Search

Intelligent optimization algorithms (IOA) refer to a kind of algorithm that is used to solve optimization problems by imitating the survival and evolution process of natural creatures or physical phenomena as the algorithm principle. The well-known intelligent optimization algorithms include genetic algorithm, particle swarm optimization, simulated annealing algorithm, etc. Intelligent optimization algorithm is a heuristic method, which is widely used in solving optimization problems and provides some new ideas for solving some practical problems. With the advan-cement of science and technology and the increase in the complexity of application scenarios, traditional intelligent optimization algorithms can no longer satisfy optimization problems in terms of solving effects and accuracy. Therefore, new and more efficient intelligent optimization algorithms are constantly being proposed. Several new intelligent optimization algorithms have been proposed at home and abroad in recent years, such as butterfly optimization algorithm (BOA), moth-flame optimization (MFO), sine cosine algorithm (SCA), grasshopper optimization algorithm (GOA), Harris hawks optimization (HHO) and sparrow search algorithm (SSA). This paper describes the basic principle, algorithm steps, corresponding improvement strategies with advantages and disadvantages of each algorithm. To objectively compare the performance of each algorithm, this paper further evaluates the performance of each algorithm through 21 test functions of 3 types and 6 indicators. Finally, this paper summarizes the characteristics of the algorithm and prospects the development direction of intelligent optimization algorithm.

Personalized learning recommendation is a research field of intelligent learning. Its goal is to provide specific learners with effective learning resources on the learning platform, thereby enhancing learning enthusiasm and learning effect. Although the existing recommendation methods have been widely used in learning scenarios, the scientific rules of learning activities make personalized learning recommendations unique in terms of personalized parameter setting, recommendation goal setting, and evaluation standard design. In response to the above-mentioned problems, the research of personalized learning recommendation in recent years is reviewed on the basis of investigating a large number of literatures. The research on personalized learning recommendation is systematically sorted out and interpreted from five aspects, i.e., the general framework of learning recommendation, learner modeling, learning recommendation object modeling, learning recommendation algorithm, and learning recommendation evaluation. Firstly, the general framework of learning recommendation system is proposed. Secondly, the ideas and methods of learner modeling are introduced. Next, the ideas and methods of learning recommendation object modeling are discussed. Then, this paper summarizes the algorithm and model of learning recommendation. The following, this paper summarizes the design and method of learning recommendation evaluation. This paper also analyzes the main ideas, implementation plans, advantages and disadvantages of the existing research in these five aspects. Finally, this paper also looks forward to the future development direction of personalized learning recommendation, which lays foundation for further in-depth research on intelligent learning.

Recognizing multi-scale objects and objects with occlusions is a key and difficult point of task in object detection. In order to detect objects with different sizes, the object detector usually uses the hierarchical structure of multi-scale feature map constructed by convolutional neural network (CNN). However, due to the small convolution layer of the bottom feature map, the top-down structure lacks the detailed information needed to capture the features of small object. The performance of these object detectors is limited. Therefore, based on the Faster R-CNN (region-convolutional neural network) framework, this paper proposes Collaborative R-CNN. This paper designs a cascaded network structure that integrates multi-scale feature maps to generate deeply fused feature information and thereby improving the ability to detect small objects. Moreover, the quantization in the RoIPooling process greatly limits the recognition ability of small objects. In order to further improve the robustness of the method, a multi-scale RoIAlign is designed to eliminate such quantization, and the ability of network to detect objects with different scales is improved by multi-scale pooling. Finally, this paper combines an adversarial network with the proposed network to generate training samples with occlusions, significantly improving the classification ability of the model, and robustness to detect occlusions. Experimental results for the PASCAL VOC 2012 and PASCAL VOC 2007 datasets demonstrate the superiority of proposed approach relative to several state-of-the-art approaches.

Mobile edge computing (MEC) supports terminal devices to offload tasks or applications to edge cloud server for processing. Edge cloud server will consume local resources when processing external tasks, so it is particularly important to build a resource pricing mechanism that charges terminal devices to reward edge cloud. The existing pricing mechanism relies on the static pricing of intermediaries, and high cost and late processing of terminal tasks make it difficult to realize the effective utilization of edge cloud computing resources. Aiming at the above problems, this paper proposes an edge cloud resource pricing mechanism based on Stackelberg game. Firstly, in view of the local task shelving problem of terminal devices due to insufficient funds during resource pricing, an auxiliary mechanism including loans and incentives is proposed to realize the timely processing of terminal devices tasks. Secondly, four price-oriented factors that affect resource pricing are proposed, and two pricing schemes, consistency and elasticity, are formulated to improve the accuracy and efficiency of pricing and prepare for dynamic pricing. Then, in order to make the dynamic pricing between terminal devices and edge cloud directly, a resource pricing mechanism model based on Stackelberg game is built, and the resource demand and pricing problem is transformed into the problem of maximum revenue of edge cloud and minimum payment cost of terminal devices. Finally, through improved reinforcement learning SARSA (state action reward state action) algorithm, the optimal strategy of resource demand and pricing is obtained. Experiments show that the pricing mechanism proposed in this paper is more than 12% better than other pricing algorithms in terms of edge cloud revenue maximization, and the edge cloud revenue under the elasticity pricing scheme is 24% better than that of the consistency pricing scheme.

Proteins, as the material basis of life, are the ultimate controller and direct performer of life activities. Most proteins perform their biological functions by binding to other proteins to form complexes. The identification of protein complexes can help people to better understand the organization and functions of the complexes, as well as the mechanisms of cells. At present, the rapid development of high-throughput experimental technology has led to huge amounts of protein-protein interaction (PPI) data. Many methods for computationally predicting protein complexes based on PPI data have been proposed. Different methods have their own characteristics and advantages, and they also have inherent drawbacks. Firstly, this paper classifies and comprehensively analyzes and reviews the existing protein complex prediction methods. Then, it introduces the commonly used evaluation indicators and main data sets in complex prediction, compares and analyzes the prediction performance of several representative methods. Finally, it summarizes state-of-the-art complex prediction methods, highlights future research directions, and puts forward several issues to be resolved in the future. It is hoped that through the analysis and comparison of various methods, this paper can provide some valuable guidance and future directions for users and researchers on using the existing methods and developing new methods of protein complex prediction.

Effective graph analysis methods can reveal the intrinsic characteristics of graph data. However, graph is non-Euclidean data, which leads to high computation and space cost while applying traditional methods. Graph embedding is an efficient method for graph analysis. It converts original graph data into a low-dimensional space and retains key information to improve the performance of downstream tasks such as node classification, link prediction, and node clustering. Different from previous studies, this paper focuses on both static and dynamic graph embedding. Firstly, this paper proposes a universal taxonomy of static and dynamic methods, including matrix factorization based methods, random walk based methods, autoencoder based methods, graph neural networks (GNN) based methods and other embedding methods. Secondly, this paper analyzes the theoretical relevance of static and dynamic methods, and comprehensively summarizes the core strategy, downstream tasks and datasets. Finally, this paper proposes four potential research directions of graph embedding.

The widely researched sentence-level event extraction methods struggle to extract all arguments of the same event from a whole document. To solve this problem, this paper proposes a joint model for document-level event extraction based on deep learning. Firstly, an entity recognition module based on multi-head self-attention mechanism is used to identify entities and their types sentence by sentence. Then, an event type detection module trained by defining the importance of different argument roles, is used to locate the event mention sentence and predict the event type without the help of event triggers. Finally, an event argument extraction module embeds every entity??s semantic vector with its type information and its distance to the event mention sentence before feeding into a context-aware Transformer, in order to extract arguments within the document scope. In addition, by training the three modules mentioned above jointly, this paper realizes an end-to-end event extraction model and avoids error propagation problems in traditional pipeline models. The experimental results on a public dataset shows that, when each document contains only one event, the proposed model achieves a 86.3% [F1-score], which outperforms state-of-the-art methods, and the training process completes rather quickly.

With the increasing popularity of mobile devices, a large amount of user check-in point of interests (POIs) data have been accumulated. The information of user check-in makes the recommendation of the next POI become a hot issue in recent years. The accuracy of the next POI recommendation is restricted by two aspects: On the one hand, the sparsity of check-in data. At present, most researchers alleviate data sparsity problem to a certain extent by introducing geographic correlation or friend evaluation information on POIs. However, not all POIs have strong geographic correlation, and only a small number of users comment on POIs where they have check-ins. On the other hand, deep learning based check-in sequence training has the problem of gradient disappearing. In order to solve these problems, this paper proposes a user??s next POI recommendation model which integrates the influence of regions and friends. Firstly, the region information of POIs is integrated into the sequence of POIs. Sequentially, this paper uses the neural network model with residual connection to embed the sequence, to avoid gradient vanishing and improve the convergence of the model. Finally, this model integrates the information of POIs visited by friends for recommendation, to improve the accuracy of POI recommendation. Experimental results show that the proposed model is more accurate than other existing models.

Multi-view subspace clustering has attracted increasing attention for revealing the inherent low-dimension structure of the data. Nevertheless, most existing methods directly fuse the multiple noisy affinity matrices from the original data, and commonly conduct clustering after obtaining a unified multi-view representation. Separating the representation learning from the clustering process can result in a suboptimal clustering result. To this end, this paper proposes a one-stage partition-fusion multi-view subspace clustering algorithm. Instead of directly fusing the noisy and redundant affinity matrices, this paper fuses the more discriminative partition-level information extracted from the affinity matrices. Moreover, this paper proposes a new framework, integrating representation learning, multiple information fusion and final clustering process. The three sub-processes promote each other to serve clustering best. The promising clustering results can lead to better representations and therefore better clustering performance. Consequently, this paper solves the resultant optimization problem through an alternative algorithm. Experiment results on four real-world benchmark datasets show the effectiveness and superior performance of the proposed method over the state-of-the-art approaches.

Trial difficulty prediction (TDP) is the task of automatically predicting the difficulty of a trial given the case text, which has a broad application prospect in judicial intelligent system. In practice, the tools of TDP rely heavily on the experience of experts, which leads different conclusions in predicting the difficulty of the trial. However, there are few related research work. To address these issues, this paper regards it as a text classification problem in natural language processing. Through the analysis, it is found that, traditional text classification methods don??t consider the structural uniqueness and logical dependence among trial elements in complaint, which makes it difficult to predict the difficulty of a trial accurately. In order to solve the mentioned challenges, this paper carefully studies indictments and considers the complex and simple trial elements for judging cases, presents an end-to-end model, MAT-TAN (mask-attention and topological association network). Specifically, this paper proposes a novel mask-attention network (MAT), to carry out fine-grained analysis of a case description text in indictments. The masking mechanism plays a role of the intelligent gatekeeper, focusing on the specific position of the trial elements in indictments. Together with the self-attention mechanism, it extracts the comprehensive and accurate characteristics of each trial element. This paper proposes a novel topological association network (TAN), which models the judicial logic dependency relationship between different elements, and effectively integrates the characteristics of different elements. Finally, the TDP is realized. The experimental results conducted on real-world datasets demonstrate that the MAT-TAN can improve the macro averaged F1 up to 0.036 compared with baselines, showing that it has a better performance in TDP.

Current research on lip recognition focuses on improving recognition accuracy and studying features of multimodal inputs. However, little attention has been paid to improving the effectiveness of lip visual features. Lip visual information plays a key role in visual speech recognition and lip recognition. It is important when audio is destroyed or has no information. How to obtain accurate and effective lip visual features is one of the most difficult tasks in lip recognition. This paper reviews the latest research work on lip recognition in recent years from three aspects: lip dataset, traditional visual feature extraction methods, and in-depth learning methods for visual feature extraction. Firstly, this paper summarizes the dataset for lip recognition. The lip dataset is divided into two types: front view and multi-view. Further two types of datasets are summarized from their characteristics, limitations, and download addresses. Secondly, this paper introduces the traditional methods of lip visual feature extraction from the perspective of pixel point, shape and mixed features. The basic idea, network structure and features of each method are mainly introduced. In the deep learning method of lip visual feature extraction, the network structure, advantages and disadvantages of four deep learning methods are mainly introduced, such as 2D CNN, 3D CNN, 2D CNN combined with 3D CNN, and other neural networks. The performance of these methods on open datasets is compared. Finally, the challenges faced by lip visual feature extraction methods and future research trends are prospected.

In the early screening of ophthalmic diseases, diabetes, heart disease and other diseases, automatic mon-itoring of fundus capillaries occupies an important position. To solve the deficiency of blood vessel segmentation caused by the imprecise expression of capillary features, this paper proposes a multi-block residual neural network model (MbResU-Net) which uses the encoding-decoding network structure. In order to reduce the information loss caused by the semantic gap between network encoder and decoder, the model uses a nonlinear network structure instead of shortcut connections to be embedded in the network. In order to obtain the detailed features of fundus blood vessels, MbResU-Net connects three U-shaped networks in a residual block, and then extracts as much as possible low-level features of the fundus tube on the premise of avoiding information loss. To ensure the seg-mentation quality, with the proper pre-processing images to the network, it designs a cross-entropy loss function incorporating the cost matrix to train the network parameters. This paper compares MbResU-Net and the existing fundus blood vessel segmentation algorithms on the DRIVE and CHASE DB1 color fundus image datasets. Exper-iments show that MbResU-Net is superior to existing methods in Sen, ACC and AUC. The Sen is 0.7987 and 0.7972, ACC is 0.9648 and 0.9726 and AUC is 0.9791 and 0.9824. Experiments prove the effectiveness and robustness of the model in the segmentation of complex curvature and small blood vessels.

Face synthesis is one of the hot topics in the field of computer vision because of its application and technical value. In recent years, the breakthrough of deep learning has attracted much attention in this field. This paper divides the research in this field into four subcategories: face identity synthesis, face movements synthesis, face attributes synthesis and face generation, and systematically summarizes the development process, status quo, and existing problems of these subcategories. First of all, for face identity synthesis, three approaches are summarized, including computer graphics, digital image processing and deep learning. This paper summarizes their respective routine processes, and analyzes the technical principles of milestone work in detail. Secondly, face movements synthesis is further divided into label driven expression editing and real face driven face reenactment, where the shortcomings and problems in each field are pointed out. Then, the development of face attribute synthesis based on generative model is introduced, especially generative adversarial network. Finally, this paper briefly describes all kinds of researches on face generation. In addition, this paper also introduces the practical application and related problems of face synthesis field and provides the possible research direction in this field.

RNA-binding protein (RBP) is a total name of a class of proteins that bind to RNA (ribonucleic acid) along with the process of RNA??s regulation metabolic. An RBP may have multiple target RNAs, and its defective expression may cause various diseases. Existing methods are mostly designed for a specific RBP binary classification model to predict whether an RNA can bind to it. But these methods do not take into account the similarity and association between different RBPs. Therefore, iDeepM uses multi-label deep learning methods to improve it. This method fuses multi-label technology and long short term memory (LSTM) network, learns the similarity between different RBPs, and predicts the binding of a given RNA to multiple RBPs. However, this method fails to perform sufficient feature learning and multi-label learning on RNA sequences, and the prediction accuracy is low. This paper continues the research ideas of iDeepM multi-label, and proposes a new method RNA-RBP multiview learning (RRMVL). For the first time, the RNA sequence view, the amino acid sequence view, the RNA sequence semantic view and the multi-gap dipeptide component view are used to compose multi-view data to deal with multi-label RBP recognition. In order to use the different learning advantages of multi-view data, this paper fuses the deep features extracted from four views and uses the principle of logistic regression to learn multi-label features from them. After that, the learnt weighted feature vectors are fed to the multi-label classifier chain to achieve the optimal multi-label chain learning effect. Experimental studies show that the prediction accuracy of the RNA-binding protein recognition model combining multi-view and multi-label learning has been significantly improved compared with the previous single-view method.

Fine-grained image categorization aims to distinguish the sub-categories from a certain category of images. Generally, fine-grained data sets have the characteristics of the intra-class similarity and inter-class variation, which makes the task of fine-grained image categorization more challenging. With the increasing development of deep learning, the methods of fine-grained image categorization based on deep learning exhibit more powerful feature representation and generalization capabilities, and can obtain more accurate and stable classification results. Therefore, deep learning has been attracting more and more attentions and research from the researchers in the fine-grained image categorization. In this paper, starting from the background of fine-grained image categorization, the difficulties and the meaning of fine-grained image categorization are introduced. Then, from the perspectives of strong supervision and weak supervision, this paper reviews the research progress of fine-grained image classification algorithms based on deep learning, and a variety of typical classification algorithms with excellent performance are introduced. In addition, the YOLO (you only look once), multi-scale CNN (convolutional neural network), and GAN (generative adversarial networks) model are further discussed in the application of fine-grained image categorization, the perfor-mance of the latest relevant fine-grained data augmentation methods is compared and an analysis of different types of fine-grained categorization methods is made under complex scenarios. Finally, by comparing and summarizing the categorization algorithms, the future improvement directions and challenges are discussed.

Generally, the traditional multi-view learning methods assume that all samples are completed in all views. However, this assumption often fails in real applications because of limited access to data, equipment malfunc-tion, as well as occlusion and so on. Thus, it is ineffective to directly use these traditional methods for addressing incomplete multi-view data. At present, several effective incomplete learning algorithms have been proposed, but they do not make full use of label information and thus reduce the discrimination of the recovered samples. Therefore, this paper proposes an incomplete multi-view classification method via discriminative and sparse representation (IMVC-DSR). Specifically, this method is based on the assumption that missing samples can be represented linearly and sparsely by a few observed samples. Meanwhile, in order to make full use of label prior information to improve the discrimination of recovered samples, this method encourages that missing samples are represented by the samples from their same classes rather than the others. Also, according to the view consistence across multi-view, this paper designs a selection operator to select the same samples from different views and meanwhile expects their linear representations are consistent with each other. Finally, experiments demonstrate the efficacy of the proposed method on five public benchmark datasets.

Bloom filter (BF) is highly efficient for membership queries, which is widely used in blockchain mem-bership queries. Aiming at the existing BFs do not exploit the data characteristics of blockchain and the features of modern processors, this paper proposes a novel bloom filter variant named blockchain bloom filter (BBF). Firstly, this paper modifies data structure which divides BBF into groups, so that the mapping range of an element is limited into a group to reduce the number of cache misses and improve cache efficiency. Secondly, a simplified three-stage Hash process is presented to eliminate computing overhead by taking advantage of blockchain data characteristics. On this basis, BBF uses single instruction multiple data (SIMD) to parallelize element insertion and query, and accelerate BBF construction and query speed, which can realize efficient query and analysis of blockchain data ultimately. The experimental result shows that BBF??s membership query speed under positive query can improve 4 times and 3 times over the other two state-of-the-art bloom filter variants, i.e., BF, OMBF（one-memory bloom filter）, which enables significant performance improvement.

Knowledge graph question answering (KGQA) is the procedure of processing natural language questions posed by users to obtain relevant answers from knowledge graph (KG) based on some form of KG. Due to the limitation of knowledge scale, computing power and natural language processing capability, the early knowledge base question answering systems were limited to closed-domain questions. In recent years, with the development of KG and the construction of open-domain question answering (QA) datasets, KG has been used for open-domain QA research and practice. In this paper, in accordance with the development of technology, the open-domain KGQA is summarized. Firstly, five rule and template based KGQA methods are reviewed, including traditional semantic parsing, traditional information retrieval, triplet matching, utterance template, and query template. This type of methods mainly relies on manually defined rules and templates to complete QA task. Secondly, five deep learning based KGQA methods are introduced, which use neural network models to complete the subtasks of QA process, including knowledge graph embedding, memory network, neural network-based semantic parsing, neural network-based query graph, and neural network-based information retrieval method. Thirdly, four general domain KG and eleven open-domain QA datasets, which KGQA commonly used are described. Fourthly, three classic KGQA datasets are selected according to the difficulty of questions to compare and analyze the performance metric of each KGQA system, and the effect between above methods. Finally, this paper looks forward to the future research directions on this topic.

In the forward secure sequential aggregate (FssAgg) signature scheme, the signer combines signatures generated in different intervals under different secret keys incrementally and sequentially in a layered “onion-like” fashion with the first signature innermost in the aggregate. In contrast with general (not forward-secure) aggregate signature schemes which aggregate signatures from multiple signers, a FssAgg signature scheme aggregates signatures of a single signer not the signatures of different signers, so the verifier uses a single public key to verify the entire aggregate. With the advantages of forward secure signature and the aggregate signature at the same time, the FssAgg signature scheme has been widely applied in logging systems and blockchain since it was proposed in 2007. Although there have been several FssAgg signature schemes, all of them are based on the classic number theory problem, which are no longer secure in the quantum era. So looking for the quantum-immune FssAgg signature is much urgent. Based on the small integer solution over lattice, a FssAgg signature in the standard model is proposed. In order to improve the efficiency, this paper uses the fixed-dimensional lattice basis delegation technique to update the keys, and the forward security is achieved. This paper uses the message addition technique and the pre-image sampling technique to embed the message and the small integer solution into the signature process, and the existentially unforgeability of the scheme is also guranteed in the standard model.

Recently the rapid development of the deep learning technique has provided a powerful tool for the clustering research, and has given rise to quite a number of deep neural network-based clustering methods. Among these methods, deep embedding clustering (DEC) has been drawing increasing attention, due to its advantage in performing deep representation learning and optimizing clustering assignment simultaneously. However, one limita-tion to DEC lies in its sensitivity to the hyper-parameter λ, which often requires manual fine-tuning. To address this problem, this paper presents an improved deep embedding clustering method with ensemble learning (IDECEL). Instead of searching for a single optimal hyper-parameter, this paper makes use of a set of diversified hyper-parameters λ to construct an ensemble of diversified base clusterings. By exploiting the concept of entropy, this paper evaluates the uncertainty of the clusters in these base clusterings and weights them accordingly. Further, this paper constructs a locally weighted bipartite graph between base clusters and data samples, and efficiently partitions it to obtain a better clustering result. Experimental results on multiple datasets show that the proposed IDECEL method not only alleviates the hyper-parameter sensitivity problem in DEC, but also exhibits more robust clustering performance than several other deep clustering and ensemble clustering methods.

Machine reading comprehension (MRC) is a research field driven by datasets. The task of MRC is to make the machine correctly answer relevant questions on the basis of understanding the natural language text. But limited by datasets, most of the machine reading comprehension tasks are single-turn question answering, and there is no dependency between question answering pairs. Conversational question answering (ConvQA) is the human-computer process that enables the machine to carry out continuous topics when it helps human to obtain information. In recent years, with the development of MRC datasets and deep neural networks, researchers combine machine reading com-prehension and conversational question answering to form a new more complex field, conversational machine com-prehension (CMC). This has greatly boosted the field of MRC. This paper summarizes the latest research progress of CMC in recent years from three aspects. Firstly, it describes the definition of CMC task, the challenges and the characteristics of the datasets. Then, it summarizes the research progress and the current architecture of the latest models, and focuses on the relevant technical methods used in conversational history embedding and conversational reasoning. Finally, this paper analyzes the current model of CMC, and prospects the future research hotspots and methods of CMC.

Video object detection is to solve the problem of object localization and recognition in every video frame. Compared with image object detection, video is featured by high redundancy, which contains a lot of local spatio-temporal information. With the rapid popularity of deep convolutional neural network in the field of static image object detection, it shows a great advantage over traditional methods in performance. Besides, it plays a due role in video-based object detection task. However, the current video object detection algorithms still face many challenges, such as improving and optimizing the performance of mainstream object detection algorithms, maintaining the spatiotemporal consistency of video sequences, and making detection of model lightweight. In view of the above problems and challenges, on the basis of investigating a large number of literature, this paper systematically sum-marizes the video object detection algorithm based on deep learning. Based on the basic methods like optical flow and detection, these algorithms are classified. In addition, in the angles of backbone network, algorithm structure and data sets etc., these methods are explored. Combined with the experimental results in the ImageNet VID data set, this paper analyzes the performance advantages and disadvantages of typical algorithms of this field, and the relations between these algorithms. As for video object detection, the problems to be solved as well as the future research direction are expounded and prospected. Video object detection has become a hot spot pursued by many computer vision scholars. More efficient and accurate algorithms will be proposed, and its development direction will be better and better.

Data pricing is the behavior of taking data as assets and pricing the assets. In the current data markets, there is little transparency and information asymmetry between buyers and sellers, resulting in confusion in data pricing. If there is a standard process and evaluation method for data pricing, buyers can obtain datasets they need with a reasonable price, moreover, it would also improve the efficiency of data trading markets. This paper retrieves the relevant literatures about data pricing in recent years, and then summarizes the definition, characteristics, development and application scenarios of data pricing. It describes the data transaction process and data transaction costs, and focuses on describing two important research hotspots that affect data pricing, data pricing strategy and data pricing model. It comprehensively evaluates the mechanisms, advantages, disadvantages, and application scenarios of the existing six data pricing strategies and five pricing models. Finally, it analyzes the challenges of data pricing from three aspects: data value evaluation, data transaction rule and data privacy protection. And it prospects the future development trends of data pricing. The research results of this paper will provide valuable reference and foundation for future relevant work.

In order to meet the computing requirements of large-scale spiking neural network (SNN), neuromorphic computing systems usually need to adopt large-scale parallel computing platforms. Therefore, how to quickly deter-mine a reasonable number of computing nodes for the SNN workload (that is, how to properly map the workload to the computing platform) to obtain the best performance, better power consumption and other indicators has become one of the key issues that a neuromorphic computing system needs to solve. Firstly, this paper analyzes the SNN workload characteristics and establishes a calculation model for it. Then for the NEST simulator, this paper further instantiates SNN load model of storage, calculation and communication. Finally, this paper designs and implements a NEST-based workload automatic mapper for SNN (SWAM). SWAM can automatically calculate the mapping result and complete the mapping, avoiding the extremely time-consuming manual trial process of workload mapping. SNN typical applications are run on three different computing platforms, ARM+FPGA, ARM, and PC clusters, and the mapping results of SWAM and LM (Levenberg-Marquardt) algorithm fitting and measured are compared. Experi-mental results show that the average mapping accuracy of SWAM reaches 98.833%. Compared with the LM method and the measured mapping, SWAM has absolute advantage in time cost.

In the published reviews of natural language pre-training technology, most literatures only elaborate neural network pre-training technologies or a brief introduction to traditional pre-training technologies, which may result in the development process of natural language pre-training dissected artificially from natural language processing. Therefore, in order to avoid this phenomenon, this paper covers the process of natural language pre-training with four points as follows. Firstly, the traditional natural language pre-training technologies and neural network pre-training technologies are introduced according to the updating route of pre-training technology. With the characteristics of related technologies analyzed, compared, this paper sums up the process of development context and trend of natural language processing technology. Secondly, based on the improved BERT (bidirectional encoder representation from transformers), this paper mainly introduces the latest natural language processing models from two aspects and sums up these models from pre-training mechanism, advantages and disadvantages, performance and so on. The main application fields of natural language processing are presented. Furthermore, this paper explores the challenges and corresponding solutions to natural language processing models. Finally, this paper summarizes the work of this paper and prospects the future development direction, which can help researchers understand the development of pre-training technologies of natural language more comprehensively and provide some ideas to design new models and new pre-training methods.

In recent years, with the popularization of Internet and technologies like big data analysis, the demand for mobile medical services has become more and more urgent, which mainly focuses on ascertaining their diseases based on symptoms and further choosing hospitals and doctors with good service quality based on diseases. In order to tackle problems above, this paper designs and implements a recommendation system for medical consultation based on knowledge graph and deep learning. Using the open data on Internet, a “disease-symptom” knowledge graph is constructed. Once given symptom description, a disease candidate set is built to help user self-diagnose. To improve the accuracy of disease diagnosis, a vector representation of entities in the knowledge graph is trained by a knowledge graph embedding model. Then the disease candidate set is expanded by selecting disease entity with the shortest Euclidean distance with diseases in the set. Combining the two above, disease diagnosis service is provided. To recommend hospitals and doctors, given open media data, this paper uses a deep learning model and combines it with existing quality evaluation indicators for medical service to achieve scoring for doctors multi-dimensional service quality automatically. Finally, this paper verifies the accuracy of the disease diagnosis service and the doctor recommendation service by constructing test sets and designing questionnaires, which reach 74.00% and 90.91%, respectively.

With the vigorous development of Internet technology and the popularization of smart devices, while the amount of multimedia data exploding, their forms become increasingly diverse. People's demand for information is no longer satisfied with single-modal data retrieval. Realizing cross-modal retrieval through knowledge collaboration of different modalities has become a research hotspot in recent years. On the basis of in-depth understanding and analysis of the research background and progress of cross-modal retrieval, with the key technology of cross-modal retrieval, public subspace modeling as the main line, this paper analyzes three types of methods of cross-modal retrieval technology: traditional statistical analysis, deep learning, and Hash learning. This paper conducts a comprehensive and multi-angle comparative analysis on the research content, key technology, limitations, applicability and characteristics from different angles, and experiments are done for more in-depth comparisons. Finally, the difficulties to be solved in cross-modal retrieval, future exploration directions, mainstream design ideas and development trends in recent years are fully prospected to provide a theoretical basis for further research.

In real scenes, due to the imperfections of equipment and systems or the existence of low-light environments, the collected images are noisy. The images will also be affected by additional noise during the compression and transmission process, which will interfere with subsequent image segmentation and feature extraction processes. Traditional denoising methods use the non-local self-similarity (NLSS) characteristics of the image and the sparse representation in the transform domain, and the method based on block-matching and three-dimensional filtering (BM3D) shows a powerful image denoising performance. With the development of artificial intelligence, image denoising methods based on deep learning have achieved outstanding performance. But so far, there is almost no relevant research on the comprehensive comparison of image denoising methods. Aiming at the traditional image denoising methods and the image denoising methods based on deep neural networks that have emerged in recent years, this paper first introduces the basic framework of the classic traditional denoising and deep neural network denoising methods and classifies and summarizes the denoising methods. Then the existing denoising methods are analyzed and compared quantitatively and qualitatively on the public denoising data set. Finally, this paper points out some potential challenges and future research directions in the field of image denoising.

Clustering by fast search and find of density peaks (DPC) is a new density clustering algorithm proposed in recent years. The core of the algorithm is based on local density and relative distance. By drawing a decision diagram, the cluster center is selected manually, and the clustering is completed. The DPC algorithm uses the cutoff distance to calculate the local density, and essentially only considers the number of neighboring nodes around it, and the algorithm uses a single-step allocation strategy, which limits the accuracy and effectiveness of the algorithm for any data set to a certain extent. To solve the above problems, this paper proposes an optimized density peaks clus-tering algorithm based on second-order k neighbors (SODPC). The algorithm calculates direct density and indirect density by introducing second-order k neighbors of nodes, redefines the calculation method of local density, and on this basis, it defines the multi-step allocation strategy of non-central nodes to complete the clustering. Through manual and real data tests, it is proven that the algorithm in this paper has a good clustering effect on irregular and uneven density data sets.

Privacy amplification means that the communication parties extract a shorter but highly confidential string [S] by negotiating on the public channel while sharing a partially confidential string S. Enemy only knows part of the information of the string S and the information that it knows about [S] is almost negligible. Recently, people use the generative adversarial networks (GANs) to realize the secure communication with the present of the adversary. This paper proposes to use the generative adversarial network to achieve a privacy amplification scheme when the adversary ability is limited. First, this paper proposes a privacy amplification implementation scenario. The two parties use the conversation information to generate identical keys, and the adversary listens to the conversation information. Then, with reference to the neural network structure in the basic encrypted communication model of Abadi et al., a privacy amplified communication model is built. The experiment tests the privacy amplified communication when the enemy knows part of the information or the opponent's computing power is weak. By modifying the activation function, increasing the complexity of the model and modifying the filter of convolutional neural network, the final results show that when the adversary gets 70% of the communication information, or when the communicator is more complex than the adversary model, both parties can negotiate a secure key to complete the function of security enhancement.

In order to solve the problem of model drift caused by the scale change of the target vehicle when the traditional correlation filtering algorithm tracks the vehicle in the video taken by the UAV (unmanned aerial vehicle), this paper proposes an improved scale adaptive vehicle tracking algorithm. The algorithm is based on nuclear correlation filtering. By constructing a spatial tracker that distinguishes scales, this paper uses two filters, one to locate the target vehicle??s position, the other to estimate the scale of the target vehicle, in order to quickly determine target-related information and achieve scale adaptation. In addition, in order to solve the problem of poor tracking performance caused by rapid deformation of the target vehicle, color features that are less sensitive to deformation are added to increase the robustness of the filter, and the statistical color feature method is adopted, which is not restricted by template features. The improved algorithm in this paper is tested on 28 vehicle-related video sequences in OTB and UAV data sets. The average distance accuracy is 80.8%, the average success rate is 82.7%, and the FPS reaches 58.24. Experiments show that the algorithm in this paper can improve the detection and tracking effect of vehicles in the drone scene, and can effectively solve the problems caused by the scale change and rapid deformation of the target vehicle. Compared with other nuclear-related filtering algorithms, it has better tracking accuracy and real-time performance.

This paper proposes a four-layer neural network based on randomly feature mapping (FRMFNN) and its fast incremental learning algorithms. First, FRMFNN transforms the original input features into randomly mapped features by certain randomly mapping algorithm and stores them in its nodes of first hidden layer. Then, the FRMFNN generates its nodes of second hidden layer using non-linear activation function on all random mapping features. Finally, the second hidden layer is linked to the output layer through the output weights. Since the weights of the first and the second hidden layers are randomly generated according to certain continuous sampling probability distribution, without the updates of the weights, and the output weights can be quickly solved by the ridge regression, avoiding time-consuming training process of the traditional back propagation neural networks. When FRMFNN can't reach the prescribed accuracy, its performance can be continuously improved by its rapid incremental algorithm, thereby avoiding retraining the whole network. This paper, a detail introduction of proposed FRMFNN and its incremental algorithms is provided. What's more, a proof of universal approximation property of FRMFNN is also given. Compared with broad learning system (BLS) and the incremental learning of extreme learning machine (ELM), the experimental results on several popular classification and regression datasets demonstrate the effectiveness of the proposed FRMFNN and its incremental learning algorithms.

Multi-view clustering is a type of multi-view learning method applied to unsupervised learning, which aims to use the feature set of different views to enhance the effect of clustering. Although multi-view clustering has been effectively applied in many fields, it still faces many challenges. For example, traditional algorithms only use visual information for clustering, ignoring the importance of hidden information. This paper proposes a multi-view fuzzy clustering algorithm (MVSH) that considers both visual and hidden information and feature weighting. This algorithm implements collaborative learning from various views under the framework of fuzzy clustering. On the one hand, personalized information is obtained by clustering features weighted under each view. On the other hand, the coefficient matrix shared by multi-view data sets is extracted by feature learning to obtain common (hidden) information. A multi-view learning with visual and hidden collaboration is realized. Using visual information and hidden information can better balance the common information and personalized information in the process of multi-view clustering collaborative learning. Experimental studies on multi-view data sets also effectively verify the above advantages of MVSH. Comparison with the performance of multiple related algorithms shows that the proposed method can achieve better or at least comparable performance. Experiments on multiple multi-view data sets prove that the strategy of merging visual and hidden information proposed in this paper has better effect than using visual and hidden information alone.

With the development of technologies such as Internet of things, cloud computing and artificial intelligence, big data becomes a research hotspot and is applied in many fields. The booming aviation field has natural big data soil, which has been paid more and more attention. In recent years, scholars have begun to study aviation oriented big data technology. Meanwhile airlines have also begun to use aviation big data to provide services for them, and promote it to the height of development strategy. Research and practice have shown that aviation big data can not only help to reduce the company??s operating costs, but also improve the quality of customer experience. In this paper, the definition of aviation big data is firstly given from the perspective of data and system, and the corresponding organization structures are described systematically. Secondly, the key technologies of aviation big data are elaborated in detail from five aspects: collection, storage management, preprocessing, analysis and virtual simulation and visualization, and some main models and algorithms are compared and analyzed. This paper describes the typical application scenarios of aviation big data from many aspects. Finally, the problems existing in aviation big data and the future research directions are analyzed in an in-depth way, so as to provide useful references for related research and applications.

As one of the most concerned technologies in the field of artificial intelligence in recent ten years, deep learning has achieved excellent results in many applications, but the current learning strategies rely heavily on a large number of labeled data. In many practical problems, it is not feasible to obtain a large number of labeled training data, so it increases the training difficulty of the model. But it is easy to obtain a large number of unlabeled data. Semi-supervised learning makes full use of unlabeled data, provides solutions and effective methods to improve the performance of the model under the condition of limited labeled data, and achieves high recognition accuracy in the task of image classification. This paper first gives an overview of semi-supervised learning, and then introduces the basic ideas commonly used in classification algorithms. It focuses on the comprehensive review of image classification methods based on semi-supervised deep learning framework in recent years, including multi- view training, consistency regularization, diversity mixing and semi-supervised generative adversarial networks. It summarizes the common technologies of various methods, analyzes and compares the differences of experimental results of different methods. Finally, this paper thinks about the existing problems and looks forward to the feasible research direction in the future.

Aiming at the problem that the population diversity of basic sparrow search algorithm decreases and it is easy to fall into local extremum in the late iteration, an improved sparrow search algorithm combining Cauchy variation and reverse learning (ISSA) is proposed. Firstly, this paper uses a Sin chaotic initialization population with an unlimited number of mapping folds to lay the foundation for global optimization. Secondly, this paper introduces the previous generation global optimal solution into the discoverer location-update method to enhance the sufficiency of global search. At the same time, the adaptive weight is added to coordinate the ability of local mining and global exploration, and the convergence speed is accelerated. Then, the Cauchy mutation operator and the opposition-based learning strategy are combined to perform disturbance mutation to generate new solutions at the optimal solution position, and the algorithm's ability to jump out of local space is enhanced. Finally, this algorithm is compared with 3 basic algorithms and 2 improved sparrow algorithms. Simulation and Wilcoxon rank and inspection are performed on 8 benchmark test functions. The optimization performance of ISSA is assessed, and time complexity analysis of ISSA is carried out. The results show that ISSA has faster convergence rate and higher precision than the other 5 algorithms. And the overall optimization capabilities are improved.

In recent years, attention has been paid to knowledge graph as auxiliary information to enhance recom-mendation increasedly. Since the goal of the knowledge graph learning task is to restore the relationship of the triples in the knowledge graph, rather than to accomplish the recommendation task, it is difficult for the knowledge graph learning task to efficiently help the recommendation task improve the recommendation performance. In addition, user's interest is easily affected by short-term environment and mood. This paper proposes a recommendation model that is a multi-task feature learning approach for knowledge graph and short-term preferences enhanced recommendation (MKASR) in response to the above two points. Firstly, the RippleNet algorithm is used to extract the relationship pairs between the user and the knowledge graph entity, and then these relationship pairs are stored in the form of knowledge graph triples for participating in training. The bidirectional GRU (gate recurrent unit) network based on the attention mechanism is adopted from the user's recent interaction sequence of items to extract the user's short-term preferences. Secondly, this paper uses the multi-task learning method to train the knowledge graph learning module and the recommendation module. And the feature representation between user and item can be obtained. Finally, these feature representations and the short-term preferences of users are taken into account to make comprehensive recommendations to users. The experiments on real MovieLens-1M and Book-Crossing datasets demonstrate that the proposed model has improved performance compared with other recommendation algorithms in AUC, ACC, Precision and Recall evaluation indexes.

Sleep disorders seriously affect human health and life. Accurate classification of sleep stages is the key to detecting and treating sleep disorders. In recent years, methods based on deep learning have surpassed traditional machine learning methods and human experts. However, the internal structure of deep learning is complex and requires to be designed by expert who is familiar with the computer and medical fields. This paper aims to analyze the relational induction biases in the existing sleep staging model based on deep learning, and explores the basic principles of sleep staging model design. This paper analyzes relational induction biases such as translation invariance, time invariance and hierarchical processing. Firstly, the model is divided into three categories according to whether it contains convolution layer with translation invariance and recurrent layer with time invariance: convolutional neural network framework, recurrent neural network framework and hybrid neural network framework. Secondly, according to the hierarchical processing method of the frame, segment and sequence in the model, a more detailed classification is carried out. Thirdly, it analyzes the impact of different relational induction biases in the model on the performance of sleep staging. It is proposed that a relational inductive bias matched the task should be introduced when the automatic sleep staging model is designed. Finally, the advantages and limitations of sleep staging based on deep learning are discussed, and it may be necessary to use more advanced relationship induction bias to express knowledge more abstractly and combine it with other artificial intelligence technologies in the future.

Citation recommendation (CR) is able to intelligently generate a paper list related to a query paper, which is of great value to researches. CR problem is related to papers?? semantic and structural information. Recently, network representation learning (NRL) based CR problem has gained extensive attention. However, existing studies all use single granularity networks to model CR, which have the disadvantages of high computational complexity and large memory consumption. For overcoming this challenge, this paper proposes a citation recommendation algorithm based on hierarchical attributed network representation learning (CR-HANRSL), which can greatly improve the efficiency of NRL while taking into account semantic and structural features of papers. First, the network is coarsened into a series of smaller networks based on papers?? attributes and author relationship repeatedly, and the super-nodes are made to contain child-nodes?? attributes after coarsening for further constructing semantic links. Second, this paper uses the single granularity NRL to learn the roughened network and graph convolutional network to refine the learned network representations. Finally, a multi-modal feature representation similarity between papers is generated to get a paper recommendation list. A mass of experimental results on AAN and DBLP two datasets show that the proposed method can improve NRL??s efficiency while learning high-quality network feature representations.

As the ability to collect and store data improving, real data are usually made up of different forms (view). Therefore, multi-view learning plays a more and more important role in the field of machine learning and pattern recognition. In recent years, a variety of multi-view learning methods have been proposed and applied to different practical scenarios. However, since most of the data points in the objective function have square residuals and a few outliers with large errors can easily invalidate the objective function, how to deal with redundant data becomes an important challenge for multi-view learning. For solving the above problems, this paper proposes a model, termed as robust auto-weighted multi-view subspace clustering. The model uses the Frobenius norm to deal with the squared error of data and uses the [?1]-norm to deal with outliers at the same time. Thus the effect of outliers and data points on model performance is effectively balanced. Furthermore, unlike traditional methods which measure the impact of different views by introducing hyper-parameters, the proposed model learns the weight of each view automatically. Since this model is a non-smooth and non-convex problem which is difficult to solve directly, this paper designs an effective algorithm to solve the problem and analyzes the convergence and computational complexity of this algo-rithm. Compared with traditional multi-view subspace clustering algorithms, the experimental results on multi-view datasets present the effectiveness of the proposed algorithm.

With the increasing amount of logs produced by nodes in CNGrid (China National Grid), traditional manual methods for user behavior analysis can no longer meet the need of daily analysis. In recent years, deep learning has shown good results in key tasks related to computer sciences, such as intrusion detection, image recognition, natural language processing and malware detection. This paper demonstrates how to apply deep learning models to user behavior analysis. To this end, this paper classifies user behavior in CNGrid and extracts a large number of user operation sequences bounded to sessions. These sequences are put into deep learning models. This paper proposes a deep learning model that combines recurrent neural network (RNN) with graph neural network (GNN) to predict the user behavior. Graph neural network can catch the hidden state of the user’s local behavior, so it can be used as preprocessing. Recurrent neural network can catch the message of time sequence. The model is built by combining GNN and RNN to acquire both advantages. In order to verify the effectiveness of the model, this paper conducts experiments of the real user behavior datasets on CNGrid and compares them with a variety of other methods. Experimental results demonstrate the effectiveness of this novel deep learning model.

Association rule mining is an important area in data mining, considering the large scale, high dimen-sionality, modal diversity and type complexity of current data, traditional association rule mining algorithms cannot meet the needs of big data. The particle swarm optimization algorithm, as an efficient intelligent algorithm, provides a new solution and has been widely used in association rule mining field in recent years. This paper introduces the basic principle of swarm optimization algorithm and the basic concept of association rules, and reviews the research progress of the swarm optimization algorithm itself. Then, this paper further summarizes the researches of the swarm optimization algorithm in association rule mining problem, including common data conversion methods, coding methods, and evaluation indexes. These improved algorithms from related researches are compared with other algorithms widely used in association rule mining, and their advantages, disadvantages, and application scenarios are discussed. After that, the existing improvement algorithms are systematically classified according to its methods, such as parameter, variation, and hybrid algorithm improvements, and the application areas of particle swarm optimization algorithms in association rule mining are also summarized, such as shopping baskets, finances, medical, industrial productions and risk assessments. At last, based on the introduction of the latest research pro-gress in this field, further research directions are discussed by analyzing the existing problems.

Aiming at the fluency problem of extractive method, the accuracy problem of abstractive method, and the important information missing problem caused by truncating the original document before document encoding, this paper proposes a two-stage long document summarization model SFExt-PGAbs. It is composed of submodular function for extractive summarization SFExt and pointer generator for abstractive summarization PGAbs. SFExt-PGAbs simulates the human process of summarizing a long document. First, SFExt is used to extract important sentences from the long document and filter the unimportant and redundant sentences to form a transitional document. Then, PGAbs receives the transitional document as input to generate a fluent and accurate summary. In order to get a transitional document that is closer to the original document-centered idea, this paper expands the two sub-aspects of positional importance and accuracy in the traditional SFExt, and designs a new greedy algorithm at the same time. In order to study the effect of different feature extractors on the quality of the generated summary, two kinds of recurrent neural networks are applied in PGAbs. The experimental results show that on the CNNDM test set, SFExt-PGAbs generates a more fluent and more accurate summary compared with the baseline model, and the ROUGE indicators are significantly improved. At the same time, the expanded sub-aspects of SFExt can extract more accurate summary.

Deep learning is a promising solution for computer vision at present. To solve the computer vision problem, it requires massive and high-quality image training datasets. Collecting and accurately labeling image datasets is a very time-consuming and expensive process. As computer vision applications become more widespread, it makes this problem even more pronounced. Image augmentation technologies are technical methods to effectively solve the problem of deep learning training under the condition of small-scale or low-quality training data. These technologies are continually accompanied with the development of deep learning and computer vision. This paper first reviews these image augmentation researches from the perspective of augmentation objects, operation spaces, label processing methods, and augmentation strategies and then concludes corresponding paradigms of current image data augmentation methods. After that, this paper proposes a taxonomy for current image data augmentation guided by the above paradigms, and reviews corresponding representative methods of each image data augmentation category. Finally, this paper makes conclusions on existing image data augmentation, points out the problems existing in the current image augmentation research and presents promising directions for future research.

ocal causal structure learning focuses on identifying the direct causes and direct effects of a given target variable without learning an entire causal network. Existing local causal structure learning algorithms are usually completed by two steps. Step 1 uses constraint-based methods to learn the Markov blanket (MB) or parents and children (PC) set of the target variable by conditional independence tests. However, due to small sample sizes, this may lead to unreliable tests and the accuracy of this step is usually not very high. Step 2 uses found V structures and Meek rules for distinguishing direct causes from direct effects of the target variable. But this step depends on the discovery of V structure extremely and synchronous sampling is also affected by limited samples. The accuracy of the algorithm is not very high. To solve the above problems, this paper proposes a hybrid local causal structure learning algorithm based on the combination of scoring and constraint. In step 1, a new PC learning algorithm SIAPC (score-based incremental association parents and children) is proposed by adding scoring idea into the constraint based algorithm. In step 2, the direction of the edge is determined by using the intersection of the orientation result obtained by PC algorithm and the orientation result obtained by grading some data sets, so as to reduce the dependence on V structure and alleviate the finite sample problem. After that, this paper uses independence test to modify the orientation results of the edges to further improve the accuracy of the algorithm, and then proposes HLCS (hybrid local causal structure learning) algorithm. Using benchmark Bayesian networks, the experimental results show that the algorithm proposed in this paper has better performance than the existing algorithms in terms of learning accuracy and reducing the data efficiency.

Visualization analysis of big data in education plays an important role in understanding and mining com-plex educational rules, which has become an important topic in the field of educational information science research. This paper first summarizes the typical characteristics of big data in education, introduces the latest research results in the application of big data in education from three aspects: promoting students?? metacognitive development, assis-ting teachers to supervise the learning process and improving managers scientific decision-making level, and describes the basic process of visual analysis using big data in education. Then, this paper focuses on five mainstream visua-lization presentation methods of big data in education, including text data visualization, multidimensional data vis-ualization, network data visualization, time series data visualization and geospatial data visualization, and gives specific application scenarios. Then it introduces the dynamic query and filter technology, scalable/deformable interface technology and multi-view linkage technology, three key interactive technology methods to implement the visuali-zation of big data in education. Finally, according to the latest research trends, the future research direction of big data visualization in education is prospected from four aspects: multi-mode education data fusion, man-machine interaction, man-machine collaboration paradigm, and the standard specification and evaluation system of visual design of education data.

The application of neural networks in software engineering has greatly eased the pressure of traditional method of extracting code features manually. Previous code feature extraction models usually regard code as natural language or heavily depend on the domain knowledge of experts. The method of transferring code into natural language is too simple and can easily cause information loss. However, the model with heuristic rules designed by experts is usually too complicated and lacks of expansibility and generalization. In regard of the problems above, this paper proposes a model based on convolutional neural network and recurrent neural network to extract code features through abstract syntax tree (AST). To solve the problem of gradient vanishing caused by the huge size of AST, this paper splits the AST into a sequence of small ASTs and then feeds these trees into the model. The model uses convolutional neural network and recurrent neural network to extract structure information and sequence information respectively. The whole procedure doesn't need to introduce the domain knowledge of experts to guide the model training and the model will automatically learn how to extract features through the codes which have been labeled classification. This paper uses the task of similar code search to test the performance of the trained encoder, the metric of Top1, NDCG and MRR is 0.560, 0.679 and 0.638 respectively. Compared with recent state-of-the-art feature extraction deep learning models and common similar code detection tools, the proposed model has significant advantages.

Video quality assessment (VQA) is based on the subjective quality assessment results of the human eye, using models to evaluate distorted videos. It is difficult for traditional assessment methods to make subjective assessment results consistent with objective assessment results. The VQA methods based on deep learning can be evaluated through independent learning of the model without adding manual features. It is of great significance to video quality monitor and assessment, and has become one of the research hotspots in computer vision. First, this paper introduces the research background and main research methods of VQA. Second, it introduces the VQA methods based on deep learning from both the full-reference and no-reference ones, and uses the convolutional neural network model to classify and compare the no-reference evaluation methods. Then it introduces the related databases and performance evaluation indices of the VQA models and compares the algorithm performance. Finally, this paper summarizes the existing problems in the current VQA researches and looks forward to the challenges and development directions in this field.

In recent years, Coflow scheduling has become a research hotspot in data center network. However, it is difficult for existing non-clairvoyant Coflow schedulers to infer the task information quickly. Therefore, small tasks cannot be scheduled in time, making it fail to minimize the average task completion time. Data center network requires effective inferring model to improve the accuracy and sensitivity in inferring Coflow size. This paper proposes a machine learning based Coflow size inferring model (MLcoflow), which utilizes an extreme learning machine (ELM) to establish Coflow size inferring model to minimize training error, and uses the incomplete infor-mation in training to increase the sensitivity. Experiment results show that the accurate score and sensitivity of ELM method are 19.8% and 10.2% higher than other algorithms on average, respectively. This paper compares several schedulers by simulation. MLcoflow-based scheduler reduces the average task completion time by up to 20.1%.

Aiming at the problems of low accuracy and low detection efficiency of high-resolution oil palm detection in deep learning, an effective and reliable solution is proposed from two aspects of algorithm optimization and heterogeneous hardware platform acceleration. Taking YOLOv3 object detection algorithm as an example, the optimization strategy of expanding feature selection and increasing multi-scale feature fusion is adopted to improve the detection accuracy of the algorithm for high-resolution oil palm. In addition, in the process of inference, plenty of applications require high performance models with strict power consumption limits. In order to solve this problem, taking the strategy of integer 8-bits quantitative weights and computational units reuse, this paper designs a high efficiency convolution computational engine based on SIMD. At the same time, through the strategy of the dimension change of the input image, vectorization, transmission to the input module in the form of written queue, this paper increases the efficiency of bus bandwidth greatly and accelerates the input module well. The experimental results show that the accuracy of the improved algorithm model is 97.84%, and a performance of 1.4 TOPS is obtained on the FPGA platform of Intel Arria 10. Compared with the i9-9980XE CPU, 7.51 times of the perform-ance and 33.02 times of energy efficiency are obtained. It is 1.2 times more efficient than Nvidia??s dedicated P40 accelerator.