Support Vector Machine active learning for 3D model retrieval

In this paper, we present a novel Support Vector Machine active learning algorithm for effective 3D model retrieval using the concept of relevance feedback. The proposed method learns from the most informative objects which are marked by the user, and then creates a boundary separating the relevant models from irrelevant ones. What it needs is only a small number of 3D models labelled by the user. It can grasp the user＇s semantic knowledge rapidly and accurately. Experimental results showed that the proposed algorithm significantly improves the retrieval effectiveness. Compared with four state-of-the-art query refinement schemes for 3D model retrieval, it provides superior retrieval performance after no more than two rounds of relevance feedback.

Accurate performance estimators for information retrieval based on span bound of support vector machines

Support vector machines have met with significant success in the information retrieval field, especially in handling text classification tasks. Although various performance estimators for SVMs have been proposed, these only focus on accuracy which is based on the leave-one-out cross validation procedure. Information-retrieval-related performance measures are always neglected in a kernel learning methodology. In this paper, we have proposed a set of information-retrieval-oriented performance estimators for SVMs, which are based on the span bound of the leave-one-out procedure. Experiments have proven that our proposed estimators are both effective and stable.

Active learning based on maximizing information gain for content-based image retrieval

This paper describes a new method for active learning in content-based image retrieval. The proposed method firstly uses support vector machine (SVM) classifiers to learn an initial query concept. Then the proposed active learning scheme employs similarity measure to check the current version space and selects images with maximum expected information gain to solicit user's label. Finally, the learned query is refined based on the user's further feedback. With the combination of SVM classifier and similarity measure, the proposed method can alleviate model bias existing in each of them. Our experiments on several query concepts show that the proposed method can learn the user's query concept quickly and effectively only with several iterations.

Applying CBR to Machine Tool Product Configuration Design Oriented to Customer Requirements

Product customization is a trend in the current market-oriented manufacturing environment. However, deduction from customer requirements to design results and evaluation of design alternatives are still heavily reliant on the designer＇s experience and knowledge. To solve the problem of fuzziness and uncertainty of customer requirements in product configuration, an analysis method based on the grey rough model is presented. The customer requirements can be converted into technical characteristics effectively. In addition, an optimization decision model for product planning is established to help the enterprises select the key technical characteristics under the constraints of cost and time to serve the customer to maximal satisfaction. A new case retrieval approach that combines the self-organizing map and fuzzy similarity priority ratio method is proposed in case-based design. The self-organizing map can reduce the retrieval range and increase the retrieval efficiency, and the fuzzy similarity priority ratio method can evaluate the similarity of cases comprehensively. To ensure that the final case has the based on grey correlation analysis is proposed to evaluate similar cases best overall performance, an evaluation method of similar cases to select the most suitable case. Furthermore, a computer-aided system is developed using MATLAB GUI to assist the product configuration design. The actual example and result on an ETC series machine tool product show that the proposed method is effective, rapid and accurate in the process of product configuration. The proposed methodology provides a detailed instruction for the product configuration design oriented to customer requirements.

Intelligent Agent-Based System for Digital Library Information Retrieval

A new information search model is reported and the design and implementation of a system based on intelligent agent is presented. The system is an assistant information retrieval system which helps users to search what they need. The system consists of four main components: interface agent, information retrieval agent, broker agent and learning agent. They collaborate to implement system functions. The agents apply learning mechanisms based on an improved ID3 algorithm.

Content-based image retrieval applied to BI-RADS tissue classification in screening mammography

AIM:To present a content-based image retrieval(CBIR) system that supports the classification of breast tissue density and can be used in the processing chain to adapt parameters for lesion segmentation and classification.METHODS:Breast density is characterized by image texture using singular value decomposition(SVD) and histograms.Pattern similarity is computed by a support vector machine(SVM) to separate the four BI-RADS tissue categories.The crucial number of remaining singular values is varied(SVD),and linear,radial,and polynomial kernels are investigated(SVM).The system is supported by a large reference database for training and evaluation.Experiments are based on 5-fold cross validation.RESULTS:Adopted from DDSM,MIAS,LLNL,and RWTH datasets,the reference database is composed of over 10000 various mammograms with unified and reliable ground truth.An average precision of 82.14% is obtained using 25 singular values(SVD),polynomial kernel and the one-against-one(SVM).CONCLUSION:Breast density characterization using SVD allied with SVM for image retrieval enable the development of a CBIR system that can effectively aid radiologists in their diagnosis.

ORDINAL REGRESSION FOR INFORMATION RETRIEVAL

This letter presents a new discriminative model for Information Retrieval (IR), referred to as Ordinal Regression Model (ORM). ORM is different from most existing models in that it views IR as ordinal regression problem (i.e. ranking problem) instead of binary classification. It is noted that the task of IR is to rank documents according to the user information needed, so IR can be viewed as ordinal regression problem. Two parameter learning algorithms for ORM are presented. One is a perceptron-based algorithm. The other is the ranking Support Vector Machine (SVM). The effec- tiveness of the proposed approach has been evaluated on the task of ad hoc retrieval using three English Text REtrieval Conference (TREC) sets and two Chinese TREC sets. Results show that ORM sig- nificantly outperforms the state-of-the-art language model approaches and OKAPI system in all test sets; and it is more appropriate to view IR as ordinal regression other than binary classification.

Web Information Retrieval： Problem and Prospects

The information access is the rich data available for information retrieval, evolved to provide principle approaches or strategies for searching. For building the successful web retrieval search engine model, there are a number of prospects that arise at the different levels where techniques, such as Usenet, support vector machine are employed to have a significant impact. The present investigations explore the number of problems identified its level and related to finding information on web. The authors have attempted to examine the issues and prospects by applying different methods such as web graph analysis, the retrieval and analysis of newsgroup postings and statistical methods for inferring meaning in text. The proposed model thus assists the users in finding the existing formation of data they need. The study proposes three heuristics model to characterize the balancing between query and feedback information, so that adaptive relevance feedback. The authors have made an attempt to discuss the parameter factors that are responsible for the efficient searching. The important parameters can be taken care of for the future extension or development of search engines.

An Improved Asymmetric Bagging Relevance Feedback Strategy for Medical Image Retrieval

Much attention has been paid to relevant feedback in intelligent computation for social computing, especially in content-based image retrieval which based on WeChat platform for the medical auxiliary. It has a good effect on reducing the semantic gap between high semantics and low semantics of images. There are many kinds of support vector machines (SVM) based relevance feedback methods in image retrieval, but all of them may encounter some problems, such as a small size of sample, an asymmetric positive sample and negative sample as well as a long feedback cycle. To deal with these problems, an improved asymmetric bagging (IAB) relevance feedback algorithm is proposed. Furthermore, we apply a new fuzzy support machine (FSVM) to cooperate with IAB. To solve the over-fitting and real-time problems, we use modified local binary patterns (MLBP) as image features. Finally, experimental results demonstrate that our method performs other methods in terms of improving retrieval precision as well as retrieval efficiency.

News Modeling and Retrieving Information: Data-Driven Approach

This paper aims to develop Machine Learning algorithms to classify electronic articles related to this phenomenon by retrieving information and topic modelling.The Methodology of this study is categorized into three phases:the Text Classification Approach(TCA),the Proposed Algorithms Interpretation(PAI),andfinally,Information Retrieval Approach(IRA).The TCA reflects the text preprocessing pipeline called a clean corpus.The Global Vec-tors for Word Representation(Glove)pre-trained model,FastText,Term Frequency-Inverse Document Fre-quency(TF-IDF),and Bag-of-Words(BOW)for extracting the features have been interpreted in this research.The PAI manifests the Bidirectional Long Short-Term Memory(Bi-LSTM)and Convolutional Neural Network(CNN)to classify the COVID-19 news.Again,the IRA explains the mathematical interpretation of Latent Dirich-let Allocation(LDA),obtained for modelling the topic of Information Retrieval(IR).In this study,99%accuracy was obtained by performing K-fold cross-validation on Bi-LSTM with Glove.A comparative analysis between Deep Learning and Machine Learning based on feature extraction and computational complexity exploration has been performed in this research.Furthermore,some text analyses and the most influential aspects of each document have been explored in this study.We have utilized Bidirectional Encoder Representations from Trans-formers(BERT)as a Deep Learning mechanism in our model training,but the result has not been uncovered satisfactory.However,the proposed system can be adjustable in the real-time news classification of COVID-19.

ML组合的CYGNSS海面风速反演质量控制模型

卷积神经网络(CNN)可用于气旋全球导航卫星系统(CYGNSS)的海面风速反演。虽然在模型训练前设置了质量控制指标来检测和削弱CYGNSS的异常观测数据,但CYGNSS观测数据中仍存在异常值导致模型反演精度降低,甚至出现错误反演结果。因此,提出一种基于机器学习(ML)组合的海面风速反演模型。在基于CNN回归模型的CYGNSS反演海面风速基础上,ML分类模型生成CNN回归结果的质量标志位,该标志位可以检测并删除CNN回归结果的异常值,进一步提高风速反演结果的数据质量,ML分类模型能够更好地考虑各种数据误差之间的相互作用,而不是单独使用每个条件的阈值,以达到更优的海面风速反演精度的效果。实验对比了Logistic回归(LR)、决策树(DT)、朴素贝叶斯模型、K最邻近(KNN)算法、神经网络(NN)模型、支持向量机(SVM)算法等6个分类模型,其中,基于KNN算法的分类模型对风速反演质量控制的效果最优。所提风速反演组合模型显著提高了反演结果的精度,在0~20 m/s区间内,异常样本过滤率为81.27%,在所有被过滤的数据中,过滤正确率为86.03%;风速反演误差的均方根误差从无ML分类模型的1.7 m/s降低到有ML分类模型的1.44 m/s,其中,训练样本为0~10 m/s的反演结果精度提升效果较为明显,证明了所提风速反演组合模型对风速质量控制的有效性。

基于机器学习模型FY⁃3D MWRI海面风速反演

风云三号D星(FY-3D)微波成像仪(MWRI)L1级亮温数据可用于全球海面风速反演,本文讨论了在晴空区和云区使用多元线性统计回归模型和机器学习模型反演海面风速的情况,在晴空区将4 d测试集分别放入多元线性统计回归模型,采用随机森林(RF),支持向量回归(SVR),卷积神经网络(CNN)和Stacking融合(SF)模型对海面风速进行反演,最优的均方根误差(RMSE)分别为1.56、1.31、1.24、1.29和1.27 m/s;在云区2 d测试集上的最优RMSE分别为2.12、1.98、1.87、1.89和1.89 m/s。为了进一步验证晴空区海面风速反演的可靠性,选取美国国家浮标数据中心(NDBC)实测的浮标风速对海面反演风速进行验证,CNN反演风速与NDBC实测风速的RMSE为0.74 m/s,决定系数(R^(2))为0.80;SF反演风速与NDBC实测风速的RMSE为0.85 m/s,R^(2)为0.74。结果证实了通过机器学习模型能够很好地完成FY-3D MWRI亮温反演全球海面风速的任务。

基于图卷积的无监督跨模态哈希检索算法

为解决当前无监督跨模态哈希检索在全局相似性矩阵构建和异构数据语义信息融合中存在的困难,提出一种基于图卷积的无监督跨模态哈希检索算法(GCUH)。采用分层次聚合的方式,将各个模态的相似性结构编码到全局相似性矩阵中,获得跨模态的成对相似性信息来指导学习。使用图卷积模块融合跨模态信息,消除邻居结构中的噪声干扰,形成完备的跨模态表征,提出两种相似性保持的损失函数约束哈希码的一致性。与基线模型相比,GCUH在NUS-WIDE数据集上使用64位哈希码执行文本检索图片任务的检索精度提升了6.3%。

基于深度学习的代码生成方法研究进展

关注根据自然语言描述生成相关代码片段的代码生成(code generation)任务.在软件开发过程中,开发人员常常会面临两种情形.一种是通用功能的实现,需要开发人员编写大量重复且技术含量较低的代码;另一种是依赖于特定任务要求,需要开发人员查询文档或使用其他工具才能完成的代码编写工作.代码生成作为最直接辅助开发人员完成编码的工作受到学术界和工业界的广泛关注.让机器理解用户需求,自行完成程序编写也一直是软件工程领域重点关注的问题之一.近年来,随着深度学习在软件工程领域任务中的不断发展,尤其是预训练模型的引入使得代码生成任务取得了十分优异的性能.系统梳理当前基于深度学习的代码生成相关工作,并将目前基于深度学习的代码生成方法分为3类:基于代码特征的方法、结合检索的方法以及结合后处理的方法.第1类是指使用深度学习算法利用代码特征进行代码生成的方法,第2类和第3类方法依托于第1类方法进行改进.依次对每一类方法的已有研究成果进行系统的梳理、分析与总结.除此之外,汇总并分析已有的代码生成工作中常用的语料库与评估方法,以便于后续研究进行实验设计.最后,对代码生成方法研究进展进行总结,并针对未来值得关注的研究方向进行展望.

基于多模态融合的开放域三维模型检索算法

为了满足开放域下海量新增模型数据的管理和检索需求,提出开放域三维模型检索算法,可以有效地利用多模态信息的语义一致性.借助无监督算法探寻未知样本间的类别信息,利用该类别信息实现网络模型的参数优化,使得网络模型在开放域条件下具有更好的模型表征性能及检索结果.提出基于Transformer结构的层级化多模态信息融合模型,有效地剔除了多模态间的冗余信息,得到鲁棒性更强的模型表征向量.在数据集ModelNet40上进行实验,通过与其他典型算法的对比实验可知,所提方法在mAP指标上优于所有的对比方法,验证了该方法在检索性能提升上的有效性.

机器学习在水环境典型水质参数遥感反演中的应用

遥感技术是一种可用于大面积水体长时序监测的有效方法,研究综述了机器学习方法在几种典型水质参数遥感反演中的应用。首先,简述了水质反演中几种常用机器学习算法的原理和优缺点。随后,介绍了机器学习模型反演叶绿素a、悬浮物质、溶解性有机质、磷和氮5种参数的研究进展,并进一步分析了面临的问题和挑战。在此基础之上,进行了总结和展望:(1)机器学习模型的反演效果普遍优于传统经验公式和半经验模型;(2)具有量化反演不确定性能力的机器学习模型(如混合密度网络和贝叶斯神经网络等),提供了更为全面和可靠的预测;(3)基于全球性大样本数据集构建的机器学习模型具有较好的泛化能力,存在产品化潜力;(4)未来的工作应主要集中于不确定性估计算法和迁移学习的推广、大气校正算法的评估,以及水环境遥感大数据的发展等。

Pobe:一种基于生成式模型的分布外文本检测方法

对于安全可靠的机器学习系统,具备检测训练集分布外(out-of-distribution,OOD)样本的能力十分必要.基于似然的生成式模型由于训练时不需要样本标签,是一类非常受欢迎的OOD检测方法.然而,近期研究表明通过似然来检测OOD样本往往会失效,并且失效原因与解决方案的探究仍较少,尤其是对于文本数据.从模型层面和数据层面分析文本上失效的原因:生成式模型的泛化性不足和文本先验概率的偏差.在此基础上,提出一种新的OOD文本检测方法Pobe.针对生成式模型泛化性不足的问题,引入KNN检索的方式,来提升模型的泛化性.针对文本先验概率偏差的问题,设计一种偏差校准策略,借助预训练语言模型改善概率偏差对OOD检测的影响,并通过贝叶斯定理证明策略的合理性.通过在广泛的数据集上进行实验,证明所提方法的有效性,其中,在8个数据集上的平均AUROC值超过99%,FPR95值低于1%.

基于人因工程学的高校图书馆自助检索机设计研究

为了增强目前高校图书馆自助检索机的使用效率,提升用户在图书检索过程中的使用体验。文章通过对高校图书馆自助检索机进行调研,发现其存在设备尺寸不符合人体工学、操作方式不符合人性化、界面设计不符合用户习惯等问题。文章充分考虑大学生的身体变化,运用人因工程学理论进行再设计。给出具体的高校图书馆自助检索机尺寸参数和界面设计策略。该设计研究可以为各大高校图书馆自助检索机设计的参考。

数字化精确堆取煤技术研究

为保证煤场堆取料的大型设备斗轮机的高效运行,并确保煤粉混合均匀,提高电厂的锅炉燃烧效率,需要实现电厂用煤的数字智能化精确堆取。煤场数字智能化是当前智慧电厂亟需解决的重要技术问题,能够为煤场管理及效能提升提供完整的技术支撑,也是煤场无人值守技术应用的前提条件。数字化煤场技术包括自动实现煤炭的分类存放与精确提取,煤场数据的实时监控与传输,以及对所有操作进行同步控制与统一优化。为进行数字化煤场中的精确堆取煤技术分析研究,在满足普遍适应性、鲁棒性及实用性前提下,提出了基于视觉的定位技术及精度提升的优化算法,探讨了将基于视觉的定位方法应用于斗轮机的具体条件和改进策略,以便合理调整好成本与精度需求之间的关系,实现满足安全要求的情况下,最大化降低成本。

基于推荐列表的缺陷文件识别

缺陷定位是缺陷修复的关键步骤,同时也是一项繁琐的软件活动。现有的静态缺陷定位技术通常将缺陷定位视为一个检索任务,即为每个缺陷报告生成一份按照程序实体与缺陷相关度降序排列的可疑文件推荐列表。然而,开发人员仍需人工一一审查从而找到真正有缺陷的文件,这增加了定位的时间和成本。为解决这个问题,提出了一个相应的解决方案。首先运行主流的基于信息检索的静态缺陷定位技术来获得一个初始的可疑文件推荐列表;然后依据问题特性提出3类领域特征,并基于这3类特征构建一个机器学习模型,尝试从列表中识别出真正有缺陷(Truly Buggy)的源代码文件。在4个开源项目(Zoo-Keeper,OpenJPA,Tomcat,AspectJ)的2558个bug上进行了实验,结果表明,在最初可疑文件推荐列表上可以获得72.6%~80.7%的真正有缺陷的文件预测准确率。同时探究了3类特征子集及各个特征在预测真正有缺陷的文件上的重要性,发现缺陷报告与源代码的关系特征更重要。