Search

Aiming at the problem that the adaptive steganalysis of images is difficult, and the existing models are difficult to make a targeted analysis of favorable areas of images, this paper proposed an image steganalysis model SA-SRNet based on selfattention. The model introduced self-attention into SRN et to guide the model to focus on favorable areas for steganalysis in the global image and the long-range dependencies of image . It solved the problem that the training of hard-attention was easy to fall into local optimum. First, the reward mechanism used reinforcement learning to guide the model to find the detection point which was most favorable for steganalysis. Secondly, the self-attention used the detection point to generate the important attention image. Finally, the replace mechanism used the important attention images to replace the mis-classified images to improve the quality of training set and the discriminant ability of the model. This paper carried out the experiments on BOSSbase 1. 01 dataset. The results show that SA-SRN et can achieve better accuracy of steganalysis than that of SRNet, and can improve the accuracy by up to 1. 8% . [ABSTRACT FROM AUTHOR]

The paper proposes a robot zero calibration device and method based on spatial parameter clustering identification to overcome the shortcomings of traditional shaft pin and laser tracker calibration. The dial indicator and spatial analytic geometry are used to construct the zero calibration equation， which analyzes the robot’s theoretical post and posture values to obtain the zero-position training parameters. The clustering module identifies multiple sets of zero-position training parameters， and the optimal zero-position parameters of each joint can be solved. The experimental results show that spatial parameter clustering can improve the absolute accuracy by 72.9%， compared to the most commonly used shaft pin calibration method. Compared with the laser tracker calibration method with the highest accuracy， the time efficiency is improved by 87.5%. This method considers the calibration accuracy and efficiency and significantly improves the adaptability to the calibration environment. [ABSTRACT FROM AUTHOR]

The registration quality assessment is usually given to human experts, which is time-consuming. The commonly used Dice score only focuses on the error at the edge of the tissue, which is difficult to assess the registration result within the tissue. To address these issues, this paper proposed a method to predict registration errors based on machine learning (PREML) in lung CT images.This method firstly constructed three types of features, such as deformation field statistical features, deformation field physiologically realistic features and image similarity features, then expanded the number of features by pooling, and finally used random forest regression to predict non-rigid registration errors. Moreover, it used an adaptive random perturbation to simulate the spatial distribution of lung registration errors to further improve the capability of error characterization of statistical features. The proposed method achieved a mean absolute error of 1.245±2.500 mm from ground truth on lung CT image datasets, outperforming the baseline method. The results show that PREML method has the advantages of high accuracy and robustness, enhancing the safety and effectiveness of registration algorithms in clinical applications. [ABSTRACT FROM AUTHOR]

Traditional relational databases use the artificial way to achieve index recommendations that can no longer meet the growing requirement of data, and using machine learning technology can effectively solve the index optimization problem of databases. The limitation that only recommended the index in a static database and couldn't update the index configuration in time, this paper proposed a multi-attribute index intelligent recommendation approach (MIRA) to recommend multi attribute index configurations for a set of workloads using reinforcement learning under databases with dynamically changing data. Experimental results on TPC-H datasets show that MIRA can effectively recommend the optimal index configuration for workload, and outperforms the defined comparison baselines and related reinforcement learning methods. [ABSTRACT FROM AUTHOR]

In the environment of big data, traditional machine learning algorithms mostly use stand-alone offline training, which is obviously unable to adapt to the continuous growth of large-scale streaming data changes. Considering the problems above, this paper proposed a distributed online integrated learning algorithm based on Flink platform, which was based on Flink distributed computing framework. The algorithm firstly performed distributed online training on the online learning algorithm through data parallelism. And then the trained multiple submodels were dynamically weighted allocation of the model through stochastic gradient descent algorithm to realize the results aggregation of multiple submodels. At the same time, the samples of the models with poor training effect were updating online. Finally, by comparing the performance of stand-alone and cluster environments on different datasets, the experimental results show that the online learning algorithm combined with the distributed ensemble training of the Flink framework can achieve the performance in the centralized training mode, and greatly improve the time efficiency of training. [ABSTRACT FROM AUTHOR]

The classification performance of the existing multi-classification tasks of finger motor imagery EEG signals is difficult to achieve usable performance. On the basis of detailed analysis of multiple components on the time scale of EEG signals, this paper designed a self-supervised sub-network for signal sub-segment extraction, and then input the sub-segment into the next sub-network for signal classification, it synthesized the two sub-networks into a self-supervised hybrid multi-task deep network. In the training phase, the sub-segment extraction sub-network extracted different sub-segments for each EEG signal, and the subsequent classification sub-network judged whether the sub-segment was the best and automatically adjusted the position of the sub-segment. It weighted the total loss function by two loss functions of two sub-networks, and extracted the best sub-segment signal and obtained the best classification effect through the overall network learning algorithm. In the verification and testing phase, the sub-segment extraction sub-network automatically extracted the corresponding sub-segment input classification sub-network according to the parameters of the training for classification. It verified the network performance on the largest SCP data of Motor-Imagery data set and the Data sets 4 of BCI Competition IV. On the SCP dataset, the average test classification accuracy of all subjects’ three finger classification tasks is more than 70%, the average test classification accuracy of four finger classification tasks is about 60%, and the average test classification accuracy of five finger classification tasks is about 50%, which is significantly improved compared with the existing reports. It is confirmed that the network can effectively extract sub-segments of motor imagery EEG signals, and has good classification effect and generalization performance. [ABSTRACT FROM AUTHOR]

Learning causal structure from observed data has important applications. An existing method for learning causal structure is to learn causal structure by testing the independence between noise and causal variables under the functional causal model assumption. However, such methods often involve highly computationally complex in process of testing independence, which affects the practicability and robustness of the structure learning algorithm. To this end, this paper proposed a causal structure learning algorithm that used higher-order cumulants as independence scores under a linear non-Gaussian model. The algorithm was mainly divided into two steps. The first step was to use the method based on conditional independence constraints to learn the Markov equivalence class of the causal structure. The second step was to define a score based on high-order cumulants, the score could determine the independence of two random variables so that the causal structure of the best independence score could be searched from the Markov equivalence class as the output of the algorithm. The advantages of this algorithm were: a)Compared with the independence test based on the kernel method, the method had lower computational complexity. b)The method based on score search could always obtain a model that best matches the data generation process, which improved the robustness of the learning method. Experimental results show that the high-order cumulant-based causal structure learning method improves the F1 score by 5% in synthetic data and learns more causal directions in real data. [ABSTRACT FROM AUTHOR]

Federated learning is designed to solve the problem of data fragmentation and data isolation based on privacy protection in distributed machine learning. In the federated learning system, participants collaboratively train a model. Each participant uses local data to train the local model, and uploads the trained local model to the server for aggregation. In the real application environment, the data distribution between nodes is often very different, resulting in the accuracy of federated learning model is low. In order to solve the influence of non-independent identically distributed data on the accuracy of the model, this paper proposed a clustering federated learning framework by using the similarity of data distribution between different nodes. Extensive experiments were conducted on Synthetic, CIFAR-10 and FEMNIST standard datasets. Compared with other federated learning methods, clustering federated learning based on data distribution greatly improves the accuracy of the model and requires less computation. [ABSTRACT FROM AUTHOR]

【Objective】 Aiming at the problems of easy falling off and easy to cause pig infection in traditional pig breeding ear tag recognition, the paper used the improved YOLOv5s model to carry out non-invasive recognition of pig face.【Method】 Firstly, the distance of K-Means was changed to 1-IOU, which improved the adaptability of the target anchor frame of the model; Secondly, the CA coordinate attention mechanism was introduced to improve the model feature extraction; Finally, BiFPN feature fusion was introduced to effectively use the features to improve the detection ability of the model. The pig face dataset used in the experiment was divided into 5 categories, with 12 756 samples after data enhancement, and the training set and test set were divided into a ratio of 9∶1.【Result】 The improved algorithm reached 0.926, 0.897 and 0.955 in accuracy, recall, and average accuracy (IOU=0.5), which were 13.2%, 3.0% and 2.2%, while the improved algorithm has better generalization ability and high recognition accuracy in single, multiple, small target, dense, and occluded scenes. 【Conclusion】 Using deep learning algorithm can get the facial information of pigs and identify them accurately, reduce the situation of missed inspection and wrong inspection, and provide better technical support for intelligent management of pigs. [ABSTRACT FROM AUTHOR]

In a multiple-services mobile edge computing network environment, the requirements and popularity of services, the computing requirement and the amount of data transferred from users to edge servers are dynamic with time. How to adaptively adjust the caching subset of total service types in the resource-limited edge server and determine the task offloading destination and resource allocation decisions to obtain the optimum overall system performance is a challenge problem. To solve this challenge, this paper firstly converted it into a Markov decision process, then proposed a soft actor-critic (SAC) -based deep reinforcement learning algorithm to jointly determine not only the discrete decisions of service caching and task offloading but also the continuous allocation of bandwidth and computing resource. The algorithm employed an effective trick of converting multiple continuous action outputs into discrete action choices to deal with the key design challenge that arised from continuous-discrete hybrid action space while improving the accuracy of algorithm decisions. Additionally, the algorithm integrated an efficient reward function that added auxiliary reward terms to improve resource utilization. Extensive numerical results show that compared with other baseline algorithms, the proposed algorithm can effectively reduce the long-term average completion delay of tasks while accessing excellent performance in terms of stability. [ABSTRACT FROM AUTHOR]

LLP is a machine learning method that puts instances into packages. It only provides instance information and label proportional information in packages, but does not provide label information. Aiming at the LLP problem of multiple related tasks, this paper proposed a label proportion integrated learning model based on transfer learning, called AT-LLP for short. The model connected related tasks by constructing shared parameters between tasks, and migrated the knowledge learned in the source task to the target task, so as to improve the learning efficiency of the target task. At the same time, the algorithm introduced the ensemble learning algorithm, which continuously adjusted the weight coefficient of the training set in the multi-iteration learning process of the classifier, so as to further train the weak classifier into a strong classifier. Experiments show that the proposed AT-LLP model has better performance than the existing LLP method. [ABSTRACT FROM AUTHOR]

Local path planning for unmanned surface vehicle ( USV) plays an important role in maritime rescue and marine transportation. Existing local path planning algorithms achieve good results in simple scenarios, but have poor performance when facing complex obstacles and sea current disturbances present in the environment. To this end, this paper proposed a reinforcement learning algorithm based on spatial and temporal sensing-enhanced deep Q-network. Firstly, it introduced a multiscale spatial attention module to capture the multiscale spatial information of distance sensors, which enhanced the perception capability of complex obstacle environments. Secondly, it used the LSTM-based current sensing module to extract the temporal sequence features of the current disturbance environment, which enhanced the perception capability of the current disturbance. In addition, by simulating the sensor and motion model of USV, it designed the reinforcement learning state space, action space and direction-guided reward function, it improved the navigation performance and convergence speed of the algorithm. Simulation experiments in complex scenarios show that the proposed algorithm improves both success rate and average arrival time metrics comparing to the original algorithm, and the algorithm shows strong adaptability to complex environment. [ABSTRACT FROM AUTHOR]

In recommender systems, graph convolutional networks have stronger information extraction capabilities for graphstructured data. However, existing graph convolutional networks mainly focus on enhancing the model structure, ignoring the importance of improving the sampling quality of the original samples and mining the implicit relationship between users and items. Aiming at the above problems, this paper proposed a graph contrastive learning recommendation algorithm based on mixed sampling. Firstly, the algorithm used a mixed sampling method to extract part of the information in positive samples and injected them into negative samples, thereby generating new informative hard negative samples. Secondly, to extract features on hard negative samples, it used the light graph convolution network to obtain node representations of users and items. Finally, it carried out a multi-task strategy to jointly optimize the recommendation supervision task and the contrastive learning task. The experiments on real datasets Yelp2018 and Amazon-book demonstrate that the proposed algorithm improves performance compared with other recommendation algorithms in recall and NDCG evaluation indexes. [ABSTRACT FROM AUTHOR]

Directed evolution can increase the rate of enzyme evolution by mimicking the natural evolutionary process and has become a key technology for enzyme engineering. Directed evolution has played an important role in biocatalysis and drug design, however the experimental screening is in great challenge due to the large number of mutant libraries caused by the randomness of mutations. In recent years, emerging technologies such as artificial intelligence and big data processing have also become crucial in biocatalysis researches. Machine learning methods are statistical learning approaches to obtain sequence/structure mappings to enzyme function in a data-driven manner, which will improve the efficiency of enzyme engineering. This paper reviews the state-of-the-art technologies involved in machine learning models, especially focusing on the research and application progresses of machine learning methods in enzyme engineering. With the advancement of machine learning algorithms and technologies, it is expected that more accurate and effective models will be proposed in the future to promote screening of new enzymes and accurate design of biocatalysts [ABSTRACT FROM AUTHOR]

Displacement prediction has been one of the hot spots in landslide studies. In recent years, with the development of computer science, artificial intelligent models have been increasingly used in displacement prediction of landslide. Compared with the commonly used machine learning methods like LSTM neural network, the XGBoost（eXtreme Gradient Boosting） model in integrated algorithms is not common in landslide displacement prediction. However, it has become a hot field of research in academia because it has higher prediction accuracy and faster operating speed in the prediction. Taking the monitoring data of the landslide disaster site in Yaoshan Village, Anxi County, Quanzhou, Fujian as an example, Python is used to build the XGBoost model in this paper. By comparing the maximum information coefficients, the feature with a higher correlation to the displacement is selected and put into the model for landslide displacement prediction. The results show as follows: the XGBoost model can predict rainfall-induced step-like landslide displacement more accurately than most machine learning models. This is because it has advantages such as introducing a regular term in the objective function to control model overfitting, and using forward validation for model dataset partitioning. Meanwhile, it is of important significance for the displacement prediction, monitoring and early warning of such landslides. [ABSTRACT FROM AUTHOR]

With the supervised deep learning algorithms, it is prone to generate artifacts when restoring the blurred solar speckle images taken by Yunnan Observatories, and it has a long training time and over-reliance on reference images, this paper proposed a weakly supervised method based on structural reparameterization combined with multi-branch module to reconstruct solar speckle images. First, deblurring model combined single-scale and multi-scale network to design, with constructing multi-branch modules to extract features of different scales, enhance detailed information, and reduce the generation of artifacts; second, each branch structure re-parameterized to make the reuse of structure parameters runs through the entire feature extraction process; after that, the deblurring model embedded in the weakly supervised training, the blurred image assorted firstly, then the degradation model used to learn different levels of degradation. Constituted paired dataset of corresponding levels, and the deblurring model used to inversely degenerate the dataset to reconstruct solar speckle images. Experimental results show that compared with the existing deblurring method, the proposed method has higher model training efficiency and less dependence on reference images, which can meet the high-resolution reconstruction requirements of solar speckle images. [ABSTRACT FROM AUTHOR]

As a key problem and hub link between problem discovery and problem solving, the current research on root cause analysis mainly includes two types of methods: data-driven and causality-driven. In view of the advantages of data-driven methods in narrowing the range of root causes, at present, root cause research mainly focusses on datadriven methods such as association rule mining, heuristic search, machine learning and deep learning. There are few root cause analysis from the perspective of causal knowledge, nor has it been based on method dimension. The inductive analysis of the root cause is carried out, and there is a lack of relevant research results. Therefore, this paper sorts out and summarizes the main results of root cause analysis in recent years, analyses the differences and advantages of root cause analysis under different method dimensions, and proposes a root cause analysis method that integrates causal knowledge, and combines asymmetric Shapley value with causal chain diagram to improve the accuracy of root cause analysis. Finally, the existing some research difficulties and development trends put forward scientific and meaningful future research directions. [ABSTRACT FROM AUTHOR]

Aiming at the existing multi-modal learning which based on attention mechanism,this paper studied the self-association between the context of the text and the spatial positional relationship of the object area of the image.Based on the analysis of existing attention networks,this paper proposed to use SA and SRA to map the text information to the image object,and finally obtained the fusion feature output.Compared with other attention mechanisms,SA and SRA can better match text information to the image object area.The model is trained and verified on the VQAv2 dataset and achieves an accuracy of 64.01% on the VQAv2 dataset. [ABSTRACT FROM AUTHOR]

Generate a brief and short summary for a document can effectively alleviate the reading pressure brought by information explosion.In recent years,Seq2 Seq model has been widely applied in various text summarization.Especially,Seq2 Seq with attention model has become a basic framework for abstractive text generation.To make the generated summary reflect the abstractive feature,this paper proposed a novel framework,which employed Seq2 Seq model with attention and coverage mechanism as basic model,and used VAE to depict the latent feature in decoding process.At last,this paper used a pointer-generator network for relieving the out-of-vocabulary(OOV) problem to generate the summary tokens successively.Experimental results on the LCSTS dataset demonstrate that the proposed framework is able to capture the latent feature of summary text and relief the OOV and tokens repetition problem,thus generating more accurate and readable summary. [ABSTRACT FROM AUTHOR]

In order to understand the latest research progress of TransE based representation learning methods in real time,this paper classified TransE based representation learning methods into four types:the method based on complex relationship,the method based on relationship path,the method based on image information,and the method based on other aspects.Then,this paper analyzed the design ideas,advantages and disadvantages of each method.At the same time,it compared and summarized the common data sets and evaluation indexes of the TransE based representation learning method,as well as the performance of various TransE based representation learning algorithms in the experiment.From the research results,PaSKoGE,NTransGH,TCE and TransD method perform the best in link prediction and triple classification tasks,which are worth promoting and further expanding,and can be further improved in path specific embedding,two-layer neural network,triple context and dynamic mapping matrix construction. [ABSTRACT FROM AUTHOR]

Aiming at the problem of inconsistent probability distribution between training samples and test samples in machine learning, this paper proposed a semi-supervised domain adaptation method based on dropout regularization to transfer neural representations from label-rich source domains to unlabeled target domains. From the perspective of semi-supervised learning, this paper added a small amount of labeled target domain data into the source domain data, so that the neural network could learn the feature distribution of target domain data while learning the feature distribution of source domain data. With the guidance of prior knowledge, neural network could fit target domain data well even without abundant label information. Experiments show that the proposed algorithm is superior to other existing algorithms in domain adaptation performance on digital datasets and robust on real datasets covering a wide range of natural categories. [ABSTRACT FROM AUTHOR]

How to select features on semi-supervised data sets by incomplete supervisory information has become a research hotspot in the field of pattern recognition and machine learning. In order to facilitate researchers to systematically understand the research status and development trend of semi-supervised feature selection, this paper reviewed the semi-supervised featu re selection methods. This paper first discussed the classification of semi-supervised feature selection methods and divided them into graph-based methods, pseudo-label-based methods, SVM-based methods and other methods according to their theoretical basis, then introduced and compared typical methods for each category, and then sorted out hot applications of semi-supervised feature selection. Finally, this paper looked forward to the future research directions of semi-supervised feature selection. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the existing super-resolution algorithm based on deep learning didn't make full use of the feature information of each level, resulting in low reconstruction accuracy and large parameter quantity, this paper proposed a double dense connection structure named densely channel attention skip connection network. In the inner structure of the network, it improved the original dense cascade block to generate a channel separable dense cascade block. The outer structure adopted a densely residual connection and attention mechanism to fuse the features extracted by the dense block to achieve the goal that less convolution layer and higher precision effect. This paper tested the network models on several benchmark datasets . The results show the proposed model has higher accuracy and fewer parameters than the other models. [ABSTRACT FROM AUTHOR]

At present, multi-agent reinforcement learning algorithms mostly adopt frameworks that are centralized in learning and decentralized in action. These frameworks may take too long to converge or may not converge at all. In order to speed up the collective learning time of multi-agents, this paper proposed a novel multi-agent group learning strategy. It used recurrent neural network (RNN) to predict the grouping matrix of multi-agents to share the experience between them, resulting in improved learning efficiency within the multi-agents group. Meanwhile, this paper proposed the concept of information trace to remedy the problem that the agents could not share information brought by the grouping. In order to strengthen the retention of excellent experience within the group, this paper proposed the practice of delaying the death time of excellent agents in the group. Finally, the results show that, compared to MADDPG, the training time is reduced by 12% in the labyrinth experiment and by 17 % in capture the flag experiment. [ABSTRACT FROM AUTHOR]

In order to improve the performance of frame disambiguation model, this paper used a neural network model to learn frame representation based on sentences in corpus different from the traditional classification algorithm extracting features manually, and employed the learned frame representation on the frame disambiguation task, which significantly improved the performance of the task. Making full use of the CFN example sentence database and being based on the hinge-loss neural network, the algorithm learnt the frame representation that could distinguish the correct frame from the error frame in the largest degree. This paper also used the WSABIE algorithm to learn the representation vector of the context of the target word, and finally used the cosine distance between the context representation vector and the frame representation vector to make a decision for the task. Experiment performed three sets of two-fold cross-validation (3 x 2 BCV) on 88 ambiguous words in CFN, and the best accuracy of frame disambiguation reach to 72. 52%. The t-test results show that the performance of the proposed method is significantly higher than other frame disambiguation methods. [ABSTRACT FROM AUTHOR]

Mobile traffic classification/ clustering is an important foundation for mobile network traffic management. However, the mobile network traffic data used by different papers were collected from different network environment. In addition, the labels and the flow statistical features of mobile traffic were different from papers. These experimental results couldn't be directly compared. This paper collected the traffic data generated by App based on MobileGT system. The two kinds of labels were built on these data (App level and function level),and two kinds of flow statistical features were independently extracted on these traffic data. This paper comprehensively researched the machine learning techniques on the traffic data with different labels and different flow statistical features. The experimental results show that the uni-direction flow based features are better than bidirection flow based features, random forest and AdaBoost are better on classifying mobile traffic data, and K-means is better on clustering mobile traffic data. [ABSTRACT FROM AUTHOR]

Based on video action recognition, this paper firstly reviewed the traditional action recognition algorithms based on RGB, including traditional algorithms and deep learning algorithms. Because of the influence of background lighting, RGB-videobased action recognition is low accuracy, but it has rich color appearance information. Then this paper analyzed and summarized action recognition algorithms based on RGB-D, which was divided into depth sequence, bone and multi-feature fusion. RGB-D video has multiple modes and can provide more information for action recognition, which performs better than RGB video but also brings new challenges. Finally, this paper prospected the common data sets and the possible future development direction. [ABSTRACT FROM AUTHOR]

Abnormal event detection is a hot topic in the field of intelligent surveillance monitoring technology research currently. As an important research content of computer vision, its main goal is to use computers to automatically detect abnormal events. Traditional methods have limitations in the weakness of low-level video feature description ability, high computational cost of anomaly detection methods and poor robustness in modeling complex scenes. In recent years, how to design a high-level semantic feature extraction method, accelerate the process of abnormal event detection, and model complex scenes such as multiple cameras have become the forefront topic of current research. Based on the research situation at home and abroad and the mainstream methods, this paper introduces the basic techniques involved in abnormal event detection in surveillance videos, and analyzes the advantages and disadvantages of various types of surveillance video feature extraction methods， feature learning models and anomaly detection methods. This paper also summarizes the commonly used benchmark datasets that can be used to abnormal event detection in surveillance videos. Finally, we discuss the difficulties, challenges and future development trends of abnormal event detection in surveillance videos. [ABSTRACT FROM AUTHOR]

With the advancement of data acquisition technology, spatiotemporal data with geographic location information are growing rapidly, so it is urgent to explore effective spatiotemporal data modeling methods. Spatiotemporal sequence prediction is one of the basic methods in spatiotemporal data modeling and widely used in many fields. At present, there is no Chinese literature on its review, so it is of great significance to summarize these methods. For the spatiotemporal sequence prediction problem, this paper reviewed its application background and development history. Then it introduced some related definitions and characteristics. According to its category, the spatiotemporal sequence prediction methods, based on traditional methods, classical machine learning methods and deep learning methods are analyzed respectively as well as their application scope, advantages and disadvantages. Finally, this paper made a review and prospect over the research directions of spatiotemporal sequence prediction, laying a theoretical foundation for researchers to further study this problem [ABSTRACT FROM AUTHOR]

Most existing class incremental learning methods employ data storage or extended network structures, but they cannot effectively alleviate the catastrophic forgetting problem due to memory resource limitation. To solve this issue, this paper proposed a brain-inspired generative replay approach. Firstly, it used VAE-ACGAN to simulate the memory self-organizing system to improve the quality of the generated pseudo-samples. Then, it used a shared parameter module and a private parameter module to protect the extracted features. Finally, the potential variables in the generator used a Gaussian mixture model to select specific replay pseudo-samples. The experimental results of the proposed method on MNIST, Permuted MNIST and CIFAR-10 show that the classification accuracy of the proposed method is 92.91%, 91.44% and 40.58%, respectively, which is significantly better than other class incremental learning methods. Furthermore, the backward transfer and forward transfer metrics reach 3.32% and 0.83% on the MNIST dataset, demonstrating that the model achieves a trade-off between task stability and plasticity, effectively preventing catastrophic forgetting. [ABSTRACT FROM AUTHOR]

Fruit picking has been one of the most important operations in modern agriculture. Manual picking cannot fully meet large-scale apple production in recent years, due to the labor-intensive and time-consuming task. Alternatively, robot picking can be expected to replace manual work. However, the conventional semi-automatic picking machine can seriously vibrate the branches and damage the fresh apples during batch harvesting in the existing orchard. It is a high demand to detect apples under complex environments, and then pick them without damaging the fruit or trees. The picking robots are also required a compatible auxiliary platform during operations in an orchard. The performance of picking robots can be assessed using the picking speed, picking success, and fruit damage rates. These parameters depend on three key technologies: target recognition, localization, and fruit separation. In this paper, an overview was provided of the history and current status of commercial apple-picking robots. Specifically, the performance and robustness were highlighted, in terms of the picking efficiency, structural composition, the type of orchard, and adaptability to the external environment. The main technical features were then summarized from the application of apple-picking robots. Since the apple picking was operated in an unstructured orchard environment, the most serious challenge remained in the application of target recognition, localization, and fruit separation techniques. Therefore, the recognition techniques were reviewed for the apples in the complex natural environments with changing light, branch and leaf shading, overlapping fruit, night-time environments, and same-colored apples. Deep learning-based apple recognition algorithms were the far-reaching prospects in complex environments, particularly in multiple scenarios. The accuracy and speed of apple recognition were improved using deep learning recognition algorithms in the field of apple picking. The second most challenge was the application of the apple localization for the fruit occlusion and overlap, while the fruit displacement by wind during the imaging process. Therefore, four aspects were introduced in the apple localization, including 2D information acquisition, depth information acquisition, apple pose acquisition and oscillating fruit localization. Point cloud processing algorithms were highlighted to achieve the precise localization of occluded apples. In addition, the applications of end-effectors were reviewed in the apple-picking robots with high efficiency and non-destructive picking. Three types of the end-effector structure (such as hand-jaw, vacuum, and shear) were summarized under fruit separation. It was found that the hand-jaw was the most commonly-used type of end-effector. Current research on the end-effectors was focused on the flexible materials for the finger surfaces, under-driven, flexible structures or control systems, such as the soft hands and perception-based servo control. Therefore, it is necessary to improve the picking speed and success rate with the low fruit damage and cost saving in the apple picking robot, in order to accelerate the process of apple picking automation. Future development was also proposed for the apple-picking robot in the urgent commercial need of the pioneering innovation in the combination of agricultural machinery and agronomy. Some optimization can be expected to be performed on the recognition algorithm, multi-sensor fusion, multi-arm cooperation, human-machine collaboration, extending device versatility, and integrating 5G and IoT technology. This finding can provide a strong reference for the current status, future development trends, and equipment innovation of apple-picking robots, together with the key technologies. [ABSTRACT FROM AUTHOR]

The traditional landslide identification is mainly through remote sensing visual interpretation and human-computer interaction identification, which has the problems of time-consuming and laborious, subjective and low extraction accuracy. In this paper, a highway in Fujian Province is taken as the experimental area, and a rapid landslide identification method based on high-resolution remote sensing data for mountainous roads in southeast Fujian is proposed. Firstly, by analyzing and studying landslide features such as hue, topography, spectrum, vegetation index and texture using sub-meter high-resolution remote sensing images, the multi-dimensional multi-scale landslide classification and identification rules applicable to the mountainous roads in southeast Fujian are established. Secondly, a road landslide identification model is constructed by combining the classification tool of machine learning algorithm with the multi-dimensional multi-scale feature screening set, and the preliminary identification of landslides based on high-resolution remote sensing images in terms of hue. Finally, the landslides are further segmented and extracted by slope, normalized vegetation index and texture feature screening set, so as to accurately identify the spatial distribution of landslides on mountainous roads. The average accuracy of landslide identification by this method reaches 85.73%, and the morphological features extracted from the landslides are complete, which can clearly show the "tongue" and "dustpan" shape of the landslides, and also clearly identify the slide walls and piles of the landslides. The research results can provide scientific reference for landslide identification and risk assessment of mountainous traffic routes in vegetation development areas in China. [ABSTRACT FROM AUTHOR]

There is dynamic complexity of preference among members in group recommendation, and the preferences of group members will affect each other. As for these questions, this paper proposes a group recommendation algorithm that combines the context information and self attention mechanism. Specifically, first of all, in order to better capture the dynamic relationship among group members, the author used self attention mechanism to automatically learn the context information of users in the group and the relationship weights among other members in the group; Secondly, according to the group matrix learned from the self attention mechanism, the author embedded the average value to obtain the group preference expression. Finally, according to group preference, the author applied the Bayesian personalized ranking method into the test for group recommendation. During the research, the author used the above mentioned method to test on two datasets, namely Movielens-1 M and CAMRa2011. As a result, compared with the benchmark algorithm, this method has improved the evaluation indexes HR and NDCG. And the experimental results show that this method has achieved better effect in group recommendation. [ABSTRACT FROM AUTHOR]

Relation detection for knowledge base question answering (KBQA) aims to select the best relation path that matches the question expressed by natural language from the candidate relations in the knowledge base and retrieve the answer to the question. To solve the problem of semantic information loss and inadequate attention interaction in existing relation detection methods, this paper proposed a multi-granularity relation detection model incorporating global-local features. The model firstly used bi-directional long short-term memory (Bi-LSTM) networks to encode the question and relation and modeled relations from multiple granularity, such as word-level and relation-level representation. Then, the model introduced a bi-directional attention mechanism to implement attentive interaction of the question and relation. Finally, it extracted global and local features by the aggregation operation and word-level interaction, respectively, to calculate the semantic similarity of the question and candidate relation. Experiments show that the proposed model achieves 93.5 % and 84.13% accuracy on SimpleQuestions and WebQuestionsSP datasets, respectively, improving the effect of relation detection. [ABSTRACT FROM AUTHOR]

At present, the congestion of highways in China is severe. Traffic flow prediction plays an important role in the intelligent transportation system. If it can achieve high-precision prediction, it will be able to efficiently manage traffic and alleviate congestion. To solve this issue, this paper proposed a multi-channel traffic flow prediction method (MCST-Transformer)considering spatiotemporal correlation. Firstly, it used Transformer to extract the internal laws of different data, and then introduced a spatial correlation module to mine the association features of different data. Finally, it integrated global information through channel attention. Using the data of Guangdong province highway, the proposed method realized the traffic flow prediction of 92 toll stations within two hours with high precision. The results show that MCST-Transformer is superior to traditional machine learning methods and time series models based on the attention mechanism. Under the prediction horizon of 120 min, MAPE decreases by 5. 1 % compared with Bayesian regression. Compared with deep learning algorithms like Seq2Seq-Att and Seq2Seq, the overall MAPE of the proposed method can also be reduced by 0. 5%. It indicates that the multi-channel approach can distinguish the characteristics of different data, so as to acquire better performance. [ABSTRACT FROM AUTHOR]

In order to resolve the problems of large scale change, susceptibility to occlusion and similar object interference when target tracking by UAV, based on SiamCAR,this paper proposed IMPSiamCAR algorithm. The algorithm used the improved ResNet50 network to extract the target features, introduced the channel attention mechanism to make the model learn the semantic information of different channels, and assigned different weights to channels according to the importance of the features, so that the algorithm could pay more attention to the regions where the tracking targets exist. Then it fed the fused target features into the region regression network for positive and negative sample classification, centrality calculation and bounding box regression. Finally, it obtained the position of the target in each frame. Test experimental results of UAV123 and OTB100 datasets show that the proposed algorithm has higher tracking accuracy and success rate compared with the comparison algorithm, and can better deal with challenges such as occlusion, similar object interference as well as scale change. Moreover the results of real-time testing on VOT2018 and UAV123 datasets show that the proposed algorithm can fulfill the UAV real-time request. [ABSTRACT FROM AUTHOR]

Memory network only focusing on attention mechanism can't tackle this situation where the context sentiment relies on the concrete target. In order to address this problem, this paper proposed a target-sensitive deep memory network model. This model exploited attention mechanism to capture the sentiment information of context corresponding to a given target. Then the interaction module integrated context sentiment expression and interaction information between context and target to the classification features. Finally, it classified the features to get sentiment polarity of target. This paper conducted experiments on the two datasets from SemEval 2014 task4. The proposed model achieved better performance than attention-based memory networks. The experimental results show that considering the interaction between context and target is effective to solve the problem that context sentiment relies on the concrete target. [ABSTRACT FROM AUTHOR]

In order to enhance the network representation learning quality with the edge signed semantic information, focusing on the weakness of existing fusion methods in dealing with complex edge signed semantic information, this paper proposed a network representation learning algorithm incorporating with edge signed semantic information, and introduced the edge signed semantic information containing positive and negative relations into the network representation learning process. Firstly, this paper designed a relationship prediction model based on the three-layer perceptron to depict different types of context link relations between nodes. Then it introduced the random walk strategy to implement context link sampling to adapt the large-scale network scenarios. Experiments on three data sets show that this algorithm can effectively model different types of context links between nodes and mine the complex semantic information contained in them. Compared with the current optimal SIDE method, the performance of the proposed algorithm is improved by 0. 31 %, 1. 3 % and 1. 85 % . [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]

This paper proposed an uncalibrated 3 D robotic arm grabbing method based on improved YOL0v2 in a multi-object environment. Firstly, in order to reduce the depth learning algorithm YOL0v2 detection multi-obj ect bounding box overlapping rate and 3D distance calculation error, it proposed an improved algorithm for YOL0v2. It used this algorithm to detect and identify the target object in the image, obtained the pos ition information of the target object in the RGB image, and then used the K-means + + clustering algorithm to quickly calculate the distance from the target object to the camera according to the depth image information, and estimate the target object size and pose. Simultaneously, it used the improved YOL0v2 to get the bounding box of the gripper and calculated the distance from the robot to the target object. Then the system estimated the distance between the fixture, camera and obj ect in the manipulator coordinate system. Finally, the system used the PID algorithm to control the gripper to grab the object according to the size and posture of the object and the distance from the object to the gripper. The detected boundary boxes of the target obj ect was more accurate based on the improved YOL0v2 than on old one. It also enhanced the distance from the fixture to the object and the size of the object as well as the accuracy of the pose estimation. In addition, in order to avoid complicated calibration, this paper proposed a non-calibration method. The learning scheme was different from the traditional uncalibrated estimation method based on Jacobian matrix, because it had good universality. A simulation experiment shows that the proposed method can accurately classify and locate the objects in the image. The Universal Robot 3 robotic arm uses this framework to verify the effectiveness of capturing objects in a cluttered environment. [ABSTRACT FROM AUTHOR]

This paper proposed constrained deep reinforcement learning( CDRL) to compute the fastest route online using taxi drivers' experience in different time period. Firstly, this paper described the extraction of experiential road segment database (ERSD). Then it introduced CDRL method, which mainly comprised of two phase: bounded condition of route and deep Q-learning algorithm. In the first phase, the task was to generate alternative constrained road segments of OD pair. In the second phase, it devised deep Q-learning algorithm to learning the experience of taxi drivers, and computed the fastest route of a given OD according to their departure time. Lastly, this paper tested an empirical studies in CBD of Guangzhou. The results show that the routes computed by CDRL method is approximately equal to shortest route ( SR) and fastest route (FR) method in travel time and route length. Furthermore, the CDRL method notably outperforms FR and SR in computing efficiency, so it is more suitable for online fastest route computation. [ABSTRACT FROM AUTHOR]

The data sources of existing link prediction algorithms are mainly based on neighbors, paths, and random walk methods. The link prediction algorithms use mainly node similarity assumptions or maximum likelihood estimates. The link prediction based on neural network is still absent. Some research achievements based on neural network show that the Deep Walk algorithm based on neural network is an efficient network representation learning algorithm, which can more effectively learn the network structure features in the network. It has been proven that DeepWalk is equivalent to factorize the target matrix. Therefore, this paper proposed a link prediction algorithm (LPMF) based on matrix factorization of Deep Walk. This algorithm based on matrix factorization used the Deep Walk algorithm to get the network representation vectors. And then, it calculated the similarities between node pairs of nod es by the cosine similarity method. Based on that, the similarity matrix of the target network was constructed. Finally, this paper used the similarity matrix to conduct the link prediction experiments on three real -world citation networks. The experimental results show that the new method is superior to the existing 20 kinds of link prediction algorithms, which fully shows that LPMF can effectively find the structural correlation between nodes in the network, and performs a more excellent performance in the actual tasks of link prediction. [ABSTRACT FROM AUTHOR]

In order to understand the development of SISR based on deep learning and grasp the hotspots and directions of the current research, this paper combed the existing model of SISR based on deep learning. Firstly, the paper introduced the related deep learning algorithm, these models based on deep learning and their evaluation index. In addition, it compared the performance of existing models through experiments, which aimed to understand the advantages of SISR model based on deep learning. Finally, the paper summarized the key issues of SISR, and prospected the future development trends. [ABSTRACT FROM AUTHOR]

This paper proposed a detection algorithm for sweeping jamming with low complexity under the constraint of the probability of false alarm and missed detection, and theoretically analyzed the performance of the detection algorithm to provide the basis for jamming detection of the actual communication system. Then, the paper proposed an anti-jamming learning algorithm based on Q-learning, which could autonomously select the best communication channel and the maximal reliable transmission time when threatened by the sweeping jamming. Finally, the simulation results show that the proposed detection algorithm can effectively detect the linear sweeping jamming, and the detection performance is basically consistent with the theoretical analysis results with low complexity. The anti-jamming learning algorithm can adaptively determine the optimal communication channel, and effectively avoid the sweeping jamming to realize the reliable information transmission. [ABSTRACT FROM AUTHOR]

In speech recognition, the convolution location-based hybrid attention mechanism can not extract location information that can be valid over long term. This paper proposed a new hybrid attention mechanism to solve this problem. Firstly it convolved with the attention score generated for the current time to extract multi-channel features, followed by obtaining the feature vectors of constant dimensions via global average pooling. Then it introduced a LSTM unit as the external memory module and used the generated feature vectors as input to generate the location vectors for the next time point. Finally this paper used the classic LAS model to verify the effect of the new hybrid attention mechanism. Experiment results show that the new hybrid attention mechanism can take full consideration of the attention scores at many past time points. Compared to the convolution location-based LAS model, the label error rate of the proposed scheme on pure and noisy speech datasets is reduced by 1.8% and 2.21%, respectively. [ABSTRACT FROM AUTHOR]

Mixed sampling data，whose different features are collected at different sampling frequencies， pervasively exists in the real world. Because the different sampling data set not only has different features，but also has different number of samples，traditional classification methods cannot be used directly. Therefore， this paper proposes a classification method for mixed sampling data based on Fisher discrimination dictionary learning to solve the classification problem of mixed sampling data. Inspired on some classification ideas for multi-view data，the proposed model compares multiple data collected at multiple sampling frequencies to multiple views of multi-view data，and designs a way to learn a sub-dictionary for each class of each view. A structured dictionary whose dictionary atoms have correspondence to the class labels is learned，so that the within-class scatter is less and the between-class scatter is bigger based on the Fisher discrimination criterion. In addition，this paper designs a specific classification scheme for the inconsistent sample size of mixed sampling data. Finally，experimental results demonstrate the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Qualitative safety evaluation and analysis can not meet individual safety requirements gradually. In order to better quantify the degree of risk in life and find the quantitative relationship between outliers and individual risk, this paper uses logical regression method to analyze the specific dependencies between independent variables and dependent variables. Taking a cement plant in Henan Province as the investigation object, an emergency early warning model based on logistic regression method is established. The linear relationship between independent variables and accident consequences is calculated and analyzed by SAS software, and compared with cluster analysis method. The results show that compared with clustering analysis, logistic regression model has better accuracy in monitoring outliers, which provides a good analysis basis for individual risk warning and accident prevention. [ABSTRACT FROM AUTHOR]

Dump yard landslide accidents occur frequently, with serious casualties and severe social impact. The scientific landslide early warning method of dump yard is of great significance for accident prevention and risk management. This paper proposes a dump landslide based on the mRBF-ELM integrated model. Early warning method, input the early warning index data of the dumping site into the trained multi-radial basis function extreme learning machine integrated model, combine the output of the stability coefficients of all sub-models, and the output result of the integrated model is after the prediction results of all sub-models are integrated output. As a comparison, the prediction results obtained by applying the mRBF-ELM integrated model are compared with the application of artificial neural network (ANN), extreme learning machine (ELM), Gaussian radial basis function extreme learning machine (RBF-ELM¹), hyperbolic tangent radial basis. The prediction results of the function extreme learning machine (RBF-ELM²) and the inverse multiple quadratic radial basis function extreme learning machine (RBF-ELM³) are compared and analyzed from the three aspects of average relative error, root mean square error, and goodness of fit to verify the reliability of the model. The results show that the early warning model has high evaluation reliability and good value in popularization and application. [ABSTRACT FROM AUTHOR]

Loop closure detection is important for visual simultaneous localization and mapping (VSLAM) systems to reduce cumulative errors and perform re-localization. In order to shorten the loop closure detection online runtime while meeting the accuracy recall rate requirement, the paper proposed a fast loop closure detection algorithm using broad autoencoder (FLCD-BA). The algorithm improved broad learning architecture to learn feature representations autonomously from the input data in an unsupervised manner, and then accomplished loop closure detection tasks. Differing from the traditional deep learning methods, the network used a pseudo-inverse ridge regression algorithm to solve the weight matrix and achieved fast reconfiguration of the network by incremental learning without retraining the entire network. The proposed algorithm experimented on three public datasets without using GPU devices, and the training time of the network greatly reduced compared to bag-of-words models as well as deep learning methods. The experimental results show that the algorithm has high accuracy and recall in detecting loop closure, and the average running time per frame in the test is only 21 ms, which provides a new algorithm for loop closure detection in visual SLAM systems. [ABSTRACT FROM AUTHOR]

Because heterogeneous information network（HIN） contains rich network structure and semantic information, recommendation systems often use HIN for recommendation. However, the current researches of recommender systems is mainly based on indirect information provided by meta-paths for recommendation, but these researches don’t make full use of direct interactive information. To make full use of this information, this paper proposed a ternary interaction model（AMMRec） that incorporated attention mechanisms and heterogeneous information network meta-paths. This method firstly used the implicit feedback matrix to construct user similarity matrix and item similarity matrix and used the representation learning method of the HIN to obtain the corresponding feature vector embeddings in HIN. Then it used the the attention mechanism to modify the embeddings and designed attention neural network to fuse representation vectors of different meta-paths. Finally it concatenated user embeddings and meta-path embeddings and item embeddings, and generated recommendation results through fully connected neural network. The experimental results on real datasets show that AMMRec improves the recommendation accuracy by up to 9.5%. In addition, AMMRec has good interpretability for the recommendation results. [ABSTRACT FROM AUTHOR]