Search

Homography estimation is a basic and important step in many computer vision tasks. Traditional homography estimation methods are based on feature point matching, which are difficult to work in weak texture images. Deep learning has been applied to homography estimation to improve its robustness, but the existing methods do not consider the multi-scale problem caused by object scale differences, resulting in limited accuracy. To solve the above problems, this paper proposes a multi-scale residual network for homography estimation. The network can extract the multi-scale feature of the image, and used the Multi-Scale Feature Fusion Module to effectively fuse the features. In addition, it further reduced the difficulty of network optimization by estimating the four-corner normalized offset. Experiments on MS-COCO dataset showed that the average corner error of this method was only 0.788 pixels, which achieved sub-pixel accuracy, and can maintain high accuracy in 99% of cases. Due to the comprehensive utilization of multi-scale features and easier to optimize, this method had significantly improved accuracy and stronger robustness. [ABSTRACT FROM AUTHOR]

To achieve high efficiency with limitation of energy consumption is a key problem in the system of UAVs. In existing UAV task allocation methods, the relevance between tasks and UAV cruise directions is neglected, which may further influence the energy consumption and delay of task accomplishment. In view of this, this paper proposed an UAV task allocation method based on the relevance analysis between tasks and cruise directions. This method includes two phases: task screening phase and conflict resolution phase based on consensus. In the first phase, this method selects out tasks without turning back for an UAV according to angles between the cruise direction of the UAV and directions of these tasks, then designs an algorithm to further select candidate tasks before interaction from the tasks without turning back according to their energy consumption utility parameters and time urgency parameters. In the second phase, this method solves tasks conflicts between UAVs after exchanging of their candidate tasks according to energy consumption utility parameters and time delay parameters of these tasks in different cruise directions of these UAVs. Simulation results verify that the proposed method can achieve lower average task energy consumption and average task delay. [ABSTRACT FROM AUTHOR]

Aspect-level sentiment analysis aims to determine the sentiment polarity towards specific aspect in reviews. However, little research has been done on the influence of complex sentences on sentiment classification. Based on this, this paper proposed an aspect sentiment classification model based on Bert and Self-attention network with relative position. Firstly, it used the dynamic weighted sampling method to balance the rare contrastive sentences, so that the model can learn more contrastive sentence feature information. Then, it jointly trained the feature representations extracted by double-head self-attention network with relative position and the feature representations obtained by the Pre-trained model with absolute position. Finally, it used the label smoothing regularization technology to stabilize the model to identify the neutral samples. It tested this model on Sub Task 2 in SemEval 2014 task, and improved both accuracy and Macro-F1 indicators of the two datasets. The experimental results show that the effectiveness of the proposed model for contrastive sentences classification, and also yield improvements in the whole test set over other benchmark models. [ABSTRACT FROM AUTHOR]

Due to the overlap of time and space coverage between tasks in mobile crowd sensing, repeated data collection may happen and cause data redundancy problem. In view of this, this paper proposed a task allocation method to reduce data redundancy within and between tasks. Firstly, this method designed a trajectory sequence prediction model based on the long short-term memory (LSTM) neural network, to predict trajectory sequences of task participants within subdivided spatial-temporal units. Then based on the trajectory prediction results, the method proposed an optimization model to minimize data redundancy. Specifically, the optimization model constrained the data redundancy within a single task by minimizing the data redundancy metric in each spatial-temporal unit, and limited the data redundancy between multiple tasks by maximizing the reuse of the sensing data of each task participant in a spatial-temporal unit. Experimental results verify that the proposed task allocation method can effectively reduce the data redundancy within and between tasks. [ABSTRACT FROM AUTHOR]

At present, most link-prediction based models on anomaly detection in social networks lack the ability to consider the influence of abnormal nodes evolution. With the limitation of network scale and complexity, the detection efficiency of traditional models is generally low. To address these issues, this paper proposed an anomaly detection method based on the spatial scale coarse granulation and weighting mechanism on abnormal nodes. Firstly, the method introduced a cohesive community discovery algorithm, Louvain algorithm, to coarsely granulate process to streamline network. Subsequently, it identified abnormal nodes with different evolution behaviors in the processed network following the quantification of abnormal evolution process. Finally, it applied the link prediction method combined with a weighting mechanism of abnormal nodes for final abnormal detection. Compared with different LinkEvent-based strategy adjustment algorithms and NESO_ED method on three real social network data sets VAST, Email-EU ( dept1 and dept2 ) and Enron, the proposed method outperforms other state-of-the-art methods, can take into account both stability and sensitivity on anomaly detection tasks and more reasonably describe the network evolution process. [ABSTRACT FROM AUTHOR]

The existing heterogeneous network embedding methods not only ignore heterogeneous edges in networks and their different effects on node embedding, but also do not consider the integration of network structure and node attributes. This paper proposed a novel heterogeneous network emhedding method with attributes information fusion, which named SHANE. It applied the sequence to sequence ( seq2seq) model to the sub-graphs divided according to edge types, seamlessly integrated the structure information and attributes information of nodes, and captured the high-order semantic information of nodes simultaneously. Experiments show that SHANE can achieve relatively significant results in link prediction tasks in two different types of datasets. [ABSTRACT FROM AUTHOR]

Since the prototype network cannot make full use of samples' semantic information, it 's difficult for model to fully excavate the transferable features in training data. As the result, the model underperforms when it is facing unfamiliar data in a new domain. To this end, this paper made improvements from three perspectives: model structure, encoding network, and metric network, and proposed a multiple-task prototypical network MTPN. In terms of model structure, on the basis of the prototype network ' s metric task, it added an auxiliary classification task to constrain the training target, which could improve the semantic feature extraction ability of the model. By using multi-task learning, model obtained a semantic representation that was more relevant to the auxiliary task. In order to improve feature transferability, this paper also proposed the LF-Transformer encoder which used hierarchical attention to fuse the underlying general encoding information. The metric network used the BiGRU-based class prototype generator to make the class prototype more representative and the distance measurement more accurate. Experiments show that MTPN achieves an accuracy of 91.62% in the sentiment classification task with few samples, which is 3.5 % higher than the existing best model. For samples in new field that have not appeared in training state, by using 5 references, the model can still obtain a classification accuracy of more than 90% on query samples. [ABSTRACT FROM AUTHOR]

In order to fully mine nodes' features and better integrate these features simultaneously in the heterogeneous information network, this paper proposed a AMF L& GRec Firstly, AMFL & GRec used the Leader Rank algorithm to extract the target node' global sequence, and used a meta-path-bas ed heterogeneous information network embedding model to extract the node' local sequence,and used the skip-gram model to learn the node' global and local features. And then it used the self-attention mechanism to learn the preference of the target nodes' local and global features to obtain the feature representation of the target node in a single meta-path. Secondly, it used the self-attention mechanism to fuse the representation of the same node under different meta-paths to obtain the final feature representation. Finally, it utilized a multi-layer perceptron to achieve the task of rating prediction. This paper conducted a large number of experiments on two real datasets. The experimental results verify that the AMFL & GRec algorithm can not only capture the micro(local) structure of densely connected nodes, but also capture the global structure of the node in the network, and finally obtain nodes' overall (local +global) characteristics. At the same time, the experimental results also prove that the AMFL & GRec's rating prediction performance is better than the baselines. It proves that in the heterogeneous information network utilizing the self-attention mechanism to consider the nodes' preferences for local and global features and meta-paths can improve the accuracy of rating prediction. [ABSTRACT FROM AUTHOR]

In attribute network, the attribute information associated with nodes is essential to improve the performance of network embedding tasks. Nevertheless, network is a graph structure, in which nodes not only contain attribute information but also embrace the rich structural information. In order to make full use of the structural information, firstly, this paper defined influential node characteristics, spatial relationships, and constructed similarity matrix based on the definition of link prediction. Then it mapped correlation similarity vector associated with nodes in the binary group to the relationship space of the adjacency matrix, so as to maintain the node vector matrix structure information feature. Based on the definition of normalized graph Laplacian, it fused the attribute information and label feature and integrated the above three kinds of information into an optimization framework. Finally, it inferenced the projection matrix by calculating the local maximum value through a second order derivative. Experimental results indicate that the proposed algorithm can effectively utilize information of the adjacency structure with the binary group of nodes, and compared with other benchmark network embedding algorithms, it also can achieve better results on the node classification task. [ABSTRACT FROM AUTHOR]

Intelligent question answering based on reading comprehension refers to letting computers read and comprehend texts like humans, extracts the text information and answers corresponding questions. The pre-training model RoBERTa-wwm-ext uses the extracted original fragments as the answers to the questions, but this method can’t solve the two situations that the answer fragments don’t exist in the original text or need to reply to the original text after summarizing. The pre-training model is used for generative model training, which can solve the problems that need to summarize the original text to a certain extent. Therefore, this paper improved the method of only using RoBERTa-wwm-ext model to extract answers. On this basis, it integrated the generative question answering model based on RAG model to answer questions that could not be handled by Roberta-wwm-ext and other extraction models. At the same time, this paper absorbed the advantages of PGN model, improved RAG model, and obtained RPGN sub model, which could make better use of reading and understanding articles to generate reasonable answers. Therefore, this paper proposed a fusion model of RPR（RAG, PGN, RoBERTa-wwm-ext）,which could be used to deal with both extractive question task and generative question answering task at the same time. [ABSTRACT FROM AUTHOR]

Recently, graph neural networks have received widespread attention due to their rich representation and reasoning capabilities. To best knowledge, current research mainly focuses on amending the convolutional strategy and network structure for higher performance, so the performance will be inevitably constrained by the limitations of the single model. Inspired by the idea of ​​ensemble learning, this paper innovatively proposed an ensemble learning framework for graph neural network (EL-GNN).Unlike regular text and images, graph data not only had features but also had rich topology information. Therefore, EL -GNN additionally supplemented the structure information during the ensemble stage rather than merely integrating the prediction results of independent classifiers. Besides, this paper further revised the ensemble strategy through reconstructing a feature-level similarity graph for subsequent assembling, which balances the feature and structure information on the basis of the assumptions of those nodes with the similar feature or easy reachability of high probability to share the same labels .The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. [ABSTRACT FROM AUTHOR]

Considering that in the multi-robot patrol task, it is a common phenomenon that the importance of the nodes to be visited is different. In view of the fact that most patrolling algorithms don't consider the difference in the importance of nodes, the idle time of all nodes tends to be the same, resulting in important nodes access frequency ordinary nodes and excessive access problems. This paper proposed a distributed patrolling strategy based on nodes' importance to optimize node access frequency and reduce the global average idle time. The robot calculated the idle time and importance of surrounding nodes, online decided the target node, estimated the time of reaching the target node, and informed nearby companions of the target node and estimated time of arrival. It increased the importance of the edge node in order to prevent some nodes from being over-visited, when the access frequency of the edge node was lower than the minimum access frequency, so that the robot could visit the node as soon as possible. Finally, this paper analyzed the influence of factors on the completion of continuous patrolling tasks of robots through simulation on the number of robots, changing environment, the importance of nodes and other factors. The results show that the number of visits to the most important nodes has increased significantly. Moreover, under the premise of the same patrolling environment and number of robots, the algorithm has fewer outliers and smaller average idle time, and the performance of multi-robot continuous patrolling is better. [ABSTRACT FROM AUTHOR]

Traditional graph regularization methods use Euclidean distance to measure the similarity of sample space, and can not accurately preserve the neighborhood information of complex data sets, which easily lead to the degradation of the generalization performance of the model in complex shape data and non convex data sets. This paper proposed an improved graph regularization algorithm, which used isometric feature mapping to preserve the neighborhood information of the sample space and help the model learn manifold. The simultaneous used of KL constraints further smoothed the external structure of the data representation, thereby capturing more sparse and advanced feature representations. Experimental results on MNIST and YaleB datasets show that compared with several popular feature extraction algorithms, the proposed algorithm can extract more meaningful and robust features. It has advantages in classification and clustering tasks, and also has better anti-interference capability. [ABSTRACT FROM AUTHOR]

Joint extraction of entities and relations from unstructured text is an important task in information extraction. The existing methods have achieved considerable performance, but are still subject to some inherent limitations, such as error propagation, redundancy of relation prediction, inability to solve the problem of relations overlap, etc. For this reason, this paper proposed a joint entity relationship extraction model BSGB( BiLSTM + SDA-GAT + BiGCN) based on graph neural network. BSGB was a two-stage predicting process. The first stage of this model extended the semantic dependency analysis to the semantic dependency graph, and proposed a graph attention network to integrate the semantic dependency graph(SDA-GAT). By stacking BiLSTM and SDA-GAT, it extracted sentence sequence and local dependent features, and performed entity span detection and preliminary relationship prediction. In the second stage, it constructed the relation-weighted GCN, which further modeled the interaction between entities and relations, and completed the final extraction of entity-relationship triples. The experimental results on the NYT dataset show that the F1 value of this model reaches 67 .1 %, which is 5. 2% higher than the baseline model in this dataset, and the prediction of the overlap relation is also significantly improved. [ABSTRACT FROM AUTHOR]

Cloud segmentation is a key task in applications of Landsat8 remote sensing images. At present, most existing deep learning based approaches usually adopt the U-type coder-encoder network structure. Although this structure can make use of spatial location information from the coding end effectively, the whole network has too many parameters and cause high computation cost. Meanwhile, the coding end only adopts simple convolution and subsampling operation, which makes it impossible to obtain high-quality contextual semantic feature information effectively. In view of the above situation, this paper proposed a new lightweight U-Net model for cloud segmentation. The whole model adopted a new U-shaped codec structure that jumped and connected the information between the shallow layer and the middle layer at the coding end, and also embedded novel optimization blocks consist of grouping convolution and attention mechanism in coding end. It also built context-semantic fusion connection to link corresponding layers in coding and decoding end. Experimental results on the common benchmark dataset 38-Cloud illustrates, compared with other mainstream cloud segmentation networks, the proposed new U-Net can achieve better segmentation precision with less model parameters. [ABSTRACT FROM AUTHOR]

Ancient murals have high artistic value and rich content.It is one of the problems for researchers to accurately classify the types of murals.Traditional mural classification tasks are arduous and require experienced researchers to complete.The existing image classification algorithms are no longer suitable for classifying mural images with strong background noise.To solve the above problems,this paper proposed a new multi-channel separable network model.Using the GoogLeNet network model as the basic framework,it used a small convolution kernel to extract the background features of the murals in a shallow layer,and then separated the larger convolution kernels such as 7×7 and 3×3 into 7×1,1×7 and 3×1,1×3 and other smaller convolution kernels for extraction the important deep-level feature information of murals.It used activation scaling to increase the stability of the network during training,and finally classified the murals through softmax,then used a small batch stochastic gradient descent algorithm to update parameters.The precision rate,recall rate and F1 value are 88.16%,90.01% and 90.38%,respectively.Compared with mainstream classification algorithms,it improves classification accuracy,generalization ability,and stability to a certain extent,which improves the efficiency of mural classification. [ABSTRACT FROM AUTHOR]

As for the special architecture of the stack processor, to improve the real-time performance, by adopting a multi-task stacking technology and the Forth self-generator technology, this paper proposed a stack-processor-based preemptive and round-robin scheduling method, which made the multi-tasks run on the Forth stack processor in a real-time manner and made up for the deficiencies of the processor in the real-time multi-task operating system. The experiments show that, compared with the current register processor-based embedded Forth real-time systems, the proposed method improves the real-time performance obviously on the three vital metrics of real-time task performance:maximum interrupt time, task context switching time, and task response time. As a result, it ensures the efficiency and security of Forth system applications, and meets the application demand for the Forth stack processor real-time multitasking operating system. [ABSTRACT FROM AUTHOR]

Aiming at the shortcomings of the existing task allocation strategies, this paper proposed a multi-task allocation strategy based on deferred acceptance in mobile crowd sensing. Firstly, based on the different needs of platform and workers, this paper proposed a method to achieve the two-side matching of tasks and workers. Secondly, considering the urgency of the task, this paper designed a hybrid deferred acceptance algorithm. Finally, in order to verify the performance of the proposed strategy, it was compared with the two baseline task allocation methods. Experimental results show that the proposed task allocation strategy is effective in improving platform utility and worker satisfaction. [ABSTRACT FROM AUTHOR]

In view of the problems of high training complexity and slow inference speed involved by the current mainstream visual question answering task which uses regional features as image representations, this paper proposed a convolutional network ( composite visionlinguistic ConvN et, CVlCN) based on composite visual language to extract the image features in visual question answering tasks. The proposed method represented image features and problem semantics into composite picture-text features through composite learning, and then calculated the attention distribution of composite picture-text features from space and channels to selectively retain visual information related to problem semantics. The experimental results show that, on the VQA-v2 dataset, the test accuracy of the proposed method on the visual question answering task is obviously improved, and the overall accuracy is 64. 4%. And the model has low computational complexity and fast inference speed. [ABSTRACT FROM AUTHOR]

Structural variation detection is an important task of genome big data analysis. Inversion is a common structural variation and is often associated with a variety of serious diseases. Therefore, obtaining the accurate position of inversion is an important task. This paper proposed a method for detecting inversions variation based on long reads and realignment. This method firstly extracted the clipped sequence of each long read, and aligned them against the genome reference sequence . Based on the original alignment information and the realignment information, this method determined the inversion signals. Finally, it clustered the inversion signals to determine the location of the inversions. The performance of this method and the other two common inversion variation detection methods were compared on two datasets. The experimental results show that this method can obtain more satisfactory detection results. [ABSTRACT FROM AUTHOR]

The path planning of automated guided vehicle (AGV) affects the efficient and smooth production in a flexible manufacturing workshop. This paper analyzed the common conflict types of multiple AGV performing transport tasks in the workshop, and then proposed a multi-AGV conflict-free path planning method based on dynamic priority strategy. This method detected the specific conflict types by time window, and considered the waiting time, the distance traveled and task urgency as the factors to quantitatively evaluate the dynamic priority, which was then utilized as the selection basis of two conflict reduction methods,i. e.,waiting and quadratic path planning. The case study indicates that when the feasible paths of each AGV are known in advance, the proposed method cannot only realize the multi-AGV conflict-free path planning, but also improve the overall efficiency of multiple AGV in completing transport tasks. [ABSTRACT FROM AUTHOR]

The detection speed of detection algorithms based on artificial design features is generally slow, and the detection accuracy needs to be improved, which can no longer meet the needs of today's industrial production. The detection technology based on deep learning cannot be deployed on resource-constrained devices because it requires a lot of computing and storage space. In response to these problems, this paper cited a channel pruning method to realize the lightweight of the YOL0v3 detection network, and obtained the pruning model Slim YOLOv3, and proposed to use Slim YOLOv3 for real-time detection tasks in industrial scenarios. The method enhanced the channel-level sparsity of the convolutional layer by applying L, regularization to the channel scaling factor, and pruned the feature channels with less information, and obtained a lightweight network model. Compared with the original model, SlimYOLOv3 gives a 60% reduction in model size and a 50% reduction in computing operations, the detection speed is 1. 7 times of the original model. It is more suitable for real-time detection of complex targets in intelligent industrial scenes. [ABSTRACT FROM AUTHOR]

Unlike long text classification, short text lacks information, and it is often difficult to obtain rich semantic features and extract complete syntactic features. Therefore, the effect of short text classification model needs to be improved . To solve this problem, this paper proposed a new model based on ResLCNN, the new model which was based on neural topic model and concatenated the features of multiple neural networks. Firstly, it extracted the topics to enrich the information of the short text. Secondly, it stored the topic information in the memory network and fused them with the sequence information as the final expression. Finally, it input the final expression into neural network with residual structure to extract local and global semantic features, and then combined the features together. The model achieves high accuracy in the Google Web search public dataset, which proves the effectiveness of the improved model in short text classification tasks. [ABSTRACT FROM AUTHOR]

This paper proposed a mobile edge computing offloading strategy based on multi-hop transmission of network connected vehicles. Through the prediction of future vehicle trajectory, it found effectively the best real-time multi-hop transmission path of vehicle network, so as to ensure that the computing task could be successfully unloaded to MEC server within the time delay requirement. The simulation results show that compared with the traditional mobile offloading strategy, the average task delay is lower, the task success rate is higher, and all aspects of performance are better than the traditional edge computing offloading strategy. Among them, the task unloading success rate increases by 10. 06% on average, and the task delay decreases by 8. 62% on average. [ABSTRACT FROM AUTHOR]

Keyword extraction plays an important role in text mining tasks . This paper aimed at proposing a keyword extraction method based on k-truss graph decomposition for short texts, named KEK. Firstly, the method applied the space vector model to present each word in the text, and constructed a text graph by taking words as nodes and the co-occurrence relationship between words as edges. Then, the method used k -truss graph decomposition technology to obtain the semantic features of the text, and conside red word frequency, word position, complex network feature to design a parameter less scoring function. Finally, it extracted keyword based on the scoring results. Experiments on real datasets show that KEK performs better than other graph-based keyword extraction methods in terms of F, score. [ABSTRACT FROM AUTHOR]

Traditional humor computing tasks rely on artificially constructed features, which are prone to loss of features, and mainly focus on humorous judgment. The Siamese bidirectional GRU attention model based on deep learning compared pairs of humorous texts to determine which sentence was more humorous. Firstly, this paper used the text processor to train the word embedding. Secondly, it used the bidirectional GRU model to obtain the annotation of each word. Finally, it performed the humorous comparison task at the fully connected layer. This paper conducted experiments on the Semeval-2017 Task6-#HashtagWars data set, and used accuracy as an evaluation indicator. The experimental results show that the model and other related models have a significant improvement in the comparison of humorous text. [ABSTRACT FROM AUTHOR]

Fault localization is a critical and difficult task for software quality assurance.With the increasing of software scale,the cost of manual fault localization becomes to be higher,so that many research tasks focus on automatic fault localization.Existing methods of program spectrum based fault localization locate fault to program statements,they are time consuming in large and complex software.To solve this problem,this paper proposed a two-phase fault localization method based on program spectrum.The first phase was coarse grained localization,located fault to program modules.The second phase was fine grained localization,located fault to program statements in the suspicious modules.Finally,it obtained the recommendation list of suspicious statements to assist the developer in debugging.Experimental results show that comparing with traditional methods,this method can guarantee the effect of localization and reduce localization cost by 10.24% on average. [ABSTRACT FROM AUTHOR]

in order to get better performance of 3 D model retrieval and classification,this paper proposed the characterization of 3 D models by multi-modal information fusion,which based on deep-learning model.The method considered the correlation of different modalities in the training step for extracting a more robust feature vector,which benefitted from the proposed correlation loss function.In addition,it applied fusion features to 3 D model retrieval and classification.This paper evaluated the proposed method on the ModelNet40 dataset for 3 D models classification task and retrieval.The comparison between the experimental results and the existing methods proves the superiority of this method,which provides a new idea for the field of 3 D model retrieval and classification. [ABSTRACT FROM AUTHOR]

In order to improve resource utilization efficiency in wireless sensor network, this paper proposed an algorithm to allocate tasks to a pair of collaboratively working sensor nodes based on their link quality. Specifically, it utilized link quality of two collaborative nodes to obtain the ability level of them. Then, allocated each node pair with a task whose task level ( e . g., measured by the computing intensity) could match with the ability level of this node pair, so that each node pair could perform task cooperatively and efficiently. Considering that a node may form different node pairs with multiple nod es, thus allocating redundant tasks( tasks with same task level), algorithm need to adjusted tasks to avoid redundant task execution. Simulation results show that the proposed algorithm can not only improve the task allocation efficiency, but also balance the energy consumption of wireless sensor network. [ABSTRACT FROM AUTHOR]

There are many problems in the process of collecting paper data using single data source, such as insufficient data comprehensiveness and limited data collection speed due to website access frequency limitation. Aiming at these problems, this paper proposed a paper data crawling technology for multi-data sources. Firstly, it used the four Chinese document service websites-How Net, Wanfang Data, Weipu, and Chaoxing as data sources, completed the task of crawling and parsing list page data for the search keywords. Then it used the task scheduling strategy to remove repeated data and balance the tasks. Finally, it used multi-threads for each data source to crawl, parse and store the detail information of the papers, and built a website for search and display. Experiments show that under the same crawling and parsing speed, this technology can complete the paper information collection task more comprehensively and efficiently, which proves the effectiveness of this technology. [ABSTRACT FROM AUTHOR]

Training text simplification model based on seq2seq requires large-scale parallel corpora. However, current task lacks large-scale and well-labeled parallel corpora. To address the above issues, this paper proposed an unsupervised text simplification algorithm that made the learning of the model only need simple and complex sentence datasets without labels. First, the method used denoising autoencoder to learn from simple sentence corpus and complex sentence corpus, respectively, to obtain a simple sentence autoencoder and a complex sentence autoencoder. Then, it combined the two autoencoders to form an initial text simplification model and a text complication model. Finally, it used back-translation to convert the unsupervised text simplification problem into a supervised problem, and iteratively optimized the text simplification model. Experiments on the standard dataset show that the method is superior to the existing unsupervised model on the general indicators BLEU and SARI, and the model has simplified effects at both the lexical and syntactic level. [ABSTRACT FROM AUTHOR]

The control cost and performance difference of various servers in the heterogeneous cloud data center will affect to the operation and maintenance management cost and QoS game balance. Since the task arrival rate is time-sensitive, this paper presented a la rge-scale task scheduling with task arrival rate awareness. Using the task arrival rate and the current state of the server in the cluster, based on saving server operation costs and load balancing of each server in task scheduling, the model could optimize the mean response time and system throughput of the heterogeneous cloud data center. By analyzing of the experimental results, it verifies that the credibility of the cluster server control model in task scheduling is at least at a confidence level of 95%. By comparing with the strategy of SRIT and RabbitMQ, the result verifies the validity and feasibility of the model. [ABSTRACT FROM AUTHOR]

Aiming at the problem of the discriminator' s poor classification accuracy on a small number of labeled samples and insufficient robustness to the local perturbation of manifolds in the generative adversarial network, this paper proposed a generative adversarial network based on variable loss and manifold regularization. The algorithm used a variable loss instead of the original discriminator to solve the adverse effect of the poorly trained classifier on the semi-supervised classification task. In addition, on the basis of variable loss of discriminator, it added manifold regular terms to improve the robustness of discriminator to local disturbance by punishing the variation of classification decision of discriminator on manifold. Using the quality of the generated samples and the semi-supervised classification accuracy as the evaluation criteria of the algorithm, it performed numerical experiments on the dataset SVHN and CIFAR-10. Comparing with other semi-supervised algorithms, the results show that the proposed algorithm can obtain higher quality generated samples and higher precision classification results with a small amount of labeled data. [ABSTRACT FROM AUTHOR]

With the development of deep learning, researchers began to explore the application of deep learning to person re-identification tasks and proposed a large number of methods, which also ushered in new challenges. In order to facilitate scholars to systematically understand the research status and development trends in this field, this paper firstly introduced the person re-identification task and existing problems. Secondly, according to training methods, it discussed the research progress of the supervised learning, semi-supervised learning/weak supervision learning and unsupervised learning for the person re-identification task. It introduced the application of GAN (generative adversarial network) and attention mechanism in person re-identification following recent research hotspots. After that, it enumerated some commonly used data sets (Market-1501, CUHK03, etc.) and their performance in this field. Finally, this paper looked forward to the future research direction of person re-identification. [ABSTRACT FROM AUTHOR]

Due to the roughness of cache replacement mechanism in Spark parallel computing framework, LRU algorithm does not consider the RDD data block reuse problem that easily results in high-reuse block to swap out of the cache and execution inefficiency. By optimizing the weight model and improving the replacement strategy, this paper put forward ERAC cache strategy, which could achieve efficient RDD autonomous cache and hierarchical replacement, included high-reuse self-caching algorithm and cache replacement hierarchical algorithm. Applied ERAC, RA and LRU in standard dataset of SNAP, experimental results show that ERAC can effective improve the memory utilization and task execution efficiency. [ABSTRACT FROM AUTHOR]

The neural network structure with memory and attention mechanism shows some reasoning abilities needed for question answering. DMN is such an architecture, which achieves high accuracy in various language tasks. However, when there is no tag during training to support facts or whether it can be applied to other patterns such as images, does the architecture achieve strong results for question answering? Based on the analysis of DMN, this paper proposed several improvements of its memory and input module. Combining these changes, it introduced a new image input module to answer visual questions. The new dynamic memory model improved the latest technology on visual question answering data sets. The experimental results show that the proposed method achieves good results on visual question answering tasks. [ABSTRACT FROM AUTHOR]

Extractive method extracts sentences from source text, which results in information redundancy. Generative method may generate non-source words, which results in grammatical problems and poor naturalness. BERT as a bidirectional Transformer model, which shows excellent performance in natural language understanding tasks. The application of text generation task remains to be explored. To address the above issues, this paper proposed a three-stage compound text summarization model based on pre-training( TSPT), which combining extractive method and generative method. The bidirectional contextual information word vectors generated by the pre-training of the source text are extracted from the sigmoid function to obtain sentence scores and extract key sentences. Finally, key sentences are rewritten as cloze filling tasks to generate final summarization. The experimental results show that it achieves high scores in CNN/ Daily Mail dataset. [ABSTRACT FROM AUTHOR]

How to allocate tasks efficiently and reasonably is one of the key problems in the research of spatial crowdsourcing. This paper proposed a MQTAM to solve the problem of low allocation efficiency of SC, and proposed an improved particle swarm optimization algorithm. The model took into account the time delay of task completion and the reliability of task completion, and used SCTAM_PSO algorithm to search the optimal assignment scheme to improve the task completion quality. The experimental results and analysis show that the MQTAM and SCTAM are effective and feasible. [ABSTRACT FROM AUTHOR]

In order to better evaluate the ability of crowdsourcing software developers, this paper analysed three basic elements of crowdsourcing software development under crowdsourcing mode, which were crowdsourcing developers, task publishers and crowdsourcing platforms. This paper proposed a model of capability value rate for software crowdsourcing developers. The model subdivided crowdsourcing tasks into three dimensions, namely sub-tasks, time phases and quality characteristics. By evaluating crowdsourcing atomic units, the model evaluated comprehensively the quality of the whole crowdsourcing software. In the evaluation process, the model fully considered the value contribution factors of developers, and deduced the calculation method of the value rate of developers' ability. The validation experiments show that the overall quality value ratio of software crowdsourcing developers tends to be consistent with the increase of task amount. The model has better stability and can effectively evaluate the ability of software crowdsourcing developers. [ABSTRACT FROM AUTHOR]

In traditional mobile cloud computing environment, task scheduling is often determined by random algorithm, adjusting working frequency by dynamic voltage regulation technology. Differentiating tasks exist in task integration, which often leads to many unreasonable task migration and heavy CPU load, and resulting in system damage and a large amount of energy consumption. For multi-workflow tasks, this paper proposed CCS algorithm, which included consolidation algorithm and multitask concurrency algorithm. By increasing the concurrency of transmission and execution between tasks, it could increase the probability of task set integration, could increase the rate of task processing, could reduce the response time of tasks, could increase the CPU usage, and control the CPU usage of host and kernel under the upper threshold. It could avoid CPU overload and optimize local algorithm according to multi-task concurrency, which adjusted task execution location, improved migration efficiency and avoided the limitations of random algorithm. Experimental results show that this algorithm can effectively improve system performance, avoid CPU overload problem, optimize energy consumption and workflow completion time. [ABSTRACT FROM AUTHOR]

Event extraction is an important and challenging task in NLP, completing the identification of event triggers and their arguments from the text. For multiple-event extraction tasks with multiple events in a sentence, this paper proposed a model based on a variant of attention mechanism called DyMAN, which could capture richer context representation and keep more valuable information than the normal attention. The experiments demonstrate that the proposed model can achieve stateof- the-art performance on the ACE 2005 corpus. Compared with the previous best model named JRNN, the DyMAN model achieved a 9. 8% improvement in trigger word classification tasks, and a 4. 5% improvement in factor classification tasks. [ABSTRACT FROM AUTHOR]

In sound event detection tasks, the target event was susceptible to background noise, and wasn't present in a significantly high portion of sound signal flow. To solve the problem, this paper proposed a new method of sound event detection based on one-dimensional dense convolutional network (DenseNet ) with multi-level attention mechanism. Firstly, it used the one-dimensional DenseNet for frame-level detection, which was effective in finding the precise onset and offset time. Then, it introduced the multi-level attention mechanism in the one-dimensional DenseNet model, which made the attention-aware features from different modules change adaptively as layers went deeper. Therefore, the model could automatically select and attend on important frames for the targets while ignoring the unrelated parts (e. g. the background noise segments) . Finally, this paper evaluated the model using DCASE 2017 Task 2 development dataset. Results show that the overall performance of the proposed method has further improvement than the conventional deep learning method. [ABSTRACT FROM AUTHOR]

At present, there is no generation model that can use simple network structure to generate high-quality specific images. For this task, this paper established the conditional boundary equilibrium generative adversarial network by combining the advantages BEGAN and adding additional condition features and the MSE loss. It used this method to extract the generation model for specific image generation. Experimental results show that compared with other supervised class generation models, this method can use simpler networks to achieve faster convergence speed and generate images with better quality and diversity. [ABSTRACT FROM AUTHOR]

In view of the loss strategy and structure of traditional discriminator is difficult to extract more abstract and task related robustness features, which leads to the shortage of semi supervised image classification, this paper proposed a generation countermeasure network based on feature recalibration. In order to learn the related features of the task, on the basis of the existing semi supervised GAN, it introduced the unsupervised canonical loss regular term of the model under different states to the discriminator, and the different output of the same input corresponding to the two branches of the training sample was punished to guide the optimization direction of the feature recalibration. In addition, it added the compression activation module to the discriminator to optimize the convolution pool structure of the traditional discriminator. The module automatically learned the importance of each characteristic channel, and could extract the features related to the task to suppress the unrelated features of the task, and realized the recalibration function of the feature, thus improving the performance of the semi-supervised image classification. [ABSTRACT FROM AUTHOR]

Aiming at the task of object recognition under resource constrained condition, this paper proposed a method of object recognition based on light weight depth network. By optimizing the design method of the network structure such as convolution operation, model parameter compression and enhancement of feature expression depth, this paper designed and implemented the lightweight network model structure named Se-DResNet for embedded platform. So that the depth network model could reduce the parameters of the model and the resources needed for operation under the condition of guaranteeing the precision. The experimental results show that the lightweight depth model has better performance than that of the basic model ResN et proposed by ILSVRC-15 champion, and it can achieve the model accuracy of 93. 5% under the condition that the model with 10. 2% fewer parameters on ImageNet-67 data set. [ABSTRACT FROM AUTHOR]

For the allocation of spatial tasks in mobile crowd sensing, the spatial distance between users and tasks directly affects the cost required to complete the tasks, while the existing research about this is considered inadequate. In order to minimize cost, the paper proposed a spatial task allocation mechanism. Firstly, based on genetic algorithm and greedy algorithm, it designed an efficient task allocation method to minimize the cost required to complete the tasks. Secondly, considering the randomness of user sensing quality,it designed a mechanism of use r sensing quality updating based on the historical quality and the quality of the current task. To verify the effectiveness of the proposed mechanism, it conducted simulations compared with the two benchmarks. The results sh ow that for spatial task allocation the proposed mechanism has better results in terms of total cost and spatial distance that the user needs to move to perform the task. [ABSTRACT FROM AUTHOR]

In order to solve the shortcomings of current Hadoop cluster inherent scheduling distributed methods in heterogeneous resource environments, this paper proposed an adaptive scheduling algorithm NCAS based on the node capability. Firstly, NCAS algorithm calculated the scheduling factor based on node performance and task characteristics; Secondly, the scheduling factor determined the amount of data and the number of task slot that each node should be assigned; Finally, NCAS algorithm dispatched data and tasks more into fast nodes and less into slow nodes. Experimental results show that, compared with the traditional scheduling algorithms, NCAS algorithm can greatly reduce the number of speculative tasks, significantly reduce the job completion time, and improve the task execution efficiency. [ABSTRACT FROM AUTHOR]

In order to obtain good results when dealing with different data types or tasks, this paper designed a convolutional n eural network (CNN) architecture based on adaptive false nearest neighbors method. It applied the center moment to CNN pooling operation and used sparse filtering algorithm to realize the unsupervised training process. It also set the size of the convolution mask (convolution kernel ) of the CNN and the unit of convolution (CNN neural) . In addition, the architecture also used adaptive false nearest neighbors method to simplify tasks such as modeling and prediction. Experimental results confirm that the improved CNN architecture is less complex and can be trained faster and it is less prone to overfitting. [ABSTRACT FROM AUTHOR]

It is making great progress in the autonomous driving field with the rapid development of road scene understanding techniques. The safety is a concerning issue with respect to the real-time and accurate performance in the related tasks which contains the road segmentation, road classification and vehicle detection. To this end, this paper proposed an approach based on deep residual learning with an encoder-decoder network structure. On the one hand, the encoder network structure used different layers of residual networks to extract the abstract features in the high dimension, which shared in the next three tasks. On the other hand, the decoder network structure adopted a mechanism of parallel computing for sub-tasks, i. e., executed the road segmentation, vehicle detection and road classification tasks simultaneously. Additionally, it used the fully convolutional networks to upsample the extracted features to specifically solve the problem of road segmentation. At last, the experimental results show that the processing rate can effectively reach more than 15 fps with the high accuracy guaranteed. [ABSTRACT FROM AUTHOR]

For the problem of lacking real-time scheduling mechanism in the embedded Forth operating system at present, this paper investigated the key technologies of multitask-scheduling of embedded operating system based on Forth virtual machine architecture. On the basis of Forth virtual machine technology, this paper defined a new interrupt task type to deal with real-time emergencies and proposed a new scheduling algorithm, which enabled the scheduling of terminal tasks, background tasks and interrupt tasks in Forth system to be successfully executed. The experimental results demonstrate that the improved Forth system can handle emergencies through real-time scheduling and has a high real-time response. It is especially suitable for the embedded environment which is required for real-time performance, so as to meet the increasingly complex application requirements of embedded system for efficient operating system and Forth technology. [ABSTRACT FROM AUTHOR]