T数据集、评价指标进行了分析和对比,并指出RGB-T跟踪中值得研究的方面.     

高光谱图像分类方法综述

在过去数十年中,高光谱图像的研究与应用已经完成了从无到有、从差到优的跨越式发展.在对其研究的众多方面中,高光谱图像分类已经成为了一个最热的研究主题.研究表明空间光谱联合的分类方法可以取得比仅依赖光谱信息的逐像素分类方法更好的分类效果.本文将对众多的空间光谱联合分类方法进行归类和分析.首先介绍高光谱图像中相邻像素间的两类空间依赖性关系,因而可将现有的空谱联合分类方法分为依赖固定邻域和自适应邻域两类;此外,还可以依据是否同时利用两类依赖关系将现有方法进一步分为单依赖和双依赖两类.另外,还可以依据空谱信息融合的不同阶段将现有的分类方法划分为预处理方法、一体化方法及后处理方法三类.最后展示几种具有代表性的空间光谱联合分类方法在真实高光谱数据集上的分类结果.     

Learning effects of interactive decision-making processes for climate change adaptation

Learning is gaining attention in relation to governance processes for contemporary environmental challenges; however, scholarship at the nexus of learning and environmental governance lacks clarity and understanding about how to define and measure learning, and the linkages between learning, social interactions, and environment. In response, this study aimed to advance and operationalize a typology of learning in an environmental governance context, and examined if a participatory decision-making process (adaptive co-management) for climate change adaptation fostered learning. Three types of learning were identified: cognitive learning, related to the acquisition of new or the structuring of existing knowledge; normative learning, which concerns a shift in viewpoints, values or paradigms, and relational learning, referring to an improved understanding of others’ mindsets, enhanced trust and ability to cooperate. A robust mixed methods approach with a focus on quantitative measures including concept map analysis, social network analysis, and self-reflective questions, was designed to gauge indicators for each learning type. A participatory decision-making process for climate change adaptation was initiated with stakeholders in the Niagara region, Canada. A pseudo-control group was used to minimize external contextual influences on results. Clear empirical evidence of cognitive and relational learning was gained; however, the results from normative learning measures were inconclusive. The learning typology and measurement method operationalized in this research advances previous treatments of learning in relation to participatory decision-making processes, and supports adaptive co-management as a governance strategy that fosters learning and adaptive capacity.     

ConvLSTM Based Temperature Forecast Modification Model for North China

The correction of model forecast is an important step in evaluating weather forecast results. In recent years,post-processing models based on deep learning have become prominent. In this paper, a deep learning model named ED-ConvLSTM based on encoder-decoder structure and ConvLSTM is developed, which appears to be able to effectively correct numerical weather forecasts. Compared with traditional post-processing methods and convolutional neural networks, ED-ConvLSTM has strong collaborative extra...     

Environmentalities of urban climate governance in Maputo,Mozambique

Interest in the role that cities can play in climate change as sites of transformation has increased but research has been limited in its practical applications and there has been limited consideration of how policies and technologies play out. These challenges necessitate a re-thinking of existing notions of urban governance in order to account for the practices that emerge from governments and a plethora of other actors in the context of uncertainty. We understand these practices to constitute adaptive governance, underpinned by social learning guiding the actions of the multiplicity of actors. The aim here is to unpack how social learning for adaptive governance requires attention to competing understandings of risk and identity, and the multiplicity of mechanisms in which change occurs or is blocked in urban climate governance. We adopt a novel lens of ‘environmentalities’ which allows us to assess the historical and institutional context and power relations in the informal settlements of Maputo, Mozambique. Our findings highlight how environmental identities around urban adaptation to climate change are constituted in the social and physical divisions between the formal and informal settlements, whilst existing knowledge models prioritise dominant economic and political interests and lead to the construction of new environmental subjects. While the findings of this study are contextually distinct, the generalizable lessons are that governance of urban adaptation occurs and is solidified within a complex multiplicity of socio-ecological relations.     

光流法及其在临近预报中的应用

短临降水预报对于暴雨和强对流天气监测预警服务具有重要意义,使用雷达回波外推方法进行短临降水预报是目前较为常用的预报方法之一,但是传统的雷达回波外推方法普遍存在数据利用率低、外推准确性差和外推模糊等问题。针对上述问题,利用陕西全省雷达拼接数据资料,选择深度学习中编码器-解码器结构,以卷积长短期记忆网络(Convolutional Long Short-Term Memory Network, ConvLSTM)作为循环单元,构造了基于全局通道注意力的ConvLSTM预测网络(Global Channel Attention based ConvLSTM, GCA-ConvLSTM);此外,为进一步提高GCA-ConvLSTM预测网络的拟合能力,使用集成学习算法对其进行改进,通过装袋算法对数据集进行采样,训练3个GCA-ConvLSTM预测网络作为基学习器,使用加权投票策略将这3种基学习器进行有效组合,最终获得了一个性能更优的组合模型。试验结果表明,基于集成学习算法改进的GCA-ConvLSTM雷达回波外推方法与现有深度学习方法相比,提升了短临降水预报方法的准确性和时效;该方法在25 dBz、35 dBz和45 dBz反射率阈值下的评估试验中分别比对比的主流深度学习模型CSI值平均高出0.149、0.192、0.085;同时该方法的外推结果拥有更加清晰的边缘和细节性纹理,减轻了外推后期模糊问题。

     

A new,valid measure of climate change understanding: associations with risk perception

The relationship between climate change understanding and other variables, including risk perception, beliefs, and worldviews, is an important consideration as we work to increase public attention to climate change. Despite significant effort to develop rigorous mechanisms for measuring affective variables, measurement of climate change understanding is often relegated to unvalidated questions or question sets. To remedy this situation, we constructed and analyzed a climate change concept inventory using a suite of validity and reliability steps, including Rasch analysis. The resultant 21-item test has a high degree of validity and reliability for measuring understanding about basic climate change processes. Inventory scores along with other variables were included in a model of climate change risk perception, providing both concurrent validity for the test and new insight into the importance of understanding, worldview, and values on risk perception. We find that environmental beliefs and cultural cognition worldview play a larger role in predicting an individual’s risk perception than knowledge. Implications for addressing climate change are considered.     

深度细粒度图像识别研究综述

细粒度图像分类是计算机视觉中一项基础且重要的工作,其目的在于区分难以辨别的对象类别（例如不同子类的鸟类、花或动物）.不同于传统的图像分类任务可以雇佣大量普通人标注,细粒度数据集通常需要专家级知识进行标注.除了视觉分类中常见的姿态、光照和视角变化因素之外,细粒度数据集具有更大的类间相似性和类内差异性,因此要求模型能够捕捉到细微的类间差异信息和类内公有信息.除此之外,不同类别的样本存在不同程度的获取难度,因此细粒度数据集通常在数据分布中表现出长尾的特性.综上所述,细粒度数据分布具有小型、非均匀和不易察觉的类间差异等特点,对强大的深度学习算法也提出了巨大的挑战.本文首先介绍了细粒度图像分类任务的特点与挑战,随后以局部特征与全局特征两个主要视角整理了目前的主流工作,并讨论了它们的优缺点.最后在常用数据集上比较了相关工作的性能表现,并进行了总结与展望.     

Disconcerting learning on climate sensitivity and the uncertain future of uncertainty

How will our estimates of climate uncertainty evolve in the coming years, as new learning is acquired and climate research makes further progress? As a tentative contribution to this question, we argue here that the future path of climate uncertainty may itself be quite uncertain, and that our uncertainty is actually prone to increase even though we learn more about the climate system. We term disconcerting learning this somewhat counter-intuitive process in which improved knowledge generates higher uncertainty. After recalling some definitions, this concept is connected with the related concept of negative learning that was introduced earlier by Oppenheimer et al. (Clim Change 89:155–172, 2008). We illustrate disconcerting learning on several real-life examples and characterize mathematically certain general conditions for its occurrence. We show next that these conditions are met in the current state of our knowledge on climate sensitivity, and illustrate this situation based on an energy balance model of climate. We finally discuss the implications of these results on the development of adaptation and mitigation policy.     

Meshless Surface Wind Speed Field Reconstruction Based on Machine Learning

We propose a novel machine learning approach to reconstruct meshless surface wind speed fields, i.e., to reconstruct the surface wind speed at any location, based on meteorological background fields and geographical information. The random forest method is selected to develop the machine learning data reconstruction model (MLDRM-RF) for wind speeds over Beijing from 2015–19. We use temporal, geospatial attribute and meteorological background field features as inputs. The wind speed field can be reconstructed at any station in the region not used in the training process to cross-validate model performance. The evaluation considers the spatial distribution of and seasonal variations in the root mean squared error (RMSE) of the reconstructed wind speed field across Beijing. The average RMSE is 1.09 m s^?1, considerably smaller than the result (1.29 m s^?1) obtained with inverse distance weighting (IDW) interpolation. Finally, we extract the important feature permutations by the method of mean decrease in impurity (MDI) and discuss the reasonableness of the model prediction results. MLDRM-RF is a reasonable approach with excellent potential for the improved reconstruction of historical surface wind speed fields with arbitrary grid resolutions. Such a model is needed in many wind applications, such as wind energy and aviation safety assessments.