The function-level code generation is an important task in the combination of software engineering and artificial intelligence, which aims to improve the productivity of software development by automatically generating function-level code based on task descriptions. However, this task currently suffers from several problems: 1) in the fine-tuning phase, current large language models cannot sufficiently capture the detailed syntactic structure of the code dataset; 2) previous retrieval-augmented methods cannot adequately consider the complex dependencies between code snippets in external code repositories; 3) existing large language models introducing self-repair mechanism tend to overfocus on previously generated erroneous code during the self-repair process. To solve these problems, we propose a retrieval filtering and thought enhancement framework for function-level code generation based on large language model. In our model, we design an abstract syntax tree mapping preprocessing module to preprocess the dataset and help large language models learn the detailed syntax information of the dataset. Furthermore, we design a retrieval filtering and thought enhancement module to retrieve the most relevant snippets of code for the task and to enhance the chain-of-thought of our model. In addition, we design a self-repair mechanism to prevent large language models from overfocusing on generated erroneous code, helping them explore more solutions to repair the erroneous code. We experimented and evaluated our model on HumanEval, MBPP, and MultiPLE benchmarks to compare with other baseline models.

Surface defect detection is an extremely challenging task. Surface defects typically exhibit weak appearances and complex boundaries. Some detection methods are limited by their high computational and storage costs when applied to devices with limited resources. Although some lightweight methods can achieve real-time inference speeds, their accuracy in detecting defects in complex scenarios remains insufficient. To address this challenge, we have designed a lightweight defect segmentation network with external semantics and high-frequency information. Firstly, we devised an External Semantic Attention (ESA) mechanism with linear complexity. This attention learns potential semantic relationships between samples by computing the correlation between input samples and external storage units in both spatial and channel dimensions, thus enhancing the network's attention to defects of weak appearances. Secondly, we designed a High-Frequency Detail Perception Block (HDPB) that enables interaction between image features and high-frequency information, allowing the network to actively learn defect boundary features and capture more details of surface defects. Extensive experiments on three publicly available datasets have shown that our method outperforms existing lightweight segmentation methods.

Document-level entity relation extraction is an important task in the field of natural language processing, which plays an important role in semantic understanding and knowledge graph construction. However, existing deep neural networks and graph neural networks models are limited by their performance and parameters number, which can not capture global semantics and have poor generalization ability. Furthermore, existing methods employing large language model for entity relation extraction do not establish good relationships among multi-tasks of entity relation extraction task, resulting in more information can not be effectively shared and transmitted between tasks. In addition, previous approaches can not effectively eliminate false entities and relationships. To solve these problems, we propose an entity relation extraction framework based on large language model and multi-tasks iterative prompt engineering. In our model, we design an iterative prompt engineering, which can better establish the relationship among multi-tasks, and ensure every task to obtain the optimal results. Moreover, we design semantic merging, group disambiguation and self-verification modules to eliminate the false entity relations and noise nodes. Additionally, we design summary prompts to provide sufficient global semantics for better text segmentation. Finally, we evaluated our model on wikiann, wikineural, ACE2005, CoNLL2003, CoNLL2004, and SciERC datasets and compared it with other baseline models.

Post-hoc explanations are important for people to understand the predictions of explanation models. One class of methods in post-hoc explanation is the generation of counterfactuals, where a hypothetical example is obtained by perturbing the inputs to show how one could obtain a different prediction from the decision model. Counter-factual explanations should satisfy several properties: One is that counterfactuals generated under specific scenarios and constraints should be feasible for users, i.e., they should accommodate different causal constraints. The other is that it is more desirable for users to have a wider variety of viable examples, i.e., counterfactual diversity. To this end, we propose a parallelizable method based on gradient optimization. We partition the input feasible domains, perform counterfactual generation independently for each feasible domain, and then parallelize the counterfactual generation process for each feasible domain. Experimental results show that our approach effectively improves the diversity, sparsity, and proximity of the generated counterfactual instances on the public datasets Adult-Income, Lending-Club, German-Credit, and COMPAS compared to other models.

The objective of Multimodal Knowledge Graph Completion (MKGC) is to forecast absent entities within a knowledge graph by leveraging additional textual and visual modalities. Existing studies commonly utilize a singular relationship embedding to depict all modalities within an entity pair, thus connecting several relationships derived from diverse modalities. However, this coupling may introduce interference from conflicting information between modalities, as the relationships between modalities for a given entity pair can be contradictory. Moreover, existing Ensemble Inference methods fail to dynamically adjust modal weights based on their differences and importance, despite the varying contributions of different modalities. In this paper, we propose the Multimodal Decouple and Relation-based Ensemble inference (MDRE) model. For each modality, we construct corresponding relationship embeddings and build separate triple representations to avoid interferences among modalities. During the training phase, we employ confidence-constrained training with temperature scaling to alleviate conflicting information in textual and visual modalities. For inference, we utilize the Relation-based Ensemble Inference method to adjust modal weights at the relationship level, thus achieving improved prediction results. Experimental results on two datasets demonstrate that MDRE outperforms existing single-modal and multimodal knowledge graph completion methods in terms of performance.

We present GauKGT5, a sequence-to-sequence model proposed for knowledge graph completion (KGC). Our research extends the KGT5 model, a recent sequence-to-sequence link prediction (LP) model. GauKGT5 takes advantage of textual characteristics inherent in the knowledge graph, exhibiting a small model size. However, KGT5’s proficiency in link prediction necessitates the ensemble with a knowledge graph embedding model, which itself poses challenges due to its substantial size and expense. By integrating the Gated Attention Unit into the KGT5 model and directly applying it to the encoder-decoder structure, we achieve improved contextual dependency capturing within the sequence, resulting in enhanced prediction accuracy, accelerated training speed, and enhanced computational efficiency. At the same time, we introduce parallel computing as a means to enhance the efficiency of model training and inference within the XPU distributed computing environment.

Object detection on unmanned aerial vehicles (DAVs)-captured scenarios play an essential role in several applications such as surveillance, environmental monitoring, security, disaster response strategies, and construction of transportation systems. Images captured by DAVs are all overhead vision including too many small objects, which are difficult to detect. Besides, the high-speed and low-altitude flight process of DAVs brings in the motion blur on the densely packed objects. The average, scale transformation and scene coverage are large, which brings difficulties in extracting and identifying useful information. To address these challenges, we propose a lightweight detection model named MA-YOLO. This article has made the following improvements based on YOLOv5:1)a multi-directional weighted pyramid structure (MiFPN) is proposed for fusing information of different scales and improves the ability to detect small objects.2) A learning-capable decoupling head (AD-head) is proposed to obtain small object information in a complex environment. Extensive experiments are conducted on the challenging VisDrone-DET2021 dataset to evaluate the performance of MA-YOLO. The obtained results show that the accuracy is better than other detection algorithms and the detection speed of VisDrone-DET2021 is improved from 85FPS to 109FPS. Thus, the MA-YOLO method pursues a trade-off between speed and accuracy compared to the state-of-the-art small object detection methods and ensures practicality on drones.

The defect detection of Printed Circuit boards (PCB) is challenging due to the complex image background, various types of defects, and small size of defects. This paper develops and evaluates the SC-YOLOv5 network for accurately detecting printed circuit boards. First, we combine the spatial attention mechanism of SA with the Efficient Channel Attention Mechanism (ECA-Net) channel attention module to construct the hybrid attention mechanism module (SCA). SCA has higher defect feature expression ability and doesn't need dimensionality reduction. Second, we analyze the feature pyramid structure of YOLOv5 and construct a multi-direction dilated convolution module (MD) for the last feature layer. MD has a rich receptive field so that MD can retain more defect information during the feature pyramid downsampling process. We perform experimental evaluations on PCB Dataset and DeepPCB datasets. Experiments show that SC-YOLOv5 improves mAP by 1.9% and 1.6 %, respectively, on the two datasets, and the detection speed can reach 120FPS. Compared with the mainstream indication defect detection algorithms, SC-YOLOv5 significantly improves accuracy.

Abstract This study takes diffusion capacity, absorptive capacity, and relationship strength as the main influencing factors, constructs models of knowledge flow in small‐world networks and scale‐free networks, and uses numerical simulation to observe the flow characteristics of explicit knowledge and tacit knowledge in different networks. The knowledge of explicit flow in different networks exhibits the phenomenon of knowledge emergence, but this phenomenon is more obvious in a small‐world network. The flow of tacit knowledge in a small‐world network has a better effect. In a scale‐free network, the quantity and frequency of knowledge flow are significantly higher than those in a small‐world network. The reason for this phenomenon is the differences in the response, connection and structure of different networks. The quantity and frequency of the flow of explicit knowledge in the same network are significantly higher than those of tacit knowledge. The reason for this phenomenon is the different types of knowledge flow in different modes, with different levels of flow difficulty and flow sustainability. First, this study visually compares the differences in flow between the two types of knowledge. Second, flow models of the two types of knowledge are constructed, and the flow characteristics of the two types of knowledge in different networks are simulated. Finally, the reasons for the differences in flow between the two types of knowledge are explained by using loosely coupled theory.

Object detection in aerial images is a critical and essential task in the fields of geoscience and remote sensing. Despite the popularity of computer vision methods in detecting objects, these methods have been faced with significant limitations of aerial images such as appearance occlusion and variable object sizes. In this letter, we explore the limitations of conventional neck networks in object detection by analyzing information bottlenecks. We propose an enhanced neck network to address the information deficiency issue in current neck networks. Our proposed neck network, which serves as a bridge between the backbone network and the head network, comprises a global semantic network (GSNet) and a feature fusion refinement module (FRM). The GSNet is designed to perceive contextual surroundings and propagate discriminative knowledge through a bidirectional global pattern. The FRM is developed to exploit different levels of features to capture comprehensive location information. We validate the efficacy and efficiency of our approach through experiments conducted on two challenging datasets, DOTA and HRSC2016. Our method outperforms existing approaches in terms of accuracy and complexity, demonstrating the superiority of our proposed method.

Many mechanisms, such as epidemic spread, rumor spread, and the spread of social emergencies, are closely related to complex network dynamics, and mining their key nodes plays an important role in understanding the structure and function of the network and maintaining its stable operation. In response to the problem that the key node identification methods in complex networks cannot comprehensively consider global and local information and ignore low-degree nodes, this study proposes a new method based on the voting mechanism. Firstly, the CI value of the network nodes is calculated using the CI algorithm, and initialized the voting ability of nodes by CI values, fully considering the local information of the nodes as well as the influence of low-degree nodes. Secondly, the concept of voting probability is introduced to distinguish the votes of network nodes for their different neighboring nodes through the voting probability, to consider more local information, and to comprehensively assess the importance of the nodes, and ultimately, it is more important to get nodes with the larger voting score. Comparing several classical key node identification methods, the experimental results show that this method can effectively identify key nodes and has a high accuracy rate in different complex networks.

Session-based recommendations already play an important role in platforms such as e-commerce and streaming media, which are designed to predict the next interaction item based on a given session. Most of the current recommendation models only use the interaction sequence of the session to capture the potential conversion patterns between items, often ignore the user’s multi-type interaction behavior that reflects the user’s fine-grained preferences. At present, most models of multi-type interaction behaviors only learn user-item multi-type interaction behaviors and item-item dependencies relatively independently, and ignore the problems of item cold start and data sparsity. These issues motivate us to propose a new model MKGTN in this paper, we apply multi-type user-item interaction behaviors and item-item dependencies to session recommendation via MLP. Simultaneously, using a multi-task learning MLT paradigm involving learning knowledge embeddings as an auxiliary task to facilitate the main task of SR. Evaluations on three datasets show that MKGTN outperforms state-of-the-art multi-action interaction models, demonstrating the superiority of our model’s performance.

Knowledge graph representation learning aims to obtain its vector representation by mapping entities and relations in knowledge graphs to a continuous low-dimensional vector space by learning methods. Most of the existing knowledge graph representation learning methods only consider the single-step relation between entities from the perspective of triples and fail to effectively utilize important information such as ordered multi-step relation paths and entity descriptions, thus affecting the ability of knowledge representation learning. We propose a knowledge graph representation learning model that integrates ordered relation paths and entity descriptions in response to the above problems. The model can integrate the triple representation in the knowledge graph, the semantic representation of entity description, and the representation of ordered relation paths for training. On the FB15K, WN18, FB15K-237, and WN18RR datasets, the proposed model and baselines are run on the link prediction task. Experimental results show that the model has higher accuracy than existing baselines, demonstrating the effectiveness and superiority of the method.

Knowledge graphs use triples containing head entity h, tail entity t, and relation r to represent real-world entities and their intrinsic relationships. In order to predict the actual missing triples in the knowledge graph, combine the strong triple representation ability of relational memory network with the powerful feature processing ability of capsule network, and use disperse dynamic routing that can improve the performance of capsule network, we propose a knowledge graph embedding model named RDMCapsE. First, the embedding vector is formed by encoding potential dependencies between entities and relations; then, different feature maps are generated using convolutional kernels with different window sizes and then reorganized into corresponding capsules; finally, the connection from the parent capsule to the child capsule is specified by the squash function and the disperse dynamic routing, and the credibility of the current triple is judged based on the score of the inner product of the child capsule and the weights. The experimental results show that compared with other models, the model in this paper can effectively improve the effect of knowledge graph completion and the classification accuracy of triples on WN18RR, FB15K-237, WN11, and FB13 datasets.

In this article, we aim to improve the performance of visual tracking by combing different features of multiple modalities. The core idea is to use covariance matrices as feature descriptors and then use sparse coding to encode different features. The notion of sparsity has been successfully used in visual tracking. In this context, sparsity is used along appearance models often obtained from intensity/color information. In this work, we step outside this trend and propose to model the target appearance by local covariance descriptors (CovDs) in a pyramid structure. The proposed pyramid structure not only enables us to encode local and spatial information of the target appearance but also inherits useful properties of CovDs such as invariance to affine transforms. Since CovDs lie on a Riemannian manifold, we further propose to perform tracking through sparse coding by embedding the Riemannian manifold into an infinite-dimensional Hilbert space. Embedding the manifold into a Hilbert space allows us to perform sparse coding efficiently using the kernel trick. Our empirical study shows that the proposed tracking framework outperforms the existing state-of-the-art methods in challenging scenarios.

With the advent of the big data era, traditional industrial production and industrial manufacturing begin to turn to intelligent manufacturing, workshop production process becomes more intelligent and automated, multi-objective job shop complex networks model based on data information, one of the new fields of intelligent manufacturing research. The main contents of this paper are two points: One is to build a multiobjective job shop complex networks model using data information; Secondly, on the basis of the model, the key node of the model is found by using the fuzzy network analysis method, and the result of the key node evaluation is more objective by choosing the fuzzy network analysis method. Finally, the simulation results show that the model can be well applied to actual industrial production, and the fuzzy network analysis method can effectively mine the key nodes in the model, which proves the rationality and effectiveness of the method.

Videos taken under low lighting condition usually have serious loss of visibility and contrast and are inconvenient for observation and analysis. To solve this problem, this paper presents a real-time night video enhancement approach. As observed that a pixel-wise inversion of a night video has quite similar appearance with the video acquired at foggy days, we use the similar idea of haze removal method to enhance the perceptual quality of the night videos. We present an improved dark channel prior model and integrate it with local smoothing and image Gaussian Pyramid operators. The experimental results demonstrate that the proposed approach can improve the perceptual quality of night videos in real-time in terms of not only enhancing details, but also effectively avoiding excessive enhancement phenomenon.

Recently, sparse representation based visual tracking have been attracting increasing interests. Although reported desired performance, whether the sparse representation constrain is really useful is not clear. In addition, the high computation complexity also limits their usage in real-time applications. In this paper, we proposed a real-time visual tracking framework using ℓ2norm regularization based collaborative representation. Our framework represents any target candidate using a set of target templates and a set of background templates respectively, then combines their reconstruction errors to track the target accurately. By constraining ℓ2norm regularization on the representation coefficients, the coefficients can be solved analytically, which makes the proposed method run in real-time. The experimental results demonstrate that the proposed approach outperforms several state-of-the-art trackers.