Text similarity calculation is a part of natural language processing and is used to calculate the similarity between two words, sentences, or texts in many application scenarios. Research on text similarity calculation plays an important role in the development of artificial intelligence. Text similarity calculation has conventionally been based on character string surfaces. With the introduction of word vectors, text similarity calculation can be modeled and calculated based on statistics and deep learning, in addition to combining it with pre-trained models. First, text similarity calculation methods can be divided into five categories: character string-based, word vector-based, pre-trained model-based, deep learning-based, and other methods. Each category is briefly introduced. Subsequently, according to the principles of the different text similarity calculation methods, common methods such as the edit distance, Hamming distance, bag of words model, Vector Space Model (VSM), Deep Structured Semantic Model (DSSM), and Simple Contrastive learning of Sentence Embedding (SimCSE) are discussed. Finally, commonly used data sets and evaluation criteria for text similarity calculation are sorted and analyzed, and the future development of text similarity calculation is prospected.

The superior performance of Transformer in natural language processing has inspired researchers to explore their applications in computer vision tasks. The Transformer-based object detection model, Detection Transformer (DETR), treats object detection as a set prediction problem, introducing the Transformer model to address this task and eliminating the proposal generation and post-processing steps that are typical of traditional methods. The original DETR model encounters issues related to slow training convergence and inefficiency in detecting small objects. To address these challenges, researchers have implemented various improvements to enhance DETR performance. This study conducts an in-depth investigation of both the basic and enhanced modules of DETR, including modifications to the backbone architecture, query design strategies, and improvements to the attention mechanism. Furthermore, it provides a comparative analysis of various detectors and evaluates their performance and network architecture. The potential and application prospects of DETR in computer vision tasks are discussed herein, along with its current limitations and challenges. Finally, this study analyzes and summarizes related models, assesses the advantages and limitations of attention models in the context of object detection, and outlines future research directions in this field.

To satisfy the performance requirements for robot path planning, an algorithm integrating improved A^* algorithm and improved Dynamic Window Approach(DWA) is proposed, which shortens the path length and improves the searching efficiency and path smoothness. To combat the challenges of the traditional A^* algorithm in complex scenarios, a new heuristic function is designed based on Manhattan distance and the diagonal distance. The weights are assigned dynamically, and the global shortest path and the least searching time are obtained. Next, an improved search strategy based on the 8-neighborhood is proposed, which involves dynamically assigning the optimal search direction to the current node, thus improving the searching efficiency and reducing the time consumption compared to the traditional 8-neighborhood 8-direction search method. Subsequently, the Floyd algorithm is employed to remove redundant nodes, reduce the steering times, and shorten the path distance. Additionally, the traditional DWA faces certain challenges; for instance, the path is not globally optimal, the path planning may fail, or the path length may increase. To solve these problems, a keypoint densification strategy is proposed to modify the deflective path. Finally, the proposed improved A^* algorithm and fusion algorithm are compared with existing methods. The simulation results show that the improved A^* algorithm can generate the shortest global path in complex environments, reducing the average steering time by 16.3% and shortening the average path searching time by 55.66%. For the fused algorithm, the average path length and average runtime shorten by 6.1% and 14.7% in the temporary obstacle environment, respectively, and shorten by 1.6% and 39.8%, respectively, in the moving obstacle environment.

Text classification is a basic task in the field of natural language processing and plays an important role in information retrieval, machine translation, sentiment analysis, and other applications. However, most deep learning models do not fully consider the rich information in training instances during inference, resulting in inadequate text feature learning. To leverage training instance information fully, this paper proposes a text classification method based on contrastive learning and attention mechanism. First, a supervised contrastive learning training strategy is designed to optimize the retrieval of text vector representations, thereby improving the quality of the retrieved training instances during the inference process. Second, an attention mechanism is constructed to learn the attention distribution of the obtained training text features, focusing on adjacent instance information with stronger relevance and capturing more implicit similarity features. Finally, the attention mechanism is combined with the model network, fusing information from adjacent training instances to enhance the ability of the model to extract diverse features and achieve global and local feature extraction. The experimental results demonstrate that this method achieves significant improvements on various models, including Convolutional Neural Network(CNN), Bidirectional Long Short-Term Memory(BiLSTM), Graph Convolutional Network(GCN), Bidirectional Encoder Representations from Transformers(BERT), and RoBERTa. For the CNN model, the macro F1 value is increased by 4.15, 6.2, and 1.92 percentage points for the THUCNews, Toutiao, and Sogou datasets, respectively. Therefore, this method provides an effective solution for text classification tasks.

The deep integration of knowledge graphs with education has promoted the development of smart education. However, there is a lack of literature on educational knowledge graphs currently, necessitating its improvement with regard to research normativity and content perspective. Four conclusions are presented from a systematic literature review of 55 important Chinese journal articles from the previous decade. First, the development of educational knowledge graphs requires five key technologies: ontology construction, knowledge extraction, knowledge representation, knowledge fusion, and knowledge reasoning. Deep learning methods are becoming a popular research topic in this context. Second, in the context of applicability, the educational knowledge graphs cover six application scenarios: personalized learning recommendations, intelligent Question-Answering (Q&A), teaching resource management, intelligent search, intelligent learning diagnosis, and classroom teaching analysis, and the horizon of applications is continuously expanding. Third, regarding application effects, the educational knowledge graphs promote personalized learning and fragmented ubiquitous learning of students while improving their learning performance as well as professionalism of teachers. Fourth, the education knowledge graphs suffer from several problems and challenges, such as single data modality, lack of quality datasets, low level of automation and borderline technology, high level of difficulty in knowledge modeling, insufficient competence care, lack of interoperability standards, and low rate of educational adoption. Hence, for further insight into the study, future research should refine the theory and establish standards, optimize techniques, achieve accurate modeling, and strengthen applications and lifting effects.

An instance segmentation algorithm (DE-YOLO) based on the improved YOLOv8 is proposed. To decrease the effect of complex backgrounds in the images, efficient multiscale attention is introduced, and cross-dimensional interaction ensures an even spatial feature distribution within each feature group. In the backbone network, a deformable convolution using DCNv2 is combined with a C2f convolutional layer to overcome the limitations of traditional convolutions and increase flexibility. This is performed to reduce harmful gradient effects and improve the overall accuracy of the detector. The dynamic nonmonotonic Wise-Intersection-over-Union (WIoU) focusing mechanism is employed instead of the traditional Complete Intersection-over-Union (CIoU) loss function to evaluate the quality, optimize detection frame positioning, and improve segmentation accuracy. Meanwhile, Mixup data enhancement processing is enabled to enrich the training features of the dataset and improve the learning ability of the model. The experimental results demonstrate that DE-YOLO improves the mean Average Precision of mask(mAP_mask) and mAP_mask@0.5 by 2.0 and 3.2 percentage points compared with the benchmark model YOLOv8n-seg in the Cityscapes dataset of urban landscapes, respectively. Furthermore, DE-YOLO maintains an excellent detection speed and small parameter quantity while exhibiting improved accuracy, with the model requiring 2.2-31.3 percentage points fewer parameters than similar models.

The objective of image matching is to establish correspondences between similar structures across two or more images. This task is fundamental to computer vision, with applications in robotics, remote sensing, and autonomous driving. With the advancements in deep learning in recent years, Two-Dimensional (2D) image matching algorithms based on deep learning have seen regular improvements in feature extraction, description, and matching. The performance of these algorithms in terms of matching accuracy and robustness has surpassed that of traditional algorithms, leading to significant advancements. First, this study summarizes 2D image matching algorithms based on deep learning features from the past ten years and categorizes them into three types: two-stage image matching based on local features, image matching of joint detection and description, and image matching without feature detection. Second, the study details the development processes, classification methods, and performance evaluation metrics of these three categories and summarizes their advantages and limitations. Typical application scenarios of 2D image matching algorithms are then introduced, and the effects of research progress in 2D image matching on its application domains are analyzed. Finally, the study summarizes the development trends of 2D image matching algorithms and discusses future prospects.

The TAPDataset pedestrian detection dataset is used in this study to address the issues of low detection accuracy, large number of algorithm parameters, and limitations of existing public datasets for small target detection in current scenic pedestrian detection. This dataset addresses the deficiencies of existing datasets regarding small target detection. Based on the YOLOv8 algorithm, a new model with high detection accuracy and low hardware requirements, called YOLOv8-L, is proposed. First, the lightweight convolution module DepthSepConv is introduced to reduce the number of parameters and computations of the model. Second, the BiFormer attention mechanism and CARAFE upsampling operator are used to enhance the model's semantic understanding of images and information fusion capability, significantly improving detection accuracy. Finally, a small target detection layer is added to extract more shallow features, effectively improving the model's performance for small target detection. The effectiveness of the algorithm is verified using the TAPDataset, VOC 2007, and TAP+VOC datasets. The experimental results show that compared with YOLOv8, the number of model parameters is reduced by 18.06% on the TAPDataset with unchanged FPS, mAP@0.5 improves by 5.51%, and mAP@0.5∶0.95 improves by 6.03%. On the VOC 2007 dataset, the number of parameters is reduced by 13.6%, with mAP@0.5 improving by 3.96% and mAP@0.5∶0.95 improving by 6.39%. On the TAP+VOC dataset, the number of parameters is reduced by 14.02%, with mAP@0.5 improving by 4.49% and mAP@0.5∶0.95 improving by 5.68%. The improved algorithm demonstrates stronger generalization performance and can be better applied to scenic pedestrian detection tasks.

Fingerprint recognition is one of the earliest and most mature biometric recognition technologies that is widely used in mobile payments, access control and attendance in the civilian field, and in criminal investigation to retrieve clues from suspects. Recently, deep learning technology has achieved excellent application results in the field of biometric recognition, and provided fingerprint researchers with new methods for automatic processing and the application of fusion features to effectively represent fingerprints, which have excellent application results at all stages of the fingerprint recognition process. This paper outlines the development history and application background of fingerprint recognition, expounds the main processing processes of the three stages of fingerprint recognition, which are image preprocessing, feature extraction, and fingerprint matching, summarizes the application status of deep learning technology in specific links at different stages, and compares the advantages and disadvantages of different deep neural networks in specific links, such as image segmentation, image enhancement, direction field estimation, minutiae extraction, and fingerprint matching. Finally, some of the current problems and challenges in the field of fingerprint recognition are analyzed, and future development directions, such as building public fingerprint datasets, multi-scale fingerprint feature extraction, and training end-to-end fingerprint recognition models, are prospected.

High-quality annotated data are crucial for Natural Language Processing(NLP) tasks in the field of Chinese scientific literature. A method of annotation based on a Large Language Model(LLM) was proposed to address the lack of high-quality annotated corpora and the issues of inconsistent and inefficient manual annotation in Chinese scientific literature. First, a fine-grained annotation specification suitable for multi-domain Chinese scientific literature was established to clarify entity types and annotation granularity. Second, a structured text annotation prompt template and a generation parser were designed. The annotation task of Chinese scientific literature was set up as a single-stage, single-round question-and-answer process in which the annotation specifications and text to be annotated were filled into the corresponding slots of the prompt template to construct the task prompt. This prompt was then injected into the LLM to generate output text containing annotation information. Finally, the structured annotation data were obtained by the parser. Subsequently, using prompt learning based on LLM, the Annotated Chinese Scientific Literature(ACSL) entity dataset was generated, which contains 10 000 annotated documents and 72 536 annotated entities distributed across 48 disciplines. For ACSL, three baseline models based on RoBERTa-wwm-ext, a configuration of the Robustly optimized Bidirectional Encoder Representations from Transformers(RoBERT) approach, were proposed. The experimental results demonstrate that the BERT+Span model performs best on long-span entity recognition in Chinese scientific literature, achieving an F1 value of 0.335. These results serve as benchmarks for future research.

Addressing the limitations of Chinese Named Entity Recognition(NER) within specific domains, this paper proposes a model to enhance entity recognition accuracy by utilizing domain-specific Knowledge Graphs(KGs) and images. The proposed model leverages domain graphs and images to improve entity recognition accuracy in short texts related to computer science. The model employs a Bidirectional Encoder Representations from Transformers(BERT)-Bidirectional Long Short-Term Memory(BiLSTM)-Attention-based model to extract textual features, a ResNet152-based approach to extract image features, and a word segmentation tool to obtain noun entities from sentences. These noun entities are then embedded with KG nodes using BERT. The model uses cosine similarity to determine the most similar nodes in the KG for the segmented words in the sentence. It retains neighboring nodes with a distance of 1 from this node to generate an optimal matching subgraph for semantic enrichment of the sentence. A Multi-Layer Perceptron(MLP) is employed to map the textual, image, and subgraph features into the same space. A unique gating mechanism is utilized to achieve fine-grained cross-modal feature fusion between textual and image features. Finally, multimodal features are fused with subgraph features by using a cross-attention mechanism and are then fed into the decoder for entity labeling. Experimental comparisons with relevant baseline models conducted on Twitter2015, Twitter2017, and a self-constructed computer science dataset are presented. The results indicate that the proposed approach achieved precision, recall, and F1 value of 88.56%, 87.47%, and 88.01% on the domain dataset compared to the optimal baseline model, its F1 value increased by 1.36 percentage points, demonstrating the effectiveness of incorporating domain KGs for entity recognition.

In the context of big data, the rapid advancement of fields such as scientific computing and artificial intelligence, there is an increasing demand for high computational power across various domains. The unique hardware architecture of the Graphics Processing Unit (GPU) makes it suitable for parallel computing. In recent years, the concurrent development of GPUs and fields such as artificial intelligence and scientific computing has enhanced GPU capabilities, leading to the emergence of mature General-Purpose Graphics Processing Units (GPGPUs). Currently, GPGPUs are one of the most important co-processors for Central Processing Units (CPUs). However, the fixed hardware configuration of the GPU after delivery and its limited memory capacity can significantly hinder its performance, particularly when dealing with large datasets. To address this issue, Compute Unified Device Architecture (CUDA) 6.0 introduces unified memory, allowing GPGPU and CPU to share a virtual memory space, thereby simplifying heterogeneous programming and expanding the GPGPU-accessible memory space. Unified memory offers a solution for processing large datasets on GPGPUs and alleviates the constraints of limited GPGPU memory capacity. However, the use of unified memory introduces performance issues. Effective data management within unified memory is the key to enhancing performance. This article provides an overview of the development and application of CUDA unified memory. It covers topics such as the features and evolution of unified memory, its advantages and limitations, its applications in artificial intelligence and big data processing systems, and its prospects. This article provides a valuable reference for future work on applying and optimizing CUDA unified memory.

This study proposes an improved Artificial Potential Field (APF) algorithm (called FC-V-APF) based on Fuzzy Control (FC) and a virtual target point method to solve the local minimum trap and path redundancy issues of the APF method in robot local path planning. First, a virtual target point obstacle avoidance strategy is designed, and the V-APF algorithm is constructed to help the robot overcome local minimum traps by adding an obstacle crossing mechanism and a target point update threshold. Second, a control strategy based on the cumulative angle sum is proposed to assist the robot in exiting a multi-U complex obstacle area. Subsequently, the V-APF and FC algorithms are combined to construct the FC-V-APF algorithm. The corresponding environment is evaluated using real-time data from the radar sensor and designed weight function, and a fuzzy controller is selected to output the auxiliary force to avoid obstacles in advance. Finally, a simulation environment is built on the Robot Operating System (ROS) platform to compare the path planning performance of the FC-V-APF algorithm with that of other algorithms. Considering path length, running time, and speed curves, the designed FC-V-APF algorithm can quickly eliminate traps, reduce redundant paths, improve path smoothness, and reduce planning time.

Achieving enhanced detection accuracy is a challenging task in the field of PCB defect detection. To address this problem, this study proposes a series of improvement methods based on PCB defect detection. First, a novel attention mechanism, referred to as BiFormer, is introduced. This mechanism uses dual-layer routing to achieve dynamic sparse attention, thereby reducing the amount of computation required. Second, an innovative upsampling operator called CARAFE is employed. This operator combines semantic and content information for upsampling, thereby making the upsampling process more comprehensive and efficient. Finally, a new loss function based on the MPDIoU metric, referred to as the LMPDIoU loss function, is adopted. This loss function effectively addresses unbalanced categories, small targets, and denseness problems, thereby further improving image detection performance. The experimental results reveal that the model achieves a significant improvement in mean Average Precision (mAP) with a score of 93.91%, 13.12 percentage points higher than that of the original model. In terms of recognition accuracy, the new model reached a score of 90.55%, representing an improvement of 8.74 percentage points. These results show that the introduction of the BiFormer attention mechanism, CARAFE upsampling operator, and LMPDIoU loss function effectively improves the accuracy and efficiency of PCB defect detection. Thus, the proposed methods provide valuable references for research in industrial inspection, laying the foundation for future research and applications.

The System of Chip (SoC) integrates multiple peripheral interfaces, the verification of which has become one of the most time-consuming steps in chip development. The PCIe protocol provides high-speed peer-to-peer serial interconnection services within the system, while supporting hot swapping, which has gradually become a universal bus protocol. When using conventional Hardware Description Languages (HDL) to validate PCIe interface designs, problems usually arise, such as difficulty in covering multiple design scenarios and boundary conditions in a short period, leading to insufficient verification. To address the above issues, this study utilizes Universal Verification Methodology (UVM) to build a PCIe interface validation platform. This platform adopts a UVM-defined framework and test classes, achieving top-level environmental integration and design of test constraints, with strong reusability and comprehensive verification. This implementation includes SoC system-level environmental integration, design, and connection of the modules to be tested, implementation of sequencer and monitor classes in the verification platform, and partial interface design. To ensure that the test cases cover as many design states and paths as possible, different functional points are divided deliberately, and constraint conditions are designed to evaluate the effectiveness and coverage of test cases using various coverage indicators. The experimental results show that the verification platform can curtail the verification cycle and increase the comprehensive coverage by more than 30%.

Multimodal sentiment analysis aims to extract and integrate semantic information from text, images, and audio data in order to identify the emotional states of speakers in online videos. Although, multimodal fusion methods have shown definite outcomes in this research area, previous studies have not adequately addressed the distribution differences between modes and the fusion of relational knowledge. Therefore, a multimodal sentiment analysis method is recommended. In this context, this study proposes the design of a Multimodal Prompt Gate (MPG) module. The proposed module can convert nonverbal information into prompts that fuse the context, filter the noise of nonverbal signals using text information, and obtain prompts containing rich semantic information to enhance information integration between the modalities. In addition, a contrastive learning framework from instance to label is proposed. This framework is used to distinguish the different labels in latent space at the semantic level to further optimize the model output. Experiments on three large-scale sentiment analysis datasets are conducted. The results show that the binary classification accuracy of the proposed method improves by approximately 0.7% compared to the suboptimal model, and the ternary classification accuracy improves by more than 2.5%, reaching 0.671. This method can provide a reference for introducing multimodal sentiment analysis in the fields of user profiling, video understanding, and AI interviews.

In recent years, remarkable advancements in Artificial Intelligence (AI) across unimodal domains, such as computer vision and Natural Language Processing (NLP), have highlighted the growing importance and necessity of multimodal learning. Among the emerging techniques, the Zero-Shot Transfer (ZST) method, based on visual-language pre-trained models, has garnered widespread attention from researchers worldwide. Owing to the robust generalization capabilities of pre-trained models, leveraging visual-language pre-trained models not only enhances the accuracy of zero-shot recognition tasks but also addresses certain zero-shot downstream tasks that are beyond the scope of conventional approaches. This review provides an overview of ZST methods based on vision-language pre-trained models. First, it introduces conventional approaches to Few-Shot Learning (FSL) and summarizes its main forms. It then discusses the distinctions between ZST and FSL based on vision-language pre-trained models, highlighting the new tasks that ZST can address. Subsequently, it explores the application of ZST methods in various downstream tasks, including sample recognition, object detection, semantic segmentation, and cross-modal generation. Finally, it analyzes the challenges of current ZST methods based on vision-language pre-trained models and outlines potential future research directions.

Cybersecurity threats are becoming increasingly prevalent with the rapid advancement of Internet technologies. Cyberattacks exhibiting high complexity and diversity, are posing significant challenges to existing defense mechanisms. As an emerging concept, situation awareness technology offers new approaches to enhancing cybersecurity defense. However, the current cybersecurity situation awareness methods suffer from limited data feature extraction capabilities and inadequate handling of long-term sequential data. To address these issues, this study proposes a fusion model that integrates Stack Sparse Auto-Encoder (SSAE), Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Attention Mechanism (AM). By utilizing SSAE and CNN to extract data features and enhancing the focus on critical information through the AM in the BiGRU model, the proposed model aims to classify the attack categories of abnormal traffic. In conjunction with the network security situational quantification indicators proposed in this study, the network security situation is quantitatively evaluated and classified. The experimental results demonstrate that the proposed fusion model outperforms traditional deep learning models in various metrics, enabling an accurate perception of the network situation.

With the increasing applicability of Internet-of-Things (IoT) technology, the number and types of IoT devices and sensors are continuously increasing. In particular, IoT water quality sensors play a vital role in the field of ecological monitoring and protection. Accordingly, this study proposes an unsupervised anomaly data detection algorithm based on explainable deep learning to address the issues of large volume, high dimensionality, and lack of labeling in the monitoring data collected by IoT water quality sensors. The algorithm uses the Auto-Encoder (AE) and SHAP algorithms to detect anomalies in multi-dimensional water quality datasets. The AE model is trained to flag data with significant reconstruction errors, and SHAP is used to interpret the AE and calculate the importance of each feature in the flagged data. Based on the importance of these features, the final anomaly value is determined for the dataset. Experimental results on an IoT water quality monitoring dataset show that the algorithm can effectively detect anomalous data with an F1 value of 0.875, outperforming existing unsupervised anomaly detection algorithms. Thus, the proposed algorithm has a practical application value for processing IoT water quality monitoring data. Furthermore, the algorithm can be applied to the anomaly detection of massive IoT monitoring data in other fields, such as meteorology and the environment.

Due to low detection accuracy for small targets in complex environments, along with false and missed detections in mainstream traffic sign detection algorithms, an improved algorithm based on YOLOv8s is proposed. This algorithm uses Pconv convolution in the backbone network and incorporates a C2faster module to achieve a lightweight network structure while maintaining network accuracy. In addition, to better utilize the information between low- and high-level features and enhance the regional context association ability, the SPPFCSPC module is designed as a spatial pyramid pooling module based on the concept of SPPF. In addition, by adding the GAM attention mechanism, the feature extraction capability of the network is further enhanced, and the detection accuracy is effectively improved. To improve the detection ability of small targets, a four-fold downsampling branch is added at the neck of the network to optimize target positioning. In addition, the Focal-EIoU loss function is used to replace the original CIoU loss function to accurately define the aspect ratio of the prediction box, which alleviates the problem of imbalance between the positive and negative samples. Experimental results show that on the CCTSDB-2021 traffic sign dataset, the improved algorithm achieved 86.1%, 73.0%, and 81.2% precision, recall, and mAP@0.5, respectively. Compared with the original YOLOv8s algorithm, increases of 0.8%, 6.3%, and 6.9% were observed, respectively. This algorithm significantly reduces false and missed detections in complex weather and harsh environments, offering better overall detection performance than the comparison algorithm, with strong practical value.

The spatiotemporal fusion super-resolution reconstruction of remote sensing images extracts information from low-resolution images with high temporal density and high-resolution images with low temporal resolution to generate remote sensing images with both high temporal and spatial resolutions. This process is directly related to the implementation of subsequent tasks such as interpretation, detection, and tracking. With the rapid advancement of Convolutional Neural Network (CNN), researchers have proposed a series of CNN-based spatiotemporal fusion methods. However, because of the inherent limitations of convolution operations, these methods still face challenges with respect to global information extraction. Inspired by the global modeling capabilities of the Swin Transformer, this paper proposes a super-resolution reconstruction model based on the Swin Transformer. In the feature extraction stage, a dual-stream structure is introduced, dividing the feature extraction network into two parts to extract temporal and spatial information separately. The performance of the model is enhanced by the global capabilities of the Swin Transformer. In the feature fusion stage, a Convolutional Block Attention Module (CBAM) that combines channel and spatial attention is introduced to enhance the important features and improve the image reconstruction accuracy. Comparative experiments are conducted on the Coleambally Irrigation Area (CIA) and Lower Gwydir Catchment (LGC) datasets using various spatiotemporal fused super-resolution reconstruction models. The results show that the proposed model achieved optimal performance across all evaluation metrics, demonstrating superior performance and enhanced generalization capabilities.

The exponential growth of diverse terminal devices, driven by advancements in mobile computing, 5th Generation mobile communication (5G), and Internet of Things (IoT) technologies, has generated vast volumes of semi-structured and unstructured data, as these devices connect to networks. Neural networks are widely applicable because of their unique advantages in mining data. To enhance data processing capabilities and inference accuracy, neural network models are often designed to be complex and thus consume substantial computational resources in both storage and execution. However, edge devices typically have limited computational resources and cannot satisfy the storage and inference requirements of complex neural network models. Therefore, they often rely on cloud computing centers to perform these tasks. This cloud-based collaboration can lead to increased response latency and network bandwidth consumption, incurring potential risks such as the infringement of user privacy through data leakage. To address these issues, this paper introduces Neural network Generation and Deployment (NGD), a method for rapidly generating and deploying tailored neural network models on edge devices, whereby neural network models that match the hardware configuration of edge devices and specific computational task requirements are generated and rapidly deployed onto target devices for local inference. Experiments are conducted on three typical edge devices, and the results demonstrate the effectiveness of NGD in generating and deploying models, affirming its practicality and effectiveness.

Existing trajectory prediction methods rely heavily on high-definition maps, which are time-consuming, costly, and complex to acquire. This makes it difficult for them to quickly adapt to the widespread adoption of intelligent transportation. To address the problem of vehicle trajectory prediction in map-free scenes, a trajectory prediction method based on spatio-temporal features of multi-modal data is proposed in this paper. Multiple spatio-temporal interaction graphs are constructed from the history of the trajectory, temporal and spatial attention are cross-utilized and deeply fused to model the spatio-temporal correlations between vehicles on the road. Finally, a residual network is used for a multi-objective and multi-modal trajectory generation. The model is trained and tested on the real dataset, Argoverse 2, and the experimental results show that compared with the CRAT-Pred, this model can improve minADE, minFDE and Miss Rate(MR) metrics in single-modal prediction by 3.86%, 3.89%, and 0.48%, and in multi-modal prediction by 0.78%, 0.96% and 0.42%. Hence, the proposed trajectory prediction method can efficiently capture the temporal and spatial characteristics of vehicle movement trajectories and can be effectively applied in related fields such as autonomous driving.

Smartphones have become an integral part of modern daily life. The Android operating system currently holds the largest market share in the mobile operating system market owing to its open-source nature and comprehensive ecosystem. Within Android smartphones, the storage subsystem plays a pivotal role, exerting a significant influence on the user experience. However, the design of Android mobile storage systems diverges from server scenarios, necessitating the consideration of distinct factors, such as resource constraints, cost sensitivity, and foreground application prioritization. Extensive research has been conducted in this area. By summarizing and analyzing the current research status in this field, we categorize the issues experienced by users of Android smartphone storage systems into five categories: host-side writing amplification, memory swapping, file system fragmentation, flash device performance, and I/O priority inversion. Subsequently, existing works addressing these five categories of issues are classified, along with commonly used tools for testing and analyzing mobile storage systems. Finally, we conclude by examining existing techniques that ensure the user experience with Android smartphone storage systems and discuss potential avenues for future investigation.

Spike sorting refers to the detection, clustering, and identification of distinct extracellular neuronal signals from different recording sites with the aim of reliably assigning them to different putative neurons. This crucial preprocessing is fundamental for neural decoding in neuroscience and represents a prominent research direction in high-bandwidth Brain-Computer Interface (BCI) studies. Conventional spike sorting algorithms involve various steps such as spike detection, spike alignment, feature extraction, and feature clustering. Currently, explosive growth in the number and density of neural electrodes presents significant challenges in terms of the efficiency and accuracy of spike sorting. To address issues such as limited feature extraction capabilities, low Signal-to-Noise Ratio (SNR), and signal superposition, algorithmic advancements, such as artificial intelligence approaches and big data spike sorting solutions have emerged as strategies for enhancing the comprehension of the intricate principles and functions underlying brain activity. This paper provides an overview linking invasive BCIs with neural encoding/decoding and spike sorting methods. Subsequently, it outlines the principles underlying various spike sorting algorithms, while discussing how brain neural signals map onto specific behaviors. Finally, the paper concludes by anticipating future challenges and trends in the development of high-bandwidth neural encoding and decoding.

With the continuous development of big data and artificial intelligence technologies, knowledge graph embedding is developing rapidly, and knowledge graph applications are becoming increasingly widespread. Knowledge graph embedding improves the efficiency of knowledge representation and reasoning by representing structured knowledge into a low-dimensional vector space. This study provides a comprehensive overview of knowledge graph embedding technology, including its basic concepts, model categories, evaluation indices, and application prospects. First, the basic concepts and background of knowledge graph embedding are introduced, classifying the technology into four main categories: embedding models based on translation mechanisms, semantic- matching mechanisms, neural networks, and additional information. The core ideas, scoring functions, advantages and disadvantages, and application scenarios of the related models are meticulously sorted. Second, common datasets and evaluation indices of knowledge graph embedding are summarized, along with application prospects, such as link prediction and triple classification. The experimental results are analyzed, and downstream tasks, such as question-and-answer systems and recommenders, are introduced. Finally, the knowledge graph embedding technology is reviewed and summarized, outlining its limitations and the primary existing problems while discussing the opportunities and challenges for future knowledge graph embedding along with potential research directions.

Large Language Models(LLMs) are currently undergoing vigorous development in the field of Natural Language Processing(NLP). However, significant challenges remain in their applications to educational digitization. To address the problem posed by the scarcity of domain-specific data and the instability of summarization leading to information loss or redundancy, this study introduces a lightweight idempotent model framework, Idempotent Generative Language Model(IGLM), for educational text summarization. The model first employs multisource training for adaptive augmentation to enhance data diversity. Subsequently, various fine-tuning procedures are applied to the downstream text summarization task. Concurrently, an idempotent summarization generation strategy is designed to mitigate the impact of text length. This strategy brings the summaries closer to idempotent form, constrains the model, mitigates biases resulting from uneven language corpora, and combines quantization techniques to generate more precise and fluent summaries under low-resource conditions. The experiments used Recall-Oriented Understudy for Gisting Evaluation(ROUGE) scores as the evaluation metric and validated the model on publicly available Chinese text summarization datasets Large-scale Chinese Short Text Summarization(LCSTS), EDUCATION, and Natural Language Processing and Chinese Computing(NLPCC). The results revealed significant enhancements in precision and coherence within this framework. Specifically, compared to the baseline model, the ROUGE-1/2/L scores were improved by 7.9, 7.4, and 8.7 percentage points on the LCSTS dataset. Moreover, on the EDUCATION dataset, the scores exhibited enhancements of 12.9, 15.4, and 15.7 percentage points for ROUGE-1/2/L. Similarly, on the NLPCC dataset, improvements of 12.2, 11.7, and 12.7 percentage points were observed for ROUGE-1/2/L. This validation confirms the model's effectiveness.

Depression, as a prevalent mental health problem, substantially impacts individual′s daily lives and well-being. Addressing the limitations of current depression detection, such as subjectivity and manual intervention, automatic detection methods based on deep learning have become a popular research direction. The primary challenge in the most accessible text modality is modelling the long-range and sequence dependencies in depressive texts. To address this problem, this paper proposes a contextual knowledge-based enhanced Transformer network model, named Robustly optimized Bidirectional Encoder Representations from Transformers approach-Bidirectional Long Short-Term Memory(RoBERTa-BiLSTM), to comprehensively extract and utilize contextual features from depressive text sequences. By combining the strengths of sequence models and Transformer architectures, the proposed model captures contextual interactions between words to provide a reference for depression category prediction and information characterization. First, the RoBERTa model is employed to embed vocabulary into a semantic vector space, and then, a BiLSTM network effectively captures long-range contextual semantics. Finally, empirical research is conducted on two large-scale datasets, DAIC-WOZ and EATD-Corpus. Experimental results demonstrate that the model achieves an accuracy exceeding 0.74 and 0.93, and a recall exceeding 0.66 and 0.56, respectively, enabling accurate depression detection.

To solve the problem of insufficient personalized monitoring in human health assessment methods and meet the demand for fine-grained health status assessment in different scenarios, a multi-scenario-based human health status assessment method is needed to achieve long-term automated monitoring. This study proposes a multi-scenario human health assessment model based on a combination of the Analytic Hierarchy Process (AHP) and Entropy Weight Method (EWM). First, health monitoring index data for the human body in four different scenarios, including exercise, rest, work/study, and recreation, are collected to construct the corresponding assessment index system. Then, the AHP and EWM weights are calculated for the assessment indicators, and the Quantum-behaved Particle Swarm Optimization (QPSO) algorithm is used to distribute the subjective and objective weights for the AHP and EWM to ensure the objectivity of the proportion of evaluation indicators. Finally, the human health state is assessed and quantified using the fuzzy comprehensive evaluation method, and the reliability and stability of the method are verified using actual monitoring data. The experimental results show that the composite scores of the proposed method under the four scenarios (exercise, rest, work/study, and recreation) are 63.78, 59.83, 58.71, and 59.21, respectively, indicating that the model has good accuracy and stability under different scenarios. The results of the physical state evaluation of the testers are analyzed, and some health suggestions are given. The model proposed in this study can comprehensively determine the health status of the human body under different scenarios and provide scientific health guidance. Thus, it provides a scientific basis for health management and disease prevention.

Image fusion is the process of combining multiple input images into a unified single image. Although visible-infrared image fusion enhances target detection accuracy, its performance often fails in low-light scenarios. This study introduces a novel fusion model, namely, DAPR-Net, which features an encoder-decoder structure with cross-layer residual connections. These connections link the encoder's output to the corresponding layer in the decoder to thereby reinforce the information flow between the convolutional layers. Within the encoder, a Dual Attention Feature Extraction Module(AFEM) is designed to better distinguish the differences between the fused image and the input visible light and infrared images while retaining crucial information from both. The experimental results show that compared with the benchmark PIAFusion model, the information entropy, spatial frequency, mean gradient, standard deviation, and visual fidelity indices on the model on LLVIP and MSRS datasets increase by 0.849, 3.252, 7.634, 10.38, and 0.293, and 2.105, 2.23, 4.099, 27.938, and 0.343, respectively. In the YOLOV5 target detection network, the average mean precision, recall rate, accuracy rate, and F1 value index of the LLVIP and MSRS datasets increased by 8.8, 1.4, 1.9, and 1.5 percentage points and 7.5, 1.4, 8.8, and 1.2 percentage points, respectively, showing significant advantages compared with other fusion methods.

Small targets often exhibit low resolution and blurriness and are easily affected by occlusions and background interference, making accurate and real-time detection of small targets challenging. In this study, to enhance the detection performance, a super-resolution-aided small-target detection algorithm based on multi-task learning called Multi-YOLO is proposed. First, a super-resolution auxiliary branch is introduced to guide the main network in extracting effective features, thereby reducing the loss of information for small targets. Second, a collaborative supervision method is employed by combining Anchor based and Anchor free detection heads to improve the detection accuracy. Additionally, a CTR3 module is used at the end of the backbone network to strengthen the correlation between the target information and position awareness. Finally, during the inference stage, only the detection branch is used to maintain the speed of inference. Experimental results show that, compared with the baseline network, Multi-YOLO achieves performance improvement on the VEDAI, COCO MiniTrain, and SPCD datasets. Specifically, on the VEDAI dataset, this method achieves a 10.9% improvement in mean Average Precision (mAP) improvement while maintaining a model size similar to that of the baseline model. Moreover, compared with mainstream single-stage object detection networks, Multi-YOLO excels in small-target detection, maintaining a remarkable balance between accuracy and speed.

In the case of detection tasks in low-light environments, owing to the influence of unfavorable factors, such as low brightness, low contrast, and noise, missed and wrong detections can occur. Hence, a low-light object detection algorithm based on image adaptive enhancement is proposed. Combining conventional image processing methods with deep learning, an image adaptive enhancement network is designed, where multiple adjustable filters are combined in cascade to gradually enhance the input low-light image, and the adjustment parameters of each filter are predicted using a convolutional neural network based on the global information of the input image. The adaptive enhancement network is combined with the YOLOv5 object detection network for end-to-end joint training such that the image enhancement effect is more conducive to object detection. As the low-light object detection process is susceptible to missed detection, the channel attention mechanism SE-Net is improved, and a feature enhancement network is designed and embedded into the end of the Neck region of the YOLOv5 network to reduce the loss of information about potential target features caused by the process of fusion of network features. Experimental results show that the proposed algorithm achieves a detection accuracy of 77.3% on the low-light dataset ExDark, which is 2.1 percentage points higher than that afforded by the original YOLOv5 object detection network, and its detection speed reaches 79 frame/s, which affords real-time detection.

Air pollution is one of the primary challenges in urban environmental governance, with PM_2.5 being a significant contributor that affects air quality. As the traditional time-series prediction models for PM_2.5 often lack seasonal factor analysis and sufficient prediction accuracy, a fusion model based on machine learning, Seasonal Autoregressive Integrated Moving Average (SARIMA)-Support Vector Machine (SVM), is proposed in this paper. The fusion model is a tandem fusion model, which splits the data into linear and nonlinear parts. Based on the Autoregressive Integral Moving Average (ARIMA) model, the SARIMA model adds seasonal factor extraction parameters, to effectively analyze and predict the future linear seasonal trend of PM_2.5 data. Combined with the SVM model, the sliding step size prediction method is used to determine the optimal prediction step size for the residual series, thereby optimizing the residual sequence of the predicted data. The optimal model parameters are further determined through grid search, leading to the long-term predictions of PM_2.5 data and improves overall prediction accuracy. The analysis of the PM_2.5 monitoring data in Wuhan for the past five years shows that prediction accuracy of the fusion model is significantly higher than that of the single model. In the same experimental environment, the accuracy of the fusion model is improved by 99%, 99%, and 98% compared with those of ARIMA, Auto ARIMA, and SARIMA models, respectively and the stability of the model is also better, thus providing a new direction for the prediction of PM_2.5.

The existing Dung Beetle Optimization(DBO) algorithm has the disadvantages of poor search accuracy and insufficient global search ability, thereby easily falling into local optima. This paper proposes a multi-strategy improved dung beetle optimization algorithm that uses a chaotic opposition-based learning strategy to initialize the dung beetle population, whereby dung beetle individuals are evenly distributed in solution space and population diversity is improved. The golden sine strategy with a nonlinear weight is introduced to improve the ball-rolling behavior and coordinate the global search and local mining ability of the algorithm. Foraging behavior is improved by referring to the position update strategy of the sparrow search algorithm, which brings the population close to the optimal position and improves convergence speed and algorithmic accuracy. Stealing behavior is improved by introducing a piecewise function, which benefits the population in the full global exploration in the early iteration stages, to avoid premature convergence of the algorithm. The Cauchy-Gaussian mutation strategy with a nonlinear weight is used to randomly perturb the current optimal position and guide the algorithm to jump out of the local optimal position. The proposed algorithm is compared with five optimization algorithms using 23 benchmark functions, 12 CEC2022 test functions, and two engineering optimization problems. The experimental results show that the proposed algorithm is superior to the other algorithms and ranks first among at least 21 benchmark functions, 10 CEC2022 test functions, and two engineering optimization problems. Compared with the original dung beetle optimization algorithm, the proposed algorithm exhibits significant improvements in convergence accuracy, convergence speed, global search ability, and stability.

Deep-learning models have achieved impressive results in fields such as image classification; however, they remain vulnerable to interference and threats from adversarial examples. Attackers can craft small perturbations using various attack algorithms to create adversarial examples that are visually indistinguishable yet can lead to misclassification in deep neural networks, posing significant security risks to image classification tasks. To improve the robustness of these models, we propose an adversarial-example defense method that combines adversarial detection and purification using a conditional diffusion model, while preserving the structure and parameters of the target model during detection and purification. This approach features two key modules: adversarial detection and adversarial purification. For adversarial detection, we employ an inconsistency enhancement technique, training an image restoration model that integrates both the high-dimensional features of the target model and basic image features. By comparing the inconsistencies between the initial input and the restored output, adversarial examples can be detected. An end-to-end adversarial purification method is then applied, introducing image artifacts during the denoising process. An adversarial detection and purification module is placed before the target model to ensure its accuracy. Based on detection outcomes, appropriate purification strategies are implemented to remove adversarial examples and improve model robustness. The method was compared with recent adversarial detection and purification approaches on the CIFAR10 and CIFAR100 datasets, using five adversarial attack algorithms to generate adversarial examples. It demonstrated a 5-9 percentage points improvement in detection accuracy over Argos on both datasets in a low-purification setting. Additionally, it exhibited a more stable defense performance than Adaptive Denoising Purification(ADP), with a 1.3 percentage points higher accuracy under Backwards Pass Differentiable Approximation(BPDA) attacks.

Knowledge tracing is a novel field that merges artificial intelligence and education with the aim of evaluating the knowledge state of students through an interactive sequence of exercises that they have completed in the past. This is a core technology for implementing large-scale personalized learning services. With the widespread application of deep learning in computer vision, Natural Language Processing (NLP), recommendation systems, and other fields, a significant number of neural network-based methods, known as Deep Knowledge Tracing(DKT) models, have emerged in the field of knowledge tracing. To address the shortcomings of existing DKT models in terms of interpretability and accuracy, this study proposes a knowledge tracing model called SLKT that integrates sequence features with learning processes and includes knowledge state, sequence feature, and prediction modules. The knowledge state module is used to simulate the learning process of students, and the sequence feature module captures their recent learning status. Through the integration of sequence features and the learning process, the SLKT model effectively solves the problems of knowledge state modeling methods that cannot consider learners' recent learning status. A dynamic Q matrix with constraints is also proposed to represent the relationship between exercises and knowledge points to better model the learning process of the learner. This ensures better interpretability, and effectively improves model accuracy. In this study, the superior performance of the proposed model is verified using the Area Under Curve (AUC) and accuracy metrics, comparing multiple sets of experiments using depth tracing models such as DKT, Dynamic Key-Value Memory Network (DKVMN), Self-Attentive Knowledge Tracing (SAKT), and Convolutional Knowledge Tracing (CKT) on public datasets in three knowledge tracing fields.

Steel surface defect detection technology in industrial scenarios has problems such as low detection accuracy and slow convergence speed, To address these issues, an improved YOLOv8 algorithm (YOLOv8-Steel surface defect detection method based on Multi-scale Cross-fertilization Network,Deformable Convolution and Coordinate Attention,YOLOv8-MDC). Firstly, the MCN feature pyramid is added into the backbone network to fully fuse the semantic information between different features; secondly, deformable convolution is introduced into the module to adaptively change the shape and position of the convolution kernel (Deformable Convolution-C2f, DCN-C2f); finally, the CA attention mechanism is added to embed the positional information into the channel attention, which solves the problem of location information loss. The experimental results on the NEU-DET dataset show that the YOLOv8n-MDC algorithm on the NEU-DET dataset has an average accuracy value (mAP@IoU=0.5) of 81.0%, which is 4.2% higher than that of the original baseline network, and the algorithm has a faster convergence speed and higher accuracy, which is better able to meet the requirements of practical industrial production.

Existing image captioning methods only use regional visual features to generate description statements and ignore the importance of grid visual features. Moreover, as these methods are two-stage approaches, image captioning quality is affected. To address this issue, this study proposes an end-to-end image captioning method based on the visual feature fusion of Transformer. First, in the feature extraction stage, the visual feature extractor is used to extract regional and grid visual features. Second, in the feature fusion stage, the regional and grid visual features are concatenated using a visual feature fusion module. Finally, the visual features are sent to the language generator to realize image captioning. All components of the method are implemented based on the Transformer model, which is a one-stage method. The experimental results on the MS-COCO dataset show that the proposed method can fully utilize the respective advantages of regional and grid visual features, with the BLEU-1, BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE metrics reaching 83.1%, 41.5%, 30.2%, 60.1%, 140.3%, and 23.9%, respectively, indicating that the proposed method is superior to mainstream image captioning methods and can generate more accurate and rich description statements.

Adversarial examples are crucial for evaluating the robustness of Deep Neural Network (DNN) and revealing their potential security risks. The adversarial example generation method based on a Generative Adversarial Network (GAN), AdvGAN, has made significant progress in generating image adversarial examples; however, the sparsity and amplitude of the perturbation generated by this method are insufficient, resulting in lower authenticity of adversarial examples. To address this issue, this study proposes an improved image adversarial example generation method based on AdvGAN, Squeeze-and-Excitation (SE)-AdvGAN. SE-AdvGAN improves the sparsity of perturbation by constructing an SE attention generator and an SE residual discriminator. The SE attention generator is used to extract the key features of an image and limit the position of perturbation generation. The SE residual discriminator guides the generator to avoid generating irrelevant perturbation. Moreover, a boundary loss based on l₂ norm is added to the loss function of the SE attention generator to limit the amplitude of perturbation, thereby improving the authenticity of adversarial examples. The experimental results indicate that in the white box attack scenario, the SE-AdvGAN method has higher sparsity and smaller amplitude of adversarial example perturbation compared to existing methods and achieves better attack performance on different target models. This indicates that the high-quality adversarial examples generated by SE-AdvGAN can more effectively evaluate the robustness of DNN.

Target detection technology is advancing, but recognizing online listening behavior remains a challenge. Inaccurate identification of online classroom conduct and high model computation owing to limited human supervision and complex target detection models pose problems. To address this, we employed an upgraded YOLOv8-based method to detect and identify online listening behaviors. This approach incorporates a Bidirectional Feature Pyramid Network (BiFPN) to fuse features based on YOLOv8n, thereby enhancing feature extraction and model recognition accuracy. Second, the C3Ghost module is selected over the C2f module on the Head side to minimize the computational burden significantly. The study demonstrates that the YOLOv8n-BiFPN-C3Ghost model achieved an mAP@0.5 score of 98.6% and an mAP@0.5∶0.95 score of 92.6% on an online listening behavior dataset. The proposed model enhanced the accuracy by 4.2% and 5.7%, respectively, compared with other classroom behavior recognition models. Moreover, the required computation amount is only 6.6 GFLOPS, which is 19.5% less than that of the original model. The YOLOv8n-BiFPN-C3Ghost model is capable of detecting and recognizing online listening behavior with greater speed and accuracy while utilizing lower computing costs. This will ultimately enable the dynamic and scientific recognition of online classroom learning among students.

Knowledge Graph Completion(KGC) is an important means of improving the quality of KGs. Existing methods for KGC are mainly divided into structure- and description-based methods. The structure-based methods perform poorly in the inference of long-tailed entities commonly found in KGs, and the description-based methods are insufficient for descriptive information utilization and negative sample information learning. To address these challenges, this paper proposes a KGC method with knowledge enhancement and contrastive learning, named KEKGC. A specific template is designed to convert triples and their descriptive information into coherent natural language statements. These statements serve as input to a Pre-trained Language Model(PLM) through a manually defined template, enhancing the language model's comprehension of the structural and descriptive knowledge of each triple. On this basis, a contrastive learning framework is incorporated to improve the efficiency and accuracy of the link prediction task. This is done by building a memory bank to store entity embedding vectors, from which positive and negative samples are selected and trained with InfoNCE loss. Experimental results on WN18RR dataset show that compared with MEM-KGC, in the link prediction task, KEKGC improves Mean Reciprocal Rank(MRR) by 5.5, improves Hits@1, Hits@3, and Hits@10 metrics by 2.8, 0.7, and 4.2 percentage points, respectively, and the accuracy of the triple classification task reaches 94.1%. Hence, this method can achieve higher prediction accuracy and better generalization ability, especially for long-tailed entities, which can effectively improve graph completion.

In recent years, the application of deep learning algorithms to modulation recognition tasks has become a research hotspot in the field of communication. However, existing methods suffer from high network complexity, high hardware requirements, and low recognition accuracy under low Signal-to-Noise Ratio (SNR). A modulation recognition method based on a two-way convolutional neural network cascade separable convolutional network called TCNN-MobileNet is proposed, which combines the Discrete Wavelet Transform (DWT) method. First, the data are preprocessed using Wavelet Transform (WT), and the signal is transmitted as input to a dual convolutional neural network for feature extraction in different dimensions. Subsequently, feature fusion is performed through the fusion layer and fed into the lightweight neural network MobileNetV1 for modulation recognition model training. Finally, 11 types of modulation recognition are classified and output through a fully connected layer. The experimental results on the publicly available dataset RML2016.10a show that the recognition accuracy of TCNN-MobileNet can reach 88.71% at a low SNR of -20 dB, 96.66% at a high SNR of 18 dB, and an average recognition accuracy of 88.37% in the SNR range of -20 dB to 18 dB, which is approximately 35% higher than classical network architectures such as ResNet18 and ResNet34. The TCNN-MobileNet recognition method can reduce the number of training parameters and the network training time while ensuring that the recognition accuracy remains unchanged, effectively simplifying the network architecture and reducing hardware requirements. This is significant for the application of lightweight neural networks in modulation recognition.

Pre-trained language models, which are trained on large-scale datasets with extensive computing power, can extract significant amounts of knowledge from unstructured text data. To address the limited information in current triplets, a method is proposed that utilizes pre-trained language models to enrich this knowledge. Initially, a textual knowledge generator is designed to enhance the semantics of the triplets by leveraging the extensive knowledge embedded in the pre-trained models. This generator transforms the information within the triplets into subgraph descriptions. Subsequently, a question type predictor is employed to determine the appropriate question words. These question words are essential for question generation as they help to locate the domain of the answer accurately, resulting in semantically coherent questions and enhanced control over the generation process. Finally, a controlled generation framework is developed to ensure that both key entities and question words appear in the generated questions, thereby increasing the accuracy of these questions. The efficacy of the proposed model is demonstrated on the public datasets WebQuestion and PathQuestion. When compared to the existing model LFKQG, the proposed model shows improvements in the BLUE-4, METEOR, and ROUGE-L metrics by 0.28, 0.16, and 0.22 percentage points, respectively, on the WebQuestion dataset, and by 0.8, 0.39, and 0.46 percentage points, respectively, on the PathQuestion dataset.

In path planning for mobile robots, challenges arise when dealing with unknown and dynamically changing environments, such as high collision rates with obstacles and susceptibility to local optima. To address these issues, this paper proposes an improved Twin Delayed Deep Deterministic (TD3) algorithm, based on TD3 policy gradient, to enhance the path-planning performance of mobile robots in unknown dynamic environments. First, a Long Short-Term Memory (LSTM) neural network is introduced and combined with the TD3 algorithm. Employing gate structures, historical state information is filtered to perceive the state changes of obstacles within the sensing range for the robot to gain a better understanding of the dynamic environment and movement patterns of obstacles. This enables the mobile robot to accurately predict and respond to the behavior of dynamic obstacles, thereby reducing the collision rate with obstacles. Second, Ornstein-Uhlenbeck(OU) exploration noise is incorporated to facilitate continuous exploration of the surrounding environment, thereby enhancing the robot's random exploration capability. Additionally, a single experience pool is divided into three separate pools-success, failure, and temporary-to improve the sampling efficiency of the effective samples and reduce training time. Finally, simulation experiments are conducted for two different scenarios involving a mixture of dynamic and static obstacles for path planning. A comparative analysis of the experimental results demonstrates that in scenario 1, the proposed algorithm reduces the convergence of the model by 100-200 rounds compared with the Deep Deterministic Policy Gradient (DDPG) and TD3 algorithms. Moreover, it shortens the path length by 0.5-0.8 units and reduces the planning time by 1-4 s. In scenario 2, the proposed algorithm reduces the convergence of the model by 100-300 rounds compared to the TD3 algorithm, shortening the path length by 1-3 units and reducing the planning time by 4-8 s. However, the DDPG algorithm fails as the mobile robot is unable to reach the destination successfully. Therefore, the improved algorithm exhibits superior path planning performance.

In speech enhancement, Auto-Encoder (AE) structures are typically used to extract features automatically. However, the features obtained in this manner are singular, redundant, and cannot adequately capture the contextual dependencies of speech signals. Therefore, a speech-enhancement method, MSF-CI, that incorporates multi-scale features and contextual information is proposed. First, a multi-scale convolutional block is used to extract multi-scale features of speech signals to solve the issue of single features. Second, the attention mechanism is applied to focus on the spatial and channel key information of the extracted features to eliminate feature redundancy. Finally, a Gated Convolutional Recurrent Neural(GCRN) network is used to learn the long-span context-dependent relations of the speech signal, whereas gated linear units are employed to improve the nonlinear learning ability and thus improve the generalization of the network. Experimental results show that the proposed MSF-CI method outperforms similar single-channel speech-enhancement models such as GRN, DPT-FSNet, and U-Net in terms of speech-perception quality and the short-term objective intelligibility of enhanced speech signals at low Signal-to-Noise Ratios(SNR) and in different noise environments. Under a SNR is 0 dB, the average speech-perception quality and average speech objective intelligibility of the proposed method are 1.49 and 0.761, respectively. The generalizability of the proposed method is verified on the Ando Tibetan corpus. Additionally, its average speech-perception quality and average speech objective intelligibility improved by 20.7% and 11.3%, respectively, with respect to noise. Therefore, the MSF-CI model not only enhances speech quality and intelligibility but also provides better generalization.

To meet the requirements of low-overhead and high-performance applications in the embedded field, a 32 bit microprocessor based on the RISC-V open-source instruction set architecture has been developed. The processor adopts a five-stage pipeline structure of sequential launch, sequential execution, and out-of-order write back and realizes a combination of integer and multiply-divide instruction set modules. To cope with pipeline conflicts, the processor adopts a dynamic branch-prediction technique and designs data-correlation control and write back disorder mechanisms. The processor is designed using Verilog and builds a System on Chip (SoC) using an Advanced High-performance Bus (AHB) and an Advanced Peripheral Bus (APB) as interconnecting bus protocols. The processor logic functions are verified by writing RV32IM assembly instruction test programs in a simulation environment. Add timing constraints and physical constraints under the Vivado synthesis tool to perform logical synthesis on the processor code and analyze the utilization of processor hardware resources, and the synthesized code stream file is downloaded to the Xilinx Artix-7 (XC7A200T-2FBG484I) Field Programmable Gate Array(FPGA) development board and runs at a main frequency of 50 MHz. In the CoreMark program at 50 MHz, the CoreMark running points reaches 3.25 CoreMark/MHz. The validation results show that the processor performance running points is the same as that of the ARM Cortex-M3 series processors, and under the premise of the same technical comparison indexes, the processor running points is better than that of the RISC-V processor comparison item. The designed processor logic function is correct and uses a low hardware overhead to achieve relatively high-performance indicators suitable for cost-constrained high-performance embedded applications.

With the development of agricultural information technology, a substantial amount of rice planting-related data has been accumulated on the Internet. To address the challenges that farmers face in quickly obtaining accurate information during the planting process, an intelligent question-answering system is constructed based on a knowledge graph, specifically for rice planting. First, relevant data are obtained through manual collection as well as web crawler technology. Natural language processing techniques, such as the named entity recognition model and an intent recognition model, are built in conjunction with front- and back-end technologies to develop an intelligent question-answering system for rice planting. Experimental results show that in the named entity recognition and intent recognition modules, the F1 values of the constructed models reach 89.17% and 96.54%, respectively, which are higher than those of other conventional models. The intelligent rice planting question-answering system, based on knowledge graph, can accurately answer most inquiries farmers encounter during the process of rice planting, facilitating the management and visualization of rice planting knowledge graph data.

Critical node detection has become an important research domain in complex networks; however, current critical node detection methods suffer from high algorithm time complexity, inaccurate critical node sets obtained, and insufficient consideration of node centrality indicators. Based on this, this study first presents a critical node-detection framework based on layer and node features. This framework introduces a layer-partitioning-based method to enhance the efficiency of selecting the initial coverage set of critical nodes, allowing for the calculation of the initial node set in linear time. Subsequently, the nodes were added back to the original network through node centrality feature indicators until the number of nodes in the solution set met a predefined threshold. To overcome the challenge of local optima during node re-addition, a node centrality index was developed, taking into account the network topology and various node attributes. Experimental results from real networks, after comparing five initial coverage set selection algorithms and five centrality indices, indicate that the proposed method, utilizing layer partitioning and node features, offers more accurate and efficient detection of critical nodes in different network types with enhanced robustness and performance over existing methods.

In intelligent situational awareness application scenarios, multi-agent angle tracking problems often occur when moving targets must be monitored and controlled. In contrast to traditional target tracking, the angle tracking task entails not only tracking the spatial coordinates of the target, but also determining the relative angles between targets. Existing control methods often exhibit unstable effects and reduced performance when addressing large-scale problems that are susceptible to environmental changes. To address this problem, the present study proposes a solution scheme based on Multi-Agent Reinforcement Learning(MARL). First, a basic model of the multi-agent angle tracking problem is established, a multi-level simulation decision-making framework is designed, and an adaptive method is proposed for this problem. As a stronger multi-agent reinforcement learning algorithm, AR-MAPPO enhances learning efficiency and model stability by dynamically adjusting the number of data reuse rounds. The experimental results show that the proposed method achieves higher convergence efficiency and better angle tracking performance than traditional methods and other reinforcement learning methods in multi-agent angle tracking tasks.

With the rapid advancement in deep learning, deepfake technology has gained significant momentum as a form of image manipulation based on generative models. The proliferation of deepfake videos and images has a detrimental sociopolitical impact, highlighting the increasing significance of deepfake detection techniques. Existing deepfake detection methods based on Convolutional Neural Networks (CNN) and Vision Transformers (ViT) commonly suffer from challenges such as large sizes of model parameters, slow training speeds, susceptibility to overfitting, and limited robustness against video compression and noise. To address these challenges, a multi-scale deepfake detection method that integrates spatial features is proposed herein. Firstly, an Automatic White Balance (AWB) algorithm is employed to adjust the contrast of input images, thereby enhancing robustness of the model. Subsequently, Multi-scale ViT (MViT) and CNN are separately utilized to extract the multi-scale global and local features, respectively, of the input images. These global and local features are then fused together using an improved sparse cross-attention mechanism to enhance the recognition performance of the model. Finally, the fused features are classified using a Multi-Layer Perceptron (MLP). According to the experimental results, the proposed model achieves frame-level Area Under the Curve (AUC) scores of 0.986, 0.984, and 0.988 on the Deepfakes, FaceSwap, and Celeb-DF (v2) datasets, respectively, demonstrating strong robustness in cross-compression experiments. Additionally, comparative experiments before and after specific model improvements have validated the gains provided by each module in terms of detection results.