The purpose of camouflage target detection is to accurately and efficiently detect camouflage targets hidden in the surrounding environment.However，most current camouflage target detection methods cannot accurately identify camouflage targets based solely on their local RGB features when encountering highly similar background interference.Considering the strong differences in frequency domain features between disguised objects and backgrounds，this paper proposes a frequency domain guided network （FGNet） that fully utilizes the advantages of RGB and frequency domain for camouflage target detection.The network is designed with a frequency domain perception module，which utilizes multi-level features extracted from the backbone network to automatically learn and enhance high-frequency and low-frequency features.Subsequently，a dual-strategy collaborative feature extraction method was designed，an edge guided feature module was designed for high-level features to extract edge features of disguised targets through high-frequency feature guidance.A texture information enhancement module was designed for low-level features to enhance the texture of disguised targets through low-frequency feature guidance.Rapid initial positioning and preservation and supplementation of details were achieved through dual-strategy collaboration.Finally，a contextual feature aggregation module was designed to enhance the feature representation of disguised targets through cross layer feature fusion and prior guided correction.Compared with the latest methods，The method proposed in this paper demonstrates superiority on three commonly used benchmark datasets：CHAMELEON，CAMO Test，and COD10K Test.

In cross-modal image-text retrieval (ITR),the transformer-based cross-modal pre-training paradigm is currently mainstream.Pre-training methods typically involve collecting large-scale data to enhance the model's performance across various downstream cross-modal tasks.To this end,a data augmentation method is proposed to generate a large amount of diverse,high-quality text-image data for pre-training; secondly,a two-stage training method combining knowledge distillation and contrastive learning is proposed,which is trained on the dataset produced by the proposed data augmentation method,thereby further enhancing model performance.The proposed model achieves state-of-the-art (SOTA) results on Chinese text-image retrieval datasets,including COCO-CN and Flickr30K-CN.

Traditional automatic indexing methods suffer from low accuracy and heavy reliance on manual review,often overlooking the potential of deep learning,particularly text representation techniques like word embeddings,in text classification and keyword extraction.This paper proposes an innovative end-to-end model based on capsule networks,addressing the challenges of processing large volumes of data and enhancing accuracy in the automatic indexing of digital resources.First,the pre-trained language model BERT is employed for text content encoding and word vector construction.Then,the incorporation of topic and attention capsule improves the performance of keyword recognition and text classification,enhancing training and inference efficiency.Finally,the end-to-end network structure capable of executing both tasks within a single framework is realized.Experiments on real digital resource datasets demonstrate that the model surpasses several existing methods in the metrics such as accuracy,recall rate,and F1 score,tackling the task of automatic indexing effectively for large-scale digital resources.

Antibodies are proteins produced by B cells that play a crucial role in the immune system.This study addresses the challenges in modeling the complementaritydetermining regions （CDRs）of antibodies by proposing a deep learning-based antibody structure prediction model （AbFold）that integrates complementary information from homologous sequences and protein language models.First，extract evolutionary constraints from homologous sequences of antibody sequences and analyze the pairwise interactions among amino acids.Concurrently，amino acid dependencies within antibody sequences are obtained from the antibody language model.Then，a fully connected neural network is designed to combine this complementary information.Finally，the structural modules are iteratively optimized using attention mechanisms and variational autoencoders to predict the threedimensional structure of antibodies.Experimental results indicate that the AbFold model effectively predicts antibody structures，with the accuracy of CDR predictions exceeding that of comparative methods.

Storing the feature embedding vectors required by large recommendation models consumes considerable memory resources,and the frequent querying and cross-layer transmission of these vectors can become a bottleneck during inference.While GPU offers TB/s-level bandwidth,leveraging GPU memory for storing and accessing embedding vectors can improve inference performance,GPU memory is costly and has limited capacity,preventing it from holding all embedding vectors.This paper addresses the characteristic data skewness observed in recommendation scenarios to develop a hybrid storage system tailored for large-scale recommendation model inference.This system optimizes inference by harnessing the high bandwidth of GPU while reducing costs through the use of DRAM as secondary storage.The experimental results demonstrate that,compared to conventional implementations utilizing memory for storing embedding vectors,our system achieves a 14-fold increase in throughput for the embedding vector component.Furthermore,when compared to other implementations that employ GPU for embedding vector storage,our system’s approach to managing the embedding table yields a 3.8x enhancement in throughput for the embedding vector component.

In password-based authentication systems,the leakage of password files poses a severe security threat.Honeywords are fake passwords injected into databases to detect password leakage incidents.Generating honeywords that are difficult for attackers to distinguish is key to the success of this technology.Real-world attackers often possess semantic awareness,and users tend to set passwords containing natural language information.Therefore,considering semantic information is crucial when generating honeywords.Contrastive learning is a method to enhance the semantic representation of pre-trained models.Based on these insights,this paper propose a novel honeyword generation technique based on contrastive learning—PassConR2T.This technique leverages the natural language information of pre-trained models to enhance password representation and further improves the semantic representation of passwords through contrastive learning.This paper design various attack scenarios and evaluate our method on a large real-world dataset.Experimental results show that PassConR2T performs excellently in coping with semantically-aware attackers,generating honeywords that are difficult to distinguish semantically.

Temporal logic is an important part of formal specification in embedded realtime systems,but existing methods of generating temporal logic from natural language requirement often focus on a single domain and lack generalization.The proposed learning-based method DeepTL can generate multiple temporal logics from natural language requirements in different domains.DeepTL includes three parts:atomic proposition recognition,temporal logic structure translation,and atomic proposition translation.After decomposed into these three parts,the translation process requires less model training data,and DeepTL makes temporal logic structure translation no longer limited to specific domains.Experiments are conducted in three domains,and the results show that DeepTL achieves high translation quality while reducing dependence on a large amount of training data and computing resources.

To enhance the accuracy of recognizing nested entities within Chinese texts,this paper proposes a novel Chinese nested named entity recognition model named DAMCNER (Data Augmentation and Multi-scale Convolution based Named Entity Recognition).The model initiates by generating embedded representations of the original input through a pre-trained model and semantically enriches these embeddings.Subsequently,a multi-head dual affine attention mechanism is utilized to construct a span feature matrix,which is further refined through multi-scale dilated convolution layers and a content-based attention mechanism.Span decoding is then performed using a multi-layer perceptron.Moreover,a data augmentation module has been designed to enhance the diversity of data samples,thereby improving the model's robustness and generalization capabilities,and further boosting the accuracy of entity recognition.Experimental results on three publicly available Chinese nested entity datasets show that the DAMCNER model outperforms existing baseline models,with an average increase of 1.52% in the F1 score.The experiments confirm that the DAMCNER model is effective in various scenarios,significantly enhancing Chinese nested named entity recognition.

The instruction set architecture(ISA)is a key technology that bridges the gap between software and hardware.The RISC-V ISA gradually stands out due to its simplicity,modularity and open-source features.It allows designers to flexibly extend it to enhance processor functionality and performance,meeting the needs of certain scenarios.The RISC-V standard organization continuously introduces new standard extensions to adapt to the ever-changing application requirements.This paper focuses on RISC-V ISA extensions,first explains the architecture of the RISC-V ISA as well as its existing extensions,and analyzes the design process of custom extensions.It then presents general methods for implementing instruction set extensions in both hardware and software.Finally,with the help of some specific cases,the paper explores the practical applications of RISC-V ISA extensions in artificial intelligence,high-performance computing and post-quantum cryptography.Finally,on this basis,this paper looks forward to the future development directions of RISC-V ISA extensions.

The influence relationships between individuals form a diffusion network where people are nodes.Understanding the topology and strength of these influence relationships in the diffusion network is crucial for understanding the mechanisms of past diffusion events and predicting future diffusion trends.However,in practice,these relationships are difficult to observe directly,and can typically only be inferred from the infection data of historical diffusion processes.Most existing methods for diffusion network inference rely on precise infection timestamps,which is hard to obtain and limits the practical applicability of these methods.This paper proposes to infer the propagation network using only the final infection states of nodes in various historical diffusion processes,rather than precise infection timestamps.First,the inference problem is modeled as a nonlinear programming problem involving the topology and strength of influence relationships.The potential parent nodes influencing each node are pruned by examining the statistical correlation of infection results between nodes,narrowing the solution space and obtaining an initial solution.Then,a differential evolution-conjugate gradient hybrid optimization method is applied to further optimize the initial solution,yielding the inferred propagation network.Experimental results demonstrate that the proposed method outperforms existing representative methods in terms of inference accuracy.

In recent years，many studies have focused on emotional support models.However，the evaluation method is only normal text generation indicators，which cannot intuitively reflect the emotional support ability of the model in real interactions.Although interactive evaluation by human is a great way to made up for this defect，it is too much cost to carry.In order to evaluate emotional support ability better，this paper proposes a model SeekerView based on the state of the help seeker to simulate emotional support interaction and designs some evaluation indicator via the model.The model can predict the emotional state of the help seeker and generate specific dialogue content based on this，from which the evaluation indicators for emotional support task can be designed，such as the number of dialogue rounds and the scores of the help seeker′s emotional relief.At the same time，this paper evaluates 7 large language models with the evaluation indicators.Compared with many other evaluation methods，the evaluation indicators designed in this paper are more consistent with the scores of the warmth dimension and the emotional validation dimension in interactive evaluation by human.

The existing review-based and rating-based methods usually use the same model to model users and projects respectively,but they are limited to the shallow feature level,and if we can fully explore the user's personalized preferences and the deep characteristics of the project,it will promote the model to learn the deeper relationship between the two representations and improve the prediction results.Consequently,we propose a personalized recommendation method that combines rating and review,which is used to deeply mine user preferences and project features.In the process of processing the comment text,the vector representation of the words in the comment text is obtained through ALBERT.Secondly,the proposed personalized attention module combines the user's personalized preference information with the comment text vector to get a deep comment-based user representation.Experiments are carried out on Amazon Digital Music,Grocery and Gourmet Food and Video Games data sets,and the NDCG index of this method is improved by 5%,11% and 8% respectively compared with the benchmark method.The code is publicly available on https://github.com/ZehuaChenLab/paperCode/tree/main/DuWenNa/PRM-RR.

Channel equalization is one of the key techniques to solve the inter-code crosstalk problem of UAV communication signals.In order to reduce the impact of inter-code crosstalk due to channel distortion during UAV communication signal transmission,accelerate the convergence speed of equalization,and reduce the bit error rate of the communication system,this paper proposes a new variable step-size dual-mode(VSSDM)blind equalization algorithm,which processes the real and imaginary parts of the signal at the receiving end in parallel during the equalization stage to correct the carrier phase rotation problem introduced by the radio channel.The error control function is adjusted according to the convergence circle switching criterion,and variable step size control is introduced to accelerate the convergence speed and further reduce the steady state error.Simulation results show that the algorithm has better convergence performance than the traditional equalization method,the ISI is reduced to -27dB in the late stage of equalization,the convergence speed is improved,and the algorithm completes the convergence in about 600 symbols.

To enhance the environmental perception capability of hexapod robots on diverse terrains,we propose a hexapod robot gait recognition network model that integrates a multi-head attention mechanism-based Temporal Convolutional Network (TCN) and a contact pre-classification module.The TCN network with multi-head attention mechanism effectively captures important features in the force-time series of the foot end and associates them with the gait of the hexapod robot.The contact pre-classification network can preliminarily extract features of the contact force signals.The multi-head attention mechanism helps the model better focus on key information,model long-term dependencies,adapt to different gait characteristics,and handle gait variations under different speeds and terrain conditions.The contact pre-classification channel can be fused with the TCN network with multi-head attention to predict,thereby improving the accuracy and robustness of gait recognition.Experiments show that the proposed model has good generalization performance,with an overall accuracy improvement of 0.9% over the TCN network with multi-head attention,and improvements of 1.6% and 0.4% compared to GRU and Transformer networks,respectively.Additionally,it exhibits higher noise resistance.

A single large non-uniform hypergraph clustering is designed to divide the nodes contained in non-uniform hypergraph into multiple clusters，so that the nodes in the same cluster are more similar，while the nodes in different clusters are less similar，which has a wide range of application scenarios.MADC（Non-uniform hypergraph clustering combining multi-scale attention and dynamic construction）is the optimal non-uniform hypergraph clustering method based on hypergraph neutral network，but MADC also has a deficiency.For example，the learning efficiency of hypergraph feature embedding needs to be improved.Based on this optimization deficiency，STHC（Self-training nonuniform hypergraph clustering），a model based on multi-scale attention and self-training networks is proposed.On the one hand，STHC proposed a parallel multi-scale attention network in the self-training hypergraph learning module to efficiently learn hypergraph feature embedding.On the other hand，the self-training network is used to jointly optimize the hypergraph feature embedding and clustering results.Extensive experiments on real datasets demonstrate that the clustering accuracy（ACC），normalized mutual information（NMI）and adjusted Rand index（ARI）of STHC model on non-uniform hypergraph clustering are better than all baseline methods.

To address the challenges of task misalignment between upstream and downstream tasks,as well as low extraction accuracy in Chinese event relation extraction using pre-trained language models,a novel model is proposed to enhance Chinese event relation extraction by coordinating knowledge and relation semantics through prompt-based learning.First,prompt templates and label encoders are constructed using the prompt learning method to bridge the gap between upstream tasks and downstream relation extraction.Next,a prompt template reconstruction strategy is designed specifically for the Chinese context,integrating relation label semantics and common background knowledge.This strategy introduces high-quality prior knowledge and external commonsense knowledge into the prompt templates,providing a dual enhancement of relation semantics and knowledge.Finally,RoBERTa encoding is employed to perform multi-label classification of event relations.Experimental results show that the proposed model achieves an F1 score of 90.88% on the CEC2.0 dataset,surpassing most current baseline and advanced models,and maintains strong performance even in few-shot scenarios.

Adequate testing is a necessary step to ensure the normal operation of modern airborne software,but traditional software testing methods often fail to meet the requirements.To this end,we design a test case automatic generation method for the field of aviation software.After constructing a variable-relationship model for aviation software requirements described in natural language,we can parse the semantics of the model to create a semantic tree of requirements,select test paths according to the safety-criticality level,and generate different coverage sets according to the method.In order to alleviate the problem that conditions are related to each other and may cause the failure of modified condition/decision coverage,we define the coupled conditions judgment and constraint criteria applicable to the demand model,propose three forms of modified condition/ decision coverage and give the corresponding coverage set generation method.In addition,we design a reasonable and effective test case selection strategy based on the equivalence classes and boundaries determined by the coverage sets to automatically generate the test case set.Finally,we demonstrate the feasibility of this paper's approach by performing formal modeling and test case generation on an example of aviation software requirements.

Joint modeling or multi-task learning can leverage large-scale speech recognition and text translation data to enhance end-to-end speech-to-text translation performance.However,most existing methods usually require architectural adjustments to the speech translation model or rely on multi-stage pre-training and fine-tuning.Moreover,the modality gap between speech and text poses challenges in using a shared encoder to process both modalities simultaneously.To address these issues,this paper proposes a simple multi-modal joint modeling framework.This framework regards the joint modeling of speech translation and text translation as multilingual neural machine translation modeling.It introduces modality-aware relative position encoding in the self-attention layer and employs a single encoder with modality awareness to encode speech and text without complicating the model's architecture,and utilizes large-scale speech recognition data,and text translation data selected by the proposed translation loss method for multi-modal joint modeling training.Experimental results on two benchmarks show that compared to baseline methods,using single encoder approach significantly improves the performance of bidirectional translation(from English and to English)on multiple speech translation tasks when jointly modeling internal and external speech recognition and text translation data.

To resolve the conflicts among constraint handling,convergence,and diversity in constrained multi-objective optimization problems,this paper proposes an constrained multi-objective particle swarm optimization algorithm with improved multi-subswarm cooperation.The proposed algorithm divides the population into one main swarm and multiple sub-swarms.The main swarm focuses on exploring the feasible region,while the subswarms concentrate on exploring the solution space without considering constraints,thereby effectively balancing the three objectives.For the subswarms,a cyclic dynamic regroup strategy is designed to periodically regroup them,enhancing diversity.For the main swarm,a dynamic jitter update strategy is introduced to help escape local optima by adding a jitter term;a distribution diversification enhancement strategy is developed,using weight vectors to select the offspring population,ensuring uniform distribution of the solution set;and a progressive boundary constraint strategy is designed to gradually shrink the search space as iterations proceed,ensuring both constraint satisfaction and convergence.Experimental results demonstrate that,compared to similar algorithms,the proposed algorithm performs best on 14 test cases and achieves optimal results in 10 of them.

In the contemporary social media environment,personalized recommendation algorithms play a crucial role.They enhance user experience by analyzing user characteristics to provide targeted content or suggest potential new friends.However,these algorithms can also lead to some unintended consequences,with the most concerning being the phenomenon of opinion polarization.While previous research on social media link recommendation mechanisms often focused on the similarity of user network structures and opinions,there has been little exploration into the heterogeneity of user activity levels and behavioral information.In response to this,this paper proposes a link-based recommendation Deffuant-Weisbuch model(LRDW),which,in addition to considering opinion conflicts among users,incorporates user activity into the recommendation strategy.Furthermore,this study explores the specific impact of different recommendation strategies on opinion polarization.Simulation results on both artificial networks and real social media networks show that recommendation mechanisms based on user activity levels can intensify opinion polarization.Additionally,this paper finds that by controlling the opinions of active users,it is possible to some extent to guide the overall direction of public opinion.Lastly,this paper introduces random elements into the recommendation strategy,effectively mitigating the polarization phenomenon.

In view of the problem that during the traffic sign detection process,driving too fast or external environmental factors such as lighting cause traffic signs to deform and blur,resulting in reduced detection accuracy and speed.The article proposes an improved RT-DETR detection algorithm for traffic signs.First,DCNv2 deformable convolution was introduced into the backbone network of the algorithm and the MCCA attention mechanism was designed,thereby proposing the DCNv2att deformable attention convolution block,which effectively generated the offset and adaptively adjusted the target object according to the shape change.Select the sampling position and capture the detailed information of the object boundary more accurately; then,to address the periodic defects existing in the position encoding method of the original algorithm Transformer encoder,a learnable position encoding is adopted,and the model can autonomously learn and adapt to the input sequence Characteristic position information representation,thereby improving the performance of the model; finally,the original GIoU loss function is replaced by the MPDIoU loss function to directly minimize the distance between the upper left and lower right points between the predicted bounding box and the ground truth bounding box,easing the problem when predicting When the bounding box is completely covered by the ground truth bounding box,the loss function cannot optimize the position and size of the anchor box.Compared with the original RT-DETR algorithm on the TT100K data set,the improved algorithm has a comprehensive average accuracy increased by 3.0 percentage points,the calculation amount has been reduced by 18%,and the FPS has been increased by 12 frames,while being lightweight.The detection accuracy and speed are improved,which is significantly better than similar comparison algorithms,and can be practically used in traffic sign detection scenarios.

Patient-ventilator asynchrony during mechanical ventilation will lead to serious prognostic problems,so it is necessary to develop a reliable automatic recognition algorithm.Many current intelligent algorithms,suffer from poor detection efficiency due to the limited datasets,with huge amounts of parameters,which are difficult to deploy on resource-constrained hardware devices.In this paper,we propose a multi-scale lightweight patient-ventilator asynchrony recognition algorithm based on feature aggregation.The algorithm adopts a lightweight parallel architecture and combines the advantages of convolution and attention mechanism to dynamically learn global dependencies while extracting local information of waveform images.At the same time,this algorithm utilizes relative position encoding to construct spatial associations,and designs personalized pruning schemes for redundant structures of the model,which improves the recognition accuracy and reduces the storage consumption,thus enhancing the computational efficiency.In the testing set,the recognition accuracy,sensitivity,and specificity of our algorithm are 0.985,0.982,and 0.986,respectively,which are better than the current algorithms and more efficient,showing high potential for application.

The self-supervised pre-training models for speech can extract general feature representations from a large amount of unlabeled data，thereby addressing the low generalization issue in spoof speech detection algorithms.However，fine-tuning these pre-training models suffers from inefficiencies in parameter usage and high training costs.Therefore，this paper proposes a method of combining self-supervised pretraining models with multiple attention-adapted fine-tuning modules for spoof speech detection，with the self-adaptive low-rank fine-tuning method as the basis.This paper designed and introduced two lightweight adaptation modules，a parallel convolutional module based on spatial attention and a serial feedforward module based on hidden dimension attention，to reduce the trainable parameters and the training overhead by inserting these two modules.These adapters are inserted to reduce the number of trainable parameters during fine-tuning and lower training costs.Moreover，these adapters compensate for the model′s deficiencies in capturing the relationships of spatial position and hidden dimensions，thereby enhancing the model′s detection performance and generalization capability.Experimental results on the ASVspoof2019 LA and PA datasets confirm the effectiveness of the proposed detection method.It reduces the equal error rate of the baseline model by 98% on the LA dataset and by 91.6% on the PA dataset.Cross-dataset testing across multiple datasets further validates the great generalization of the proposed detection method.

At present,there is no publicly available multi-view sign language dataset in China,and sign language recognition research mainly centers on single-view data,and the model recognition effect is poor due to the gesture occlusion problem.Aiming at these problems,a Multi-View Chinese Isolated Sign Language Dataset (MV-CISL) is created;based on this dataset,an isolated sign language recognition method with multi-view feature fusion is proposed,which uses an end-to-end multi-stream network based on improved 3D-ResNet18 to extract feature information from different viewpoints and integrate these feature information through decision-level fusion;to improve the network recognition performance,the proposed network is subjected to migration learning using the CSL-500 single-viewpoint dataset,which is applied to the MV-CISL dataset.The experimental results show that the proposed method outperforms the single-view and dual-view methods in terms of performance;the effectiveness of the method is further validated on the multi-stream network backbone models ResNet+LSTM,ResNet+BiLSTM,3D-MobileNet and 3D-ShuffleNet;and the data collection cost is lower and the performance is better than the methods based on frontal-view RGB and deep information fusion.

Conventional opportunistic crowdsensing requires centralized task allocation by the platform,which poses high demands on platform performance and risks of privacy leakage.To address these issues,this paper presents an innovative solution aimed at decentralized opportunistic sensing scenarios,where task allocation with global budget constraints is implemented to maximize task coverage.Diverging from conventional methods,this solution allows participants to autonomously choose whether to take part in sensing tasks,thus constructing a multi-agent collaborative system.In decentralized settings,ensuring global constraints while achieving efficient task allocation remains a significant challenge.To overcome this challenge,two methods are proposed:one applies the ant colony algorithm to task allocation issues,acquiring the objective evaluation function through communication and updating pheromones to facilitate agent collaboration and suit decentralized scenarios;the other is a decision communication approach based on the QMIX framework,where the output of the agent networks in QMIX serves as suggested actions,accompanied by a decision communication layer that negotiates according to the suggested actions and their values,thereby adhering to global constraints.Experimental results on real datasets demonstrate that the two methods proposed in this paper can match the centralized planning methods in terms of task coverage rate and exhibit commendable comprehensive performance in terms of time consumption and other metrics.

The rapid advancement of autonomous driving technology has propelled the utilization of lidar.Lidar plays a pivotal role in the realm of autonomous driving owing to its exceptional capabilities in environment perception,navigation,and obstacle avoidance.With the continuous progress in artificial intelligence and deep learning technology,3D data processing technology has achieved remarkable outcomes and been implemented across various scenarios.However,as the application of this technology expands further,safety concerns have become increasingly prominent,such as potential misidentification of non-existent objects by moving vehicles.Nevertheless,existing research predominantly focuses on individual tasks and lacks a comprehensive discussion on security issues,particularly regarding backdoor attacks.Therefore,this paper presents an all-encompassing evaluation and analysis of LiDAR-based security concerns in autonomous driving systems with special emphasis on adversarial attacks and backdoor attacks.Firstly,this paper elucidates the operational principles of lidar along with its applications in autonomous driving tasks encompassing object classification,object detection,and semantic segmentation.Specifically exploring 55 relevant papers within this review enables us to systematically introduce attack methods and defense strategies under different tasks.Furthermore,we provide 11 available datasets along with 7 evaluation metrics for researchers' convenience while also presenting 7 commonly utilized models and 4 simulation platforms that serve as valuable resources and tools for their work.Finally,considering current challenges alongside future opportunities allows us to envision prospective research directions for ensuring safe and reliable implementation of lidar technology within autonomous driving systems—aiming to offer guidance and reference points for researchers engaged in enhancing safety applications through lidar.

The Crawfish Optimization Algorithm(COA)is a novel swarm intelligence optimization algorithm that proposed in 2023 and has gained attention due to its simulation of crawfish behavior and temperature regulation.However,COA suffers from diversity degradation,insufficient exploration capability,susceptibility to local optima,and low optimization accuracy.To address these issues,this study proposes a Chaotic Accumulative Difference-Enhanced Crawfish Optimization Algorithm(CE-COA).First,the CE-COA algorithm enhances population diversity by introducing Piecewise chaotic mapping to initialize the positions of individuals.Second,during the hibernation and competition cave stages,elite cave populations are incorporated to prevent the algorithm from getting trapped in local optima,thereby enhancing its ability to discover potential solutions.Furthermore,in the foraging stage,a cumulative difference-feeding strategy is employed to fully consider the differential information differences between individuals and food sources,further improving the algorithm's optimization accuracy.In the experimental analysis phase,the CE-COA algorithm is validated using the CEC2022 test suite and some test functions from CEC2017.The algorithm's performance is evaluated using qualitative analysis,the Wilcoxon rank-sum test,and the Friedman test.Additionally,practical applications of the CE-COA algorithm are demonstrated through case studies involving two engineering design problems and a Wireless Sensor Network(WSN)node coverage problem.The experimental results consistently demonstrated superior performance of the CE-COA algorithm.

To reconstruct high-quality 3D facial models from single 2D facial images,this paper proposes a Deformable Normalized Point Cloud-based 3D face reconstruction method,DNPR.By breaking away from the fixed topology of explicit face prior models,DNPR represents faces as point clouds of geometry and appearance in a canonical space,independent of pose.By optimizing learnable points in the canonical space,it enables effective deformation and rendering within mapping and shading spaces,accelerating convergence towards the target face shape.Dense points in the later stages of training can well reproduce texture details.Comparative experiments on the benchmark dataset provided by IMavatar show that DNPR achieves an average SSIM of 0.854 and an average PSNR of 23.863,showing a certain advantage over many traditional implicit reconstruction methods.Furthermore,with a training speed of 0.06h per epoch,it demonstrates a 4 percentage point improvement over PointAvatar.

In the training of Mixture of Experts(MoE)models,the introduction of expert parallelism significantly alleviates memory pressure on individual nodes and enhances model performance.However,expert parallelism training faces issues of high communication costs caused by frequent token transmissions across nodes and imbalances in node load.To solve this problem,this paper introduces a popularity-based expert prefetching strategy(Prefetch-Expert,PE).The strategy uses the popularity of experts to intelligently predict and prefetch the necessary experts required for current training,thus improving training efficiency.Additionally,for cases where prefetching is unsuccessful,the PE strategy incorporates an asynchronous schedule mechanism that allows for the fetching of other experts while expert computations,facilitating an overlap of expert communication and computation,and effectively reducing communication delays caused by network contention.Extensive experiments conducted on CIFAR-100,WikiText-103,and SQUAD datasets demonstrate that the PE strategy reduces the convergence time of mainstream deep learning models by at least 30% compared to conventional approaches.

Deep Q-Network(DQN)has been widely used in UAV obstacle avoidance path planning task,for the traditional DQN sampling process due to the existence of sample information is not fully utilized,resulting in slow convergence speed,proposed a tree sampling Dueling-DQN based UAV three-dimensional obstacle avoidance path planning scheme.Firstly,the UAV obstacle avoidance planning network system model and simulation environment model in 3D space are described;then,the obstacle avoidance algorithm,power consumption algorithm,and UAV action ensemble are designed;finally,the Dueling-DQN algorithm combined with tree sampling is proposed,which uses the binary tree structure to store the priority samples,and combines with the reward function and the greedy strategy to obtain the obstacle avoidance flight paths of the UAV.The experimental results show that the scheme achieves the highest average reward value while obtaining a better planning path compared to traditional DQN and DDQN(Double Deep Q-Network,DDQN).Under ten obstacle difficulty levels,the minimum number of steps required to reach the target point and the lowest probability of collision are achieved compared to A*,RRT,and ACO algorithms.The simulation results verify the effectiveness of the proposed UAV 3D obstacle avoidance path planning scheme in dealing with the UAV obstacle avoidance planning problem in 3D space.

The research on drug repositioning is of positive significance for improving the efficiency of drug discovery.However,most of the current research focuses on training drugs and diseases under the same supervision mode,ignoring their inherent differences in biological characteristics.At the same time,the highly sparse drug-disease network also seriously interferes with the model prediction.Therefore,this paper proposes a drug repositioning prediction method GL-DDI based on graph convolution neural network and contrastive learning.Firstly,the multi-source similarity between drugs and diseases is integrated through attention mechanism,and the meta-path between drugs and diseases is constructed.Then,the feature extraction of drug and disease subspace is supervised by meta-path scoring matrix.Finally,by comparing the prediction results of drug subspace and disease subspace,a contrastive learning model is constructed to predict the potential indications of drugs.The experimental results show that the area under receiver operating characteristic curve and PR curve of this model are 0.9392 and 0.9442,respectively,which is superior to the existing methods.

Low-altitude logistics is a typical application to develop the new quality productive forces of logistics productivity.In this paper,a three-dimensional path planning model of urban UAV distribution is built around the problem of efficient transportation of materials in low-altitude environment.This model pays attention to the timeliness and cost requirements of distribution activities,reflects the terrain characteristics of urban scenes,and can realize the efficient and low-energy material distribution of drones in urban environment.In order to solve the flight path of the model,a self-adaptive disturbance particle swarm optimization (ADPSO) algorithm is proposed.Latin hypercube sampling,adaptive parameter adjustment and adaptive t-distribution perturbation strategies are introduced respectively to solve the problem that PSO is prone to local optimization and improve the global search performance of the algorithm.Finally,through data experiments and contrast simulation,the results show that the model constructed in this paper and the proposed method can more effectively realize low-altitude material distribution in urban areas under multiple scenarios,especially in complex environments,and the path is shortened by 12.10% compared with the original algorithm.

In medical imaging report generation,differences in semantic spaces between images and text lead to mapping challenges,causing misjudgments and less professional,less fluent reports.We propose a cross-modal contrastive learning method.It first retrieves and compares medical images with patient history to generate an initial report.This report is then refined using a cross-modal semantic synchronization memory unit,which stores features of the input images and real reports.It calculates contrastive loss between the generated report and sample points in the memory unit,which not only narrows the gap with positive samples but also widens the gap with negative samples.This process optimizes the visual-text feature extractor,ultimately producing a report that closely aligns with the image information.Experimental results show that compared to the DeltaNet model,our method improves BLEU-4 by 1.3% and CIDEr by 11.5% on the IU X-Ray and MIMIC-CXR datasets,enhancing accuracy and readability.

As real-time embedded systems become increasingly complex,the need for system time determinism becomes more pressing.The logical execution time (LET) model,as a real-time programming abstraction model,decouples the system behavior from the physical execution time by determining the time interval from reading inputs to writing outputs.The LET model distinguishes itself from the zero execution time (ZET) model and bounded execution time (BET) model by centering on delaying the release of task outputs until the end of the task cycle to eliminate output jitter.Researchers have developed programming languages such as Giotto based on the LET model,and proposed communication synchronization mechanisms and memory access optimization methods for multi-core platforms.Meanwhile,extended forms such as System Level LET address non-zero communication latency and clock synchronization in distributed environments.The contribution of this paper is to provide a systematic overview of the theoretical foundation of LET model,its application implementation,and its extension and optimization,so as to provide reliable technical references for the design of real-time systems in the fields of next-generation smart grid-connected vehicles,industrial controllers,and so on.

The improvement of traditional deep Q learning training algorithm usually focuses on the optimization of reward function，and relatively lacks the self-optimization of strategy and the dynamic adjustment of convergence gradient.Aiming at this problem，this paper proposes a hybrid algorithm PPO-Dueling DQN based on Dueling DQN algorithm.On the one hand，this algorithm can use strategy gradient descent and adaptive KL divergence penalty mechanism to realize synchronous update of strategy function loss and value function loss，and then optimize the model′s loss function and strategy selection.On the other hand，it can extract the state value and action advantage in the game process more real-time，so as to avoid relying on a single index for strategy update and effectiveness evaluation.Through comparative experiments，the optimization effect of PPO-Dueling DQN algorithm for network game model is verified in terms of learning ability，convergence speed and adaptive efficiency，and the parameter analysis of discount factor is carried out to better evaluate the model efficiency.The experimental results show that the proposed algorithm has certain performance advantages compared with the benchmark model.

Math word problem solving is an important task in the field of machine reading and mathematical reasoning,aiming to generate correct numerical answers based on the problem text.The core of this task is to understand the text semantics and generate the corresponding math expressions.Recent advancements in deep learning and the integration of pre-trained language models have significantly improved semantic understanding and problem-solving capabilities.However,existing models still have issues such as insufficient capture of structured relationships in the text and a lack of in-depth modeling of the implicit logic between numerical information.To address these problems,we proposes a Number-Enhanced Graph Convolutional Network (NEGCN).First,NEGCN constructs a numerical comparison graph to capture the magnitudes,types,and interrelationships of numerical values,effectively integrating numerical information with textual semantic features and enhancing the model's numerical reasoning ability.Secondly,NEGCN introduces a structure information extraction mechanism based on graph convolutional networks,converting math word problem texts into graph representations with syntactic dependency relations,and learning syntactic structural features through graph convolutional neural networks to enrich semantic information.Finally,NEGCN employs interactive attention to strengthen semantic interactions between different clauses,extracting global semantic information and enhancing the model's overall understanding of math word problems.Experimental results on the MAWPS and Math23K datasets prove that the overall performance of the NEGCN model is better than that of the comparison models.

Humanoid robots，with their highly realistic human-like appearance and flexible，stable motion capabilities，are capable of deeply engaging in human social activity scenarios.Humanrobot interaction is a necessary capability for humanoid robots to enter the real world and become companions in people′s lives，education，and work.However，current robot interaction technologies mainly cater to the general population，lacking humanrobot interaction schemes tailored for individuals with hearing or speech impairments.To address this，we have designed a lightweight sign language recognition deep network （SignNet），which can be rapidly integrated into robotic hardware platforms，endowing them with sign language recognition abilities.Furthermore，by combining the sign language recognition model with existing expression control algorithms and text-to-speech synthesis algorithms，the robot can not only recognize sign language but also respond through facial expressions and synthesized speech.Experimental results demonstrate that SignNet achieves the highest recognition accuracy of up to 98.2% across multiple sign language datasets （WLASL，CSL，SLR500）.Additional testing reveals that the robot can accurately recognize sign language and provide appropriate feedback，providing an effective human-computer interaction channel for people with hearing or speech impairments.

With the rise of internet media，the supervision of social networks has become increasingly difficult，leading to the proliferation of fake news on online platforms.Identifying fake news accurately and generating interpretable results have become significant research areas.Existing deep learning-based methods for explainable fake news detection suffer from a lack of information and an excessive dependence on social context.The paper proposes an explainable fake news detection model based on the news environment by constructing an environment that reflects the focus of mainstream media and the distribution of public attention.First，the news environment is constructed，and perceptual learning is conducted on the constructed environment to capture objective information within the news environment at the time of fake news creation.Second，we screen out important sentences from the constructed news environment using sentence importance scores for explanation generation.Finally，we integrate the environmental perception vector as auxiliary information to complete the fake news detection.Experimental results on two public datasets show that the model outperforms baseline models in both macF1 and ROUGE metrics.

In recent years，EEG emotion recognition has shown great promise for applications in psychotherapy and human-computer interaction.However，most existing studies have not fully explored the coupling and complementary features of complex spatiotemporal-frequency patterns in EEG signals.In this paper，we propose a network model for decoupling spatio-temporal-frequency fusion features based on multiple attentional mechanisms to purposefully capture the complementary spatio-temporalfrequency domain features of EEG signals.The model extracts key discriminative information in the signals more effectively by decoupling the multi-domain fusion features into a temporal flow module，a spatial enhancement module，and a frequency-domain flow module.This is achieved while aggregating the spatial attention mechanism and frequency-domain attention mechanism into the network model.Extensive experiments conducted on the DEAP dataset show that the model achieves accuracies of 93.68% and 92.96% in the arousal and valence dimensions，respectively，outperforming existing models and demonstrating its superiority in improving emotion recognition performance.

Hierarchical text classification is an important subtask in the field of text classification， but the complex hierarchical structure of labels makes it challenging.The most advanced method currently uses pretrained language model to encode text and combines it with graph encoder to process label structure information.However， independent modeling of text and label information may lead to low information utilization， and different strategies in the pretraining and finetuning stages also limit the knowledge of model mining pretrained models.This article proposes a hierarchical text classification method that combines contrastive learning and prompt tuning.The hierarchical label information is embedded into the text encoding process， and a prompt template is designed to integrate label information， capture the correlation between text and labels， and bridge the gap between pretrained models and downstream tasks.By using contrastive learning， positive samples are generated based on label information to strengthen the model′s learning and retention of key features， effectively guiding the learning of text feature representation.Extensive experiments on two publicly available datasets have demonstrated the effectiveness of the method.