基于多源信息的全局滤波器剪枝

针对现有神经网络剪枝方法未全面评估滤波器的重要性以及跨层滤波器的重要性间存在一定差异的问题,提出了一种基于多源信息的全局滤波器剪枝算法,建立了特征和权重信息间的连接.首先,根据特征信息较为丰富和权重信息受数据噪音影响低的特点,分别以特征间相关性和权重熵来评估滤波器的相对和绝对重要性.然后,将每层中不同压缩比例的滤波器看作一个整体,评估其对模型的全局重要性,按照压缩需求跨层剪掉模型中最不重要的部分.最后,采用知识蒸馏的方式来恢复剪枝后模型的精度,不依赖其他数据集就能完成模型的压缩与微调.为了验证所提方法的适用性,针对DeepLabV3、DABNet和U-Net网络在三个语义分割数据集上进行了大量的实验.也针对多种深度的ResNet网络在图像分类数据集上进行了验证.实验结果表明,通过多源信息可以更精确的评估单层中滤波器的重要性,通过全局重要性来指导跨层剪枝可以使模型的关键信息损失降到最低.     

基于Logistic回归的数据分类问题研究

数据分类是数据科学的一个重要研究方向,Logistic回归是最基本的分类算法之一.线性回归和Logistic回归都属于线性模型,本文介绍了两者的联系,详细阐述了模型的目标函数和参数训练过程.在经典的模式识别数据集Iris上,应用Logistic回归模型基于部分特征和全部特征维度进行了建模和分类预测.实验结果表明,适合比例的训练集和测试集切分,较高权重特征组合的建模可以获得较高的分类准确率.     

基于无预训练卷积神经网络的红外车辆目标检测

为解决基于卷积神经网络的目标检测算法对预训练权重的过度依赖,特别是数据稀缺条件下的红外场景目标检测,提出了融入注意力模块来缓解不进行预训练所带来的检测性能下降的方法。本文基于YOLO v3算法,在网络结构中融入模仿人类注意力机制的SE和CBAM模块,对提取的特征进行通道层面和空间层面的重标定。根据特征的重要程度,自适应地赋予不同权重,最终提升检测精度。在构建的红外车辆目标数据集上,注意力模块能够显著提升无预训练卷积神经网络的检测精度,融入了CBAM模块的网络检测精度为86.3 mAP。实验结果证明了注意力模块能够提升网络的特征提取能力,使网络摆脱对预训练权重的过度依赖。     

基于多模型融合的航空电子产品故障预测方法

针对复杂机载环境应力条件下航空电子产品故障预测所面临的退化趋势差异大、训练数据样本量小等问题,提出了一种改进长短期记忆(Long Short-Term Memory, LSTM)神经网络模型与集成学习框架相结合的故障预测方法,以满足现代综合航空电子系统智能调度管理与自主维护保障的需求。该方法在LSTM模型中引入Dropout机制,构建基于不同历史数据集的差异性LSTM模型组,以解决故障预测时序信息记忆问题与小样本条件下数据驱动模型训练过拟合问题;采用Adaboosting算法计算模型权重,并基于实时数据动态调整,以滤除复杂机载环境应力引入的预测误差,解决多模型融合的性能差异问题。最后,采用NASA公开的锂电池退化数据集进行仿真验证,实验结果表明,相较于传统BP神经网络、经典LSTM和LSTM基模型,该方法具有更高的趋势拟合度和预测精度。     

综合项目评分和属性的个性化推荐算法

针对传统协同过滤算法存在的数据稀疏性和冷启动问题,提出了一种综合项目评分和属性的个性化推荐算法.该算法在衡量项目相似性时,同时考虑用户评分和项目属性特征,并根据评分数据的实际稀疏情况动态调整两者的影响权重;预测评分时,利用用户对项目属性的偏好度来衡量其对未评分邻居项的喜好程度,并产生最终推荐.基于MovieLens数据集进行的实验结果表明,该算法使得最近邻的确定更加准确,系统推荐质量明显改善.     

基于粒子群优化和支持向量机的电力负荷预测

提出支持向量机的粒子群优化算法的用电量预测方法.其中,采用粒子群优化算法选取较优的支持向量机训练参数组合.以江西省2008年7月～10月的用电量数据以及相关特征数据作为实验数据,实验结果表明该算法电量负荷预测精度高于BP神经网络.     

基于随机森林的流处理检查点性能预测

物联网(IoT)的发展引起流数据在数据量和数据类型两方面不断增长。由于实时处理场景的不断增加和基于经验知识的配置策略存在缺陷,流处理检查点配置策略面临着巨大的挑战,如费事费力,易导致系统异常等。为解决这些挑战,该文提出基于回归算法的检查点性能预测方法。该方法首先分析了影响检查点性能的6种特征,然后将训练集的特征向量输入到随机森林回归算法中进行训练,最后,使用训练好的算法对测试数据集进行预测。实验结果表明,与其它机器学习算法相比,随机森林回归算法在CPU密集型基准测试,内存密集型基准测试和网络密集型基准测试上针对检查点性能的预测具有误差低,准确率高和运行高效的优点。     

基于共享知识模型的跨领域推荐算法

互联网的普及使得大量信息不断累积,推荐系统作为解决信息过载的有效手段,能够帮助人们迅速准确地筛选出感兴趣的内容.但是由于用户项目评分数据过于稀疏,新用户或新商品存在"冷启动"问题,使得传统的推荐算法计算复杂性过高、准确性较低.考虑到用户会在互联网不同领域使用各类应用,在不同领域积累了大量行为数据和评价信息.而从用户群体的角度来说,在不同领域间存在着用户群体的偏好相似性,因此如果通过在不同领域中共享代表偏好的知识模型,将有助于提升在新领域推荐的准确性,解决冷启动问题.本文提出了基于共享知识模型的跨领域推荐算法SKP （Sharing Knowledge Pattern）,通过对各个领域中用户-项目的评分矩阵分解,得到用户的潜在特征矩阵和项目的潜在特征矩阵,对用户和项目的潜在特征分别聚类,得到了用户分组对项目分组的评分知识模型,最终利用目标领域的个性知识模型和各个领域的共性知识模型来得出推荐结果.本文对三个不同领域的数据集进行了分析和划分,并在物理集群环境下进行了实验.结果表明,通过利用数据稠密的辅助领域数据,本文提出的SKP算法与已有的单领域算法、跨领域算法相比,具有更高的准确率和更低的RMSE值.     

基于跨域学习的单样本目标检测方法

针对单样本目标检测样本量较少的问题，提出了一种基于跨域学习的方法。该方法从数据增强的角度出发，增加其他域的数据集作为辅助，增强网络学习能力，同时为解决不同域间存在差异的问题，提出了一种基于图片尺度和实例尺度的跨域学习算法，分别对输入的图片特征与检测网络的候选特征增加域分类器模型，用于增强网络对跨域数据的背景和目标的域适应能力。在两个不同的跨域场景进行实验，其中在PASCAL VOC数据集上与目前主流的单样本目标检测算法进行比较，超过目前最好算法2.8个百分点，从而证明了本文方法可以有效提高单样本目标的检测性能。     

引入特征交互的红外与可见光图像自适应融合北大核心CSCD

在图像融合领域,现有的基于卷积神经网络(CNN)或Transformer架构的方法存在两个局限性:首先,浅层纹理特征与深层语义特征之间无法有效聚合;其次,红外与可见光特征的权重比例无法自适应变化。本文提出一种引入特征交互的红外与可见光图像自适应融合方法。首先,构建一种基于Transformer的特征交互模块,聚合跨尺度特征信息,增强特征表达能力。其次,设计一种融合模块,自适应地调整特征权重比例。所提出的融合方法通过两阶段训练策略完成。第一个阶段,应用创新的特征交互概念训练编码器,增强特征表达,重建特征图像。第二个阶段,基于设计的权重自适应调整模块训练红外与可见光特征融合任务。公开数据集的实验结果表明,与现有方法相比,本方法在主观和客观的评价方面均优于其他典型方法。     

Cross-project software defect prediction based onmulti-source data sets

Cross-project defect prediction (CPDP) uses one or more source projects to build a defect prediction model and applies the model to the target project. There is usually a big difference between the data distribution of the source project and the target project, which makes it difficult to construct an effective defect prediction model. In order to alleviate the problem of negative migration between the source project and the target project in CPDP, this paper proposes an integrated transfer adaptive boosting (TrAdaBoost) algorithm based on multi-source data sets (MSITrA). The algorithm uses an existing two-stage data filtering algorithm to obtain source project data related to the target project from multiple source items, and then uses the integrated TrAdaBoost algorithm proposed in the paper to build a CPDP model. The experimental results of Promise's 15 public data sets show that: 1) The cross-project software defect prediction model proposed in this paper has better performance in all tested CPDP methods; 2) In the within-project software defect prediction (WPDP) experiment, the proposed CPDP method has achieved the better experimental results than the tested WPDP method.     

数据驱动的软件缺陷预测研究综述

数据驱动的软件缺陷预测是提高软件测试效率、保证软件可靠性的重要途径之一,近几年已成为实证软件工程的研究热点.首先介绍了数据驱动软件缺陷预测的研究背景;然后总结了已有软件缺陷数据属性度量方法的特点,并按照软件开发中缺陷预测的使用场景,以数据来源为主线从基于版本内数据、跨版本数据和跨项目数据实现缺陷预测三个方面对近10年（2005~2015）已有的研究工作进行分类归纳和比较;最后对该领域未来的研究趋势进行了展望.     

Evaluating the Relative Performance of Engineering Design Projects: A Case Study Using Data Envelopment Analysis

This paper presents a case study of how Data Envelopment Analysis (DEA) was applied to generate objective cross-project comparisons of project duration within an engineering department of the Belgian Armed Forces. To date, DEA has been applied to study projects within certain domains (e.g., software and R&D); however, DEA has not been proposed as a general project evaluation tool within the project management literature. In this case study, we demonstrate how DEA fills a gap not addressed by commonly applied project evaluation methods (such as earned value management) by allowing the objective comparison of projects on actual measures, such as duration and cost, by explicitly considering differences in key input characteristics across these projects. Thus, DEA can overcome the paradigm of project uniqueness and facilitate cross-project learning. We describe how DEA allowed the department to gain new insight about the impact of changes to its engineering design process (redesigned based on ISO 15288), creating a performance index that simultaneously considers project duration and key input variables that determine project duration. We conclude with directions for future research on the application of DEA as a project evaluation tool for project managers, program office managers, and other decision-makers in project-based organizations.     

MTMVC: Semi-supervised 3D hand pose estimation using multi-task and multi-view consistency

The high performance of state-of-the-art deep learning methods for 3D hand pose estimation heavily depends on a large annotated training set. However, it is difficult and time-consuming to obtain the annotations for 3D hand poses. To leverage unannotated images to reduce the annotation cost, we propose a semi-supervised method based on Multi-Task and Multi-View Consistency (MTMVC) for hand pose estimation. First, we obtain the joints based on heatmap prediction and coordinate regression parallelly and encourage their consistency. Second, we introduce multi-view consistency to encourage the predicted poses to be rotation-invariant. Thirdly, to make the network pay more attention to the hand region, we propose a spatially weighted consistency. Experiments on four public datasets showed that our proposed MTMVC outperformed existing semi-supervised hand pose estimation methods, and by only using half of the annotations, the accuracy of our method was comparable to those of several state-of-the-art fully supervised methods.     

Semi-supervised Software Defect Prediction Using Task-Driven Dictionary Learning

We present a semi-supervised approach for software defect prediction.The proposed method is designed to address the special problematic characteristics of software defect datasets,namely,lack of labeled samples and class-imbalanced data.To alleviate these problems,the proposed method features the following components.Being a semi-supervised approach,it exploits the wealth of unlabeled samples in software systems by evaluating the confidence probability of the predicted labels,for each unlabeled sample.And we propose to jointly optimize the classifier parameters and the dictionary by a task-driven formulation,to ensure that the learned features (sparse code) are optimal for the trained classifier.Finally,during the dictionary learning process we take the different misclassification costs into consideration to improve the prediction performance.Experimental results demonstrate that our method outperforms several representative stateof-the-art defect prediction methods.     

Assembling three one-camera images for three-camera intersection classification

Determining whether an autonomous self-driving agent is in the middle of an intersection can be extremely difficult when relying on visual input taken from a single camera. In such a problem setting, a wider range of views is essential, which drives us to use three cameras positioned in the front, left, and right of an agent for better intersection recognition. However, collecting adequate training data with three cameras poses several practical difficulties; hence, we propose using data collected from one camera to train a three-camera model, which would enable us to more easily compile a variety of training data to endow our model with improved generalizability. In this work, we provide three separate fusion methods (feature, early, and late) of combining the information from three cameras. Extensive pedestrian-view intersection classification experiments show that our feature fusion model provides an area under the curve and F1-score of 82.00 and 46.48, respectively, which considerably outperforms contemporary three- and one-camera models.     

A data reduction scheme for active authentication of legitimate smartphone owner using informative apps ranking

Smartphones are vulnerable to fraudulent use despite having strong authentication mechanisms. Active authentication based on behavioral biometrics is a solution to protect the privacy of data in smart devices. Machine-learning-based frameworks are effective for active authentication. However, the success of any machine-learning-based techniques depends highly on the relevancy of the data in hand for training. In addition, the training time should be very efficient. Keeping in view both issues, we’ve explored a novel fraudulent user detection method based solely on the app usage patterns of legitimate users. We hypothesized that every user has a unique pattern hidden in his/her usage of apps. Motivated by this observation, we’ve designed a way to obtain training data, which can be used by any machine learning model for effective authentication. To achieve better accuracy with reduced training time, we removed data instances related to any specific user from the training samples which did not contain any apps from the user-specific priority list. An information theoretic app ranking scheme was used to prepare a user-targeted apps priority list. Predictability of each instance related to a candidate app was calculated by using a knockout approach. Finally, a weighted rank was calculated for each app specific to every user. Instances with low ranked apps were removed to derive the reduced training set. Two datasets as well as seven classifiers for experimentation revealed that our reduced training data significantly lowered the prediction error rates in the context of classifying the legitimate user of a smartphone.     

Defect detection method of magnetic disk image based on improved convolutional neural network

This paper proposes an efficient method for defect detection of magnetic disk image based on improved convolutional neural network. We build a model named DiskNet on the basis of VGGNet-19, in which the optimal activation function is selected predictively through a weighted probability learning curve model (WP-Model). First, we use Markov Chain Monte Carlo (MCMC) to infer the predicted value and determine prediction probability. Then, the evaluation point (EP) is determined by the effective information of training curve. In the process of DiskNet training, when the prediction probability is higher than the threshold, the neural network will select the current activation function. If the training epochs exceed the EP and the threshold is not reached, the original activation function will be used. The experimental results show that the accuracy of the proposed method in detecting defects on the magnetic disk image data set is 96.9%.     

基于轻量级注意机制的人脸检测算法

提出一个新的基于轻量级注意力机制的网络框架。在YOLOv3主干网络的基础上,使用深度卷积和点卷积代替标准卷积设计特征提取网络,加快模型的训练,提高检测的速度,然后引入注意力机制模块进行模型速度和精度的权衡,最后通过增加多尺度提取更多网络层的特征信息,同时使用K-means++聚类算法进一步优化网络参数。实验结果表明,该方法可以显著提高人脸检测模型的性能,在Wider Face数据集上可以达到94.08%的准确率和83.97%的召回率,且平均检测时间只需0.022 s,相比原始YOLOv3算法提高了4.45倍。     

Multimedia annotation via semi-supervised shared-subspace feature selection

With the rapid development of social network and computer technologies, we always confront with high-dimensional multimedia data. It is time-consuming and unrealistic to organize such a large amount of data. Most existing methods are not appropriate for large-scale data due to their dependence of Laplacian matrix on training data. Normally, a given multimedia sample is usually associated with multiple labels, which are inherently correlated to each other. Although traditional methods could solve this problem by translating it into several single-label problems, they ignore the correlation among different labels. In this paper, we propose a novel semi-supervised feature selection method and apply it to the multimedia annotation. Both labeled and unlabeled samples are sufficiently utilized without the need of graph construction, and the shared information between multiple labels is simultaneously uncovered. We apply the proposed algorithm to both web page and image annotation. Experimental results demonstrate the effectiveness of our method.