The synergy between foveal and peripheral processing is fundamental to the efficiency of biological vision. While hybrid Convolutional Neural Network (CNN)-Transformer architectures aim to capture both local and global features, they often rely on static, predefined structures that struggle to dynamically align information and adaptively allocate computational resources, ultimately limiting their performance. To address this limitation, we introduce the Central-Peripheral Vision Transformer (CPVT), a novel architecture that explicitly and hierarchically mimics this biological dichotomy. CPVT employs fine-grained, convolutionally modulated attention in its shallow layers to emulate foveal vision, while seamlessly transitioning to a coarse-grained, global attention mechanism in deeper layers to emulate peripheral vision. This design is enhanced by two specialized Feed-Forward Networks that facilitate synergistic information interaction. Rigorously validated on diverse medical imaging benchmarks, CPVT achieves state-of-the-art performance, attaining classification accuracies of 87.98% on the International Skin Imaging Collaboration (ISIC) 2018 challenge dataset and 90.41% on the Kvasir dataset. These results demonstrate that an adaptive, hierarchical integration of biological vision principles can significantly enhance machine perception for medical image analysis.

The Chamfer Distance (CD) is a cornerstone objective function for point cloud completion, yet its inherent symmetric weighting mechanism limits the quality of the generated results. By penalizing local detail deviations and global coverage deficiencies equally, standard CD often causes structural defects such as point aggregation and incomplete spatial structures. We introduce the Flexible-weighted Chamfer Distance (FCD), which decouples CD into local precision and global completeness sub-objectives. FCD employs an asymmetric weighting strategy that prioritizes global structural integrity, steering the optimization away from sub-optimal solutions. As a plug-and-play module with negligible overhead, extensive experiments on state-of-the-art networks demonstrate that FCD significantly enhances global distribution metrics while preserving local precision. Specifically, on the ShapeNet55 benchmark using AdaPoinTr, FCD reduces the Density-aware Chamfer Distance (DCD) by approximately 12.4% (from 0.613 to 0.537), effectively mitigating point clustering. Similarly, on the PCN dataset, the proposed method reduces the Earth Mover's Distance (EMD) from 23.79 to 21.40, demonstrating superior global uniformity compared to the standard CD baseline. Furthermore, FCD demonstrates excellent generalization. When applied to diverse tasks and datasets, including real-world scans (KITTI), industrial components (ABC), and point cloud upsampling (PU-GAN), it yields significant quantitative gains and produces visually more uniform and structurally complete point clouds. These results underscore FCD's potential as a versatile objective function for the broader point cloud generation domain.

Compared with monotherapy, combination drug therapy plays a crucial role in clinical treatment. However, the exponential expansion of the drug combination space has rendered traditional exploration methods for synergistic drug combinations inadequate. Recently, numerous efficient and accurate computational approaches have been developed to predict anticancer drug synergy, particularly those leveraging hypergraphs to model the multifaceted relationships between drug combinations and cell lines, which have demonstrated remarkable potential. Nevertheless, existing hypergraph-based methods fail to account for the heterogeneity of anticancer synergy hypergraphs and overlook the underlying similarities among drugs and cell lines, thereby limiting their ability to fully capture the complex interactions between drug combinations and cell lines. To address these limitations, we propose an Adaptive Heterogeneous Hypergraph Representation Learning Method (HHGSynergy) for predicting anticancer drug synergy, enabling more precise identification of synergistic drug combinations. Specifically, our framework first constructs drug/cell line similarity-based synergy hypergraphs based on the foundational anticancer synergy hypergraph, thereby establishing a comprehensive heterogeneous hypergraph. Next, a node importance calculation module is employed to learn both local and global importance weights of nodes, effectively capturing the structural characteristics of the hypergraph. Finally, a type-specific multi-head attention mechanism is utilized to iteratively update node embeddings, adaptively learning the significance of heterogeneous hyperedges. Experimental results demonstrate that HHGSynergy achieves state-of-the-art performance in both classification and regression tasks across diverse experimental scenarios, outperforming existing leading models. Case studies further underscore its potential for discovering novel synergistic drug combinations.

Circular RNA (circRNA) is a widely distributed class of non-coding RNA molecules that have been shown to play a significant role in cancer development and drug resistance, significantly influencing cellular sensitivity to therapeutic drugs and treatment outcomes. However, traditional biomedical experimental methods are limited by low efficiency and high costs when verifying the association between circular RNA and drug sensitivity. Therefore, developing an efficient and accurate computational method to predict new associations between circRNA and drug sensitivity has become an urgent need in current research. To address this, this study proposes HECLCDA, a novel method based on heterogeneous cross-scale contrastive learning. To construct a comprehensive initial information base for drugs and circRNAs, circRNA gene sequence similarity, drug structural inclusion similarity (SIS), and Gaussian kernel similarity were integrated.Based on the integrated and complete known information of circRNAs and drugs, a heterogeneous graph was built. The model used the Heterogeneous Graph Transformer to extract heterogeneous network topological information, effectively distinguishing the heterogeneity of nodes and edges. The model broke through the information relationship between node attributes and network topology at two scales, and innovatively introduced a cross-scale contrastive learning mechanism in a sparse labeling scenario. Using self-supervised signals, we aimed to enhance the discriminative power of node embeddings and maximize the mutual information between paired nodes at different scales. Cross-validation experiments demonstrated that HECLCDA performs excellently on real data and can efficiently predict drug sensitivity. Additionally, case studies further validate the model's effectiveness in predicting potential circRNA-drug sensitivity associations.

The diverse applications of internet of things (IoT) have significantly increased the demand for efficient and reliable wireless networks, making power consumption a critical concern. Reconfigurable intelligent surface (RIS) have been proposed as a solution to mitigate power consumption in wireless communication systems by dynamically adjusting the signal propagation direction between transmitters and receivers. Due to the operational status of IoT devices and the complex association relationships between RISs and devices, a dynamic and highly variable communication environment is typically resulted, which renders power consumption optimization more challenging, as compared to conventional methods that do not incorporate RISs. This paper addresses the optimization of power consumption and IoT device coverage rate in an indoor communication scenario to improve system performance. We design an Adaptive Hybrid Optimization Strategy based on the association between RISs and devices to maximize the device coverage rate. Additionally, we optimize the phase shifts of multiple RISs to minimize system power consumption using the relaxation transformative method while satisfying the coverage rate constraint. Extensive simulation results demonstrate that, in an indoor environment with several obstacles, the proposed algorithm achieves a higher device-centric coverage rate compared to a solution without RIS and exhibits lower power consumption compared to strategies that rely more on base stations.

The widespread deployment of large language models (LLMs) has intensified concerns around intellectual property (IP) protection, as model theft and unauthorized redistribution become increasingly feasible.To address this, model fingerprinting aims to embed verifiable ownership traces into LLMs.However, existing methods face inherent trade-offs between stealthness, robustness, and generalizability-being either detectable via distributional shifts, vulnerable to adversarial modifications, or easily invalidated once the fingerprint is revealed.In this work, we introduce CTCC, a novel rule-driven fingerprinting framework that encodes contextual correlations across multiple dialogue turns-such as counterfactual-rather than relying on token-level or single-turn triggers.CTCC enables fingerprint verification under black-box access while mitigating false positives and fingerprint leakage, supporting continuous construction under a shared semantic rule even if partial triggers are exposed.Extensive experiments across multiple LLM architectures demonstrate that CTCC consistently achieves stronger stealth and robustness than prior work.Our findings position CTCC as a reliable and practical solution for ownership verification

This paper introduces a transformer-based multi-branch attention network (MBTFuse) as a solution to the challenges faced by existing image fusion methods. These methods often overlook the complementary and enhancing nature of intra-modal and inter-modal features, resulting in the loss of shared features and distortion of unique information in the fused results. To address these issues, the proposed MBTFuse network enables joint learning of inter-modal and intra-modal relationships in multi-modal image fusion. The generator network of MBTFuse comprises two information refinement branches and an attention-driven branch. The former two branches combine the convolutional neural network and the transformer architecture to extract semantic information within each modality. This integration ensures a comprehensive understanding of the modalities and their individual characteristics. The attention-driven branch compensates for specific information loss during global similarity analysis by extracting shared features between modalities. This step enhances the fusion process by preserving important shared information. After extensive experiments, results on multiple datasets show that MBTFuse produces competitive fusion results compared with existing fusion frameworks.

With the ever-growing attention on communication security, machine learning-based network intrusion detection system (NIDS) is widely utilized to meet different security requirements. However, most of the existing methods manually extract or learn features from raw traffic, which is usually expensive, complicated, and time-consuming. Moreover, this also brings unprecedented challenges for preserving users’ privacy in the communication process, making it difficult for existing solutions to be deployed in practice due to the privacy requirements from legal policies. This paper proposes a privacy-preserving graph neural network (named NIGNN) for NIDS, which can encode the local structure and traffic features. To address the privacy issues pertaining to the application of graph representation learning, we design a privacy message-passing mechanism with formal privacy guarantees, in which sensitive information potentially contained in graph vertices will be kept private. Specifically, we design a privacy-enhancement graph representation that introduces a degree-sensitive item in vertex-based aggregation to reduce noise. Our theoretical analysis shows that NIGNN can provide a provable privacy guarantee. Extensive experiments demonstrate NIGNN's performance in maintaining a sound privacy-accuracy trade-off.

Event extraction (EE) is an important natural language processing task. With the passage of time, many powerful and effective models for event extraction tasks have been developed. However, there has been limited research on complex overlapping event extraction. Therefore, we propose a new cascade decoding model: A Joint Learning Framework for Cascade Decoding with Multi-Feature Fusion and Conditional Enhancement for Overlapping Event Extraction. 1) In this model, we introduce a cascade decoding mechanism with multi-feature fusion to better capture the interaction between decoding layers. 2) Additionally, we introduce an enhanced conditional layer normalization (ECLN) mechanism to enhance the interaction between subtasks. Simultaneously, the use of a cascade decoding model effectively addresses the problem of overlapping events. The model successively performs three subtasks, type detection, trigger word extraction and argument extraction. All three subtasks learned together in a framework, and a new conditional normalization mechanism is used to capture dependencies among these subtasks. The experiments are conducted using the overlapping event benchmark, FewFC dataset. The experimental evaluation demonstrates that our model achieves a higher F1 score on the overlapping event extraction task compared to the original overlapping event extraction model.

The embedding of Knowledge Graphs (KGs) in hyperbolic space has recently received great attention in the field of deep learning because it can provide more accurate and concise representations of hierarchical structures compared to Euclidean spaces and complex spaces. Although hyperbolic space embeddings have shown significant improvements over Euclidean spaces and complex space embeddings in handling the task of KG embedding, they still face challenges related to the uneven distribution and insufficient alignment of high-dimensional sparse data. To address this issue, we propose the CONHyperKGE model, which leverages contrastive learning to optimize the embedding distribution in hyperbolic space. This approach enables better capture of hierarchical structures, improved handling of symmetry, and enhanced treatment of sparse matrices. Our proposed method is evaluated on four standard KG Embedding (KGE) datasets: WN18RR, FB15k-237, Kinship, and UMLS. After extensive experimental verification, our method has improved its performance on all four datasets. Notably, on the low-dimensional Kinship dataset, our method achieves an average Mean Reciprocal Rank (MRR) improvement of 2% over the original method, while on the high-dimensional WN18RR dataset, an average MRR improvement of 1% is observed compared to the original method.

In this paper, we study multimodal metaphor detection to obtain real semantic meaning from multiple heterogeneous information sources . The existing approaches mainly suffer from two drawbacks. (1) They focus on textual aspects, overlooking the characteristics of visual metaphor information. (2) Efficient methods for fusing multimodal metaphor features are lacking. To address the first issue, we propose a visual information enhancement method based on dual-granularity visual feature fusion , obtaining complete metaphorical visual features. To achieve bidirectional interaction among multimodal metaphor features, we further develop a multi-interactive crossmodal residual network (MCRN) that fuses the consistent and complementary information between different modalities and design a progressive fusion strategy to enhance the iterative fusion ability of the model. We extensively evaluate the proposed method on the popular Met-meme metaphor detection benchmark, outperforming the existing state-of-the-art methods by a large margins; i.e., we achieve F1 score improvements ranging from 1.47% to 2.55% under different languages. In addition, we further extend the evaluation to the Sarcasm dataset to validate the ability of the model to perceive semantic contrasts and meaning transformations, and the experimental results are superior to those of a strong baseline model .

Abstract Few-shot Event Detection (FSED) is a sub-task of Event Detection that aims to accurately identify event types with limited training instances and enable smooth transfer to newly-emerged event types. Recently, the dominant works have used the prototypical network to accomplish this task and employ contrastive learning to alleviate the issue of semantically-close categories. Nevertheless, these methods still suffer from two serious problems: (1) inadequate learning of prototype representations resulting from limited training data; (2) hard-easy sample imbalance and categories imbalance caused by the large number of non-trigger word("O" tags) in the token-level classification task. To address the problems, this paper proposes the Multi-channels Prototype and Contrastive learning method with Conditional Adversarial attack, which introduces the improved multi-channels prototype and contrastive networks to alleviate the categories and hard-easy samples imbalance. Moreover, we devise a constrained adversarial attack to improve the problem of limited training data. Extensive experimental results show that our model performs better than other FSED methods. All the code and data will be available for online public access.

With the ever-growing interest in modeling complex graph structures, graph neural networks (GNN) provide a generalized form of exploiting non-Euclidean space data. However, the global graph may be distributed across multiple data centers, which makes conventional graph-based models incapable of modeling a complete graph structure. This also brings an unprecedented challenge to user privacy protection in distributed graph learning. Due to privacy requirements of legal policies, existing graph-based solutions are difficult to deploy in practice. In this paper, we propose a privacy-preserving graph neural network based on local graph augmentation, named LGA-PGNN, which preserves user privacy by enforcing local differential privacy (LDP) noise into the decentralized local graphs held by different data holders. Moreover, we perform local neighborhood augmentation on low-degree vertices to enhance the expressiveness of the learned model. Specifically, we propose two graph privacy attacks, namely attribute inference attack and link stealing attack, which aim at compromising user privacy. The experimental results demonstrate that LGA-PGNN can effectively mitigate these two attacks and provably avoid potential privacy leakage while ensuring the utility of the learning model.

Recently, three-dimensional (3D) point-cloud analysis has been extensively utilized in the domain of machine vision, encompassing tasks include shape classification and segmentation. However the inherent disorder in point clouds poses a challenge in capturing relationships among points, particularly when dealing with mutilated and occluded data. To this end, We propose the Point Geometry Transformation (PointGT) method for 3D point-cloud classification and part segmentation, by exploring the underlying geometric structure in the local and global of points. Specifically, the efficacy of PointGT arises from the integration of a local abstraction (LA) module and an optimization strategy. The LA module is tailored to address the localized features inherent to point clouds. This module encapsulates the multidimensional attributes of local edge and inside points. The bi-directional cross-attention mechanism amalgamates these two constituents into the native channel with the primary objective of optimizing the exploitation of edge and inside delineations, thereby judiciously mitigating noise artifacts. Ultimately, the channel residual connections disseminate the postdownsampling point attributes, thereby inheriting the edge and inside delineations gleaned via post bi-directional attention. The effectiveness of the proposed method was verified through the validation of point-cloud classification and segmentation datasets. The empirical findings confirmed the efficacy of PointGT; accuracies of 93.2% and 87.8% were achieved for the ModelNet40 and ScanObjectNN datasets, respectively.

Extensive work has demonstrated the effectiveness of Vision Transformers. The plain Vision Transformer tends to obtain multi-scale features by selecting fixed layers, or the last layer of features aiming to achieve higher performance in dense prediction tasks. However, this selection is often based on manual operation. And different samples often exhibit different features at different layers (e.g., edge, structure, texture, detail, etc.). This requires us to seek a dynamic adaptive fusion method to filter different layer features. In this paper, unlike previous encoder and decoder work, we design a neck network for adaptive fusion and feature selection, called ViT-Controller. We validate the effectiveness of our method on different datasets and models and surpass previous state-of-the-art methods. Finally, our method can also be used as a plug-in module and inserted into different networks.

Graph Neural Networks (GNNs) as an emerging technique have shown excellent performance in a variety of fields, such as social networks and recommendation systems. However, GNNs may have to overcome privacy concerns as large amounts of information about their training datasets may be compromised. In this paper, we develop a privacy-preserving GNN to enforce privacy preservation, which utilizes a private Functional Mechanism (FM) to train the learning model. This mechanism perturbs the polynomial approximation of the objective function to enforce Differential Privacy (DP) in the GNN model. We show that our method can maximize the accuracy of the results with comparable prediction power to the unperturbed results while satisfying the privacy guarantees.

Studies have shown that IncRNA-miRNA interactions can affect cellular expression at the level of gene molecules through a variety of regulatory mechanisms and have important effects on the biological activities of living organisms. Several biomolecular network-based approaches have been proposed to accelerate the identification of lncRNA-miRNA interactions. However, most of the methods cannot fully utilize the structural and topological information of the lncRNA-miRNA interaction network. In this article, we proposed a new method, ISLMI, a prediction model based on information injection and second order graph convolution network(SOGCN). The model calculated the sequence similarity and Gaussian interaction profile kernel similarity between lncRNA and miRNA, fused them to enhance the intrinsic interaction between the nodes, using SOGCN to learn second-order representations of similarity matrix information. At the same time, multiple feature representations obtain using different graph embedding methods were also injected into the second-order graph representation. Finally, matrix complementation was used to increase the model accuracy. The model combined the advantages of different methods and achieved reliable performance in 5-fold cross-validation, significantly improved the performance of predicting lncRNA-miRNA interactions. In addition, our model successfully confirmed the superiority of ISLMI by comparing it with several other model algorithm.

In previous deep-learning-based methods, semantic segmentation has been regarded as a static or dynamic per-pixel classification task, i.e., classify each pixel representation to a specific category. However, these methods only focus on learning better pixel representations or classification kernels while ignoring the structural information of objects, which is critical to human decision-making mechanism. In this paper, we present a new paradigm for semantic segmentation, named structure-aware extraction. Specifically, it generates the segmentation results via the interactions between a set of learned structure tokens and the image feature, which aims to progressively extract the structural information of each category from the feature. Extensive experiments show that our StructToken outperforms the state-of-the-art on three widely-used benchmarks, including ADE20K, Cityscapes, and COCO-Stuff-10K.

The lightweight MLP-based decoder has become increasingly promising for semantic segmentation. However, the channel-wise MLP cannot expand the receptive fields, lacking the context modeling capacity, which is critical to semantic segmentation. In this paper, we propose a parametric-free patch rotate operation to reorganize the pixels spatially. It first divides the feature map into multiple groups and then rotates the patches within each group. Based on the proposed patch rotate operation, we design a novel segmentation network, named PRSeg, which includes an off-the-shelf backbone and a lightweight Patch Rotate MLP decoder containing multiple Dynamic Patch Rotate Blocks (DPR-Blocks). In each DPR-Block, the fully connected layer is performed following a Patch Rotate Module (PRM) to exchange spatial information between pixels. Specifically, in PRM, the feature map is first split into the reserved part and rotated part along the channel dimension according to the predicted probability of the Dynamic Channel Selection Module (DCSM), and our proposed patch rotate operation is only performed on the rotated part. Extensive experiments on ADE20K, Cityscapes and COCO-Stuff 10K datasets prove the effectiveness of our approach. We expect that our PRSeg can promote the development of MLP-based decoder in semantic segmentation.

As the demand for Internet of Things (IoT) technologies continues to grow, IoT devices have been viable targets for malware infections. Although deep learning-based malware detection has achieved great success, the detection models are usually trained based on the collected user records, thereby leading to significant privacy risks. One promising solution is to leverage federated learning (FL) to enable distributed on-device training without centralizing the private user records. However, it is non-trivial for IoT users to label these records, where the quality and the trustworthiness of data labeling are hard to guarantee. To address the above issues, this paper develops a semi-supervised federated IoT malware detection framework based on knowledge transfer technologies, named by FedMalDE. Specifically, FedMalDE explores the underlying correlation between labeled and unlabeled records to infer labels towards unlabeled samples by the knowledge transfer mechanism. Moreover, a specially designed subgraph aggregated capsule network (SACN) is used to efficiently capture varied malicious behaviors. The extensive experiments conducted on real-world data demonstrate the effectiveness of FedMalDE in detecting IoT malware and its sufficient privacy and robustness guarantee.

The advent of 5G brought new opportunities to leapfrog beyond current Industrial Internet of Things (IoT). However, the ever-growing IoT has also attracted adversaries to develop new malware attacks against various IoT applications. Although deep-learning-based methods are expected to combat the sophisticated malwares by exploring the latent attack patterns, such detection can be hardly supported by battery-powered end devices, such as Android-based smartphones. Edge computing enables the near-real-time analysis of IoT data by migrating artificial intelligence (AI)-enabled computation-intensive tasks from resource-constrained IoT devices to nearby edge servers. However, owing to varying channel conditions and the demanding latency requirements of malware detection, it is challenging to coordinate the computing task offloading among multiple users. By leveraging the computation capacity and the proximity benefits of edge computing, we propose a hierarchical security framework for IoT malware detection. Considering the complexity of the AI-enabled malware detection task, we provide a delay-aware computational offloading strategy with minimum delay. Specifically, we construct a coordinated representation learning model, named by Two-Stream Attention-Caps, to capture the latent behavioral patterns of evolving malware attacks. Experimental results show that our system consistently outperforms the state-of-the-art systems in detection performance on four benchmark datasets.

Long non-coding RNA(lncRNA) can interact with microRNA(miRNA) and play an important role in inhibiting or activating the expression of target genes and the occurrence and development of tumors. Accumulating studies focus on the prediction of miRNA-lncRNA interaction, and mostly are concerned with biological experiments and machine learning methods. These methods are found with long cycles, high costs, and requiring over much human intervention. In this paper, a data-driven hierarchical deep learning framework was proposed, which was composed of a capsule network, an independent recurrent neural network with attention mechanism and bi-directional long short-term memory network. This framework combines the advantages of different networks, uses multiple sequence-derived features of the original sequence and features of secondary structure to mine the dependency between features, and devotes to obtain better results. In the experiment, five-fold cross-validation was used to evaluate the performance of the model, and the zea mays data set was compared with the different model to obtain better classification effect. In addition, sorghum, brachypodium distachyon and bryophyte data sets were used to test the model, and the accuracy reached 0.9850, 0.9859 and 0.9777, respectively, which verified the model's good generalization ability.

Most current metaphor detection methods use restricted context, such as modeling the context of a single sentence. Considering the language environment of metaphors, we argue that combining broader discourse features has a greater impact on the improvement of metaphor detection performance. We propose a metaphor detection method based on fine-grained discourse, which embeds the current sentence and surrounding context in a weighted manner. With the help of fine-grained discourse, our model learns local and remote information as a reference for decision-making, and provides an efficient and natural method for metaphor detection tasks. Experimental results on VU Amsterdam Metaphor Corpus show that our technique surpasses the state-of-the-art models.

The traditional method of extracting entities and relations not only disregards the dependency between the two subtasks of entities and relations but also facilitates the cumulative propagation of errors. To solve these problems, a method – called MSBD – of joint extraction of entities and relations based on the Character Graph Convolutional Network (CGCN) and Multi-Head Self-Attention Mechanism (MS) is proposed. First, a new tagging scheme is used to tag the positions of entities and relations in the text. Second, to depict the hierarchical structure information between the entities and the relations in the text and the internal structure information of the entity, the CGCN is designed to obtain the character vector of the text. MS is introduced into the coding framework of Bidirectional Long Short-Term Memory (BiLSTM) to capture the relative position information between two entities in the text and represent the subspace information. The Dense Connected Convolutional Network (Dense Net) is embedded in the decoding framework to enhance the reuse and transmission of key information and achieve joint extraction of entities and relations in the text. The experimental results show that the P, R and F values are significantly improved for extracting entities and relations.

In the task of person re-identification, the attention mechanism and fine-grained information have been proved to be effective. However, it has been observed that models often focus on the extraction of features with strong discrimination, and neglect other valuable features. The extracted fine-grained information may include redundancies. In addition, current methods lack an effective scheme to remove background interference. Therefore, this paper proposes the feature refinement and filter network to solve the above problems from three aspects: first, by weakening the high response features, we aim to identify highly valuable features and extract the complete features of persons, thereby enhancing the robustness of the model; second, by positioning and intercepting the high response areas of persons, we eliminate the interference arising from background information and strengthen the response of the model to the complete features of persons; finally, valuable fine-grained features are selected using a multi-branch attention network for person re-identification to enhance the performance of the model. Our extensive experiments on the benchmark Market-1501, DukeMTMC-reID, CUHK03 and MSMT17 person re-identification datasets demonstrate that the performance of our method is comparable to that of state-of-the-art approaches.

Deep neural network models for Uyghur personal pronoun resolution learn semantic information for personal pronoun and antecedents, but tend to be short-sighted—they ignore the importance of each feature. In this article, we propose a Uyghur personal pronoun resolution model based on Attention mechanism, Convolutional neural networks and Gated recurrent unit (ATCG). Our model studies the grammatical structure and semantic features of Uyghur, and extracts 11 key features for Uyghur resolution task. Attention mechanism can focus on the importance of words in sentences. Gated Recurrent Unit (GRU) is applied in this model to achieve the interdependent features with long distance. The ATCG model effectively makes up for the shortcomings of relying only on the features of the content level and achieves better classification performance. Experimental results on Uyghur resolution dataset show that our model surpasses the state-of-the-art models.

In the traditional remote sensing image recognition, the traditional features (e.g., color features and texture features) cannot fully describe complex images, and the relationships between image pixels cannot be captured well. Using a single model or a traditional sequential joint model, it is easy to lose deep features during feature mining. This article proposes a new feature extraction method that uses the word embedding method from natural language processing to generate bidirectional real dense vectors to reflect the contextual relationships between the pixels. A bidirectional independent recurrent neural network (BiIndRNN) is combined with a convolutional neural network (CNN) to improve the sliced recurrent neural network (SRNN) algorithm model, which is then constructed in parallel with graph convolutional networks (GCNs) under an attention mechanism to fully exploit the deep features of images and to capture the semantic information of the context. This model is collectively named an improved SRNN and attention-treated GCN-based parallel (SAGP) model. Experiments conducted on Populus euphratica forests demonstrate that the proposed method outperforms traditional methods in terms of recognition accuracy. The validation done on public data set also proved it.

To solve the problems from using single-layer features in traditional water body identification models, such as the lack of local descriptors, large quantization errors, and the lack of semantic information descriptions, a multimodal model is proposed based on the different levels of feature knowledge. First, based on the multidescriptor hierarchical feature, the middle-level local feature extraction of the water body is achieved, and, combined with the convolutional neural network, the high-order global features of the water body are extracted. Then, the image features are hierarchically normalized, and multimodal RBM self-encoding is used for fusion to reduce the quantization error of each layer feature in the encoding process. Finally, the generative model of the Multimodal Model is used to expand the data and filter the multilayered features after fusion. In addition, the semantic information of a water body is further discovered by using the encoder and decoder of the discriminant model and is classified by employing SoftMax. The results show that compared with the traditional water body identification methods, the proposed method improves the recognition accuracy and image description ability.

To improve the accuracy and automation of malware Uniform Resource Locator (URL) recognition, a joint approach of Convolutional neural network (CNN) and Long-short term memory (LSTM) based on the Attention mechanism (JCLA) is proposed to identify and detect malicious URL. Firstly, the URL features including texture information, lexical information and host information are extracted and filtered, and pre-processed with encode. Then, the feature matrix more relevant to the output are chose according to the weight of the attention mechanism and input to the constructed parallel processing model called CNN_LSTM, combinating CNN and LSTM to get local features. Next, the extracted local features are merged to calculate the global features of the URLs to be detected. Finally, the URLs are classified by the SoftMax classifier using global features, the accuracy of the model in malicious URL recgonition is 98.26%. The experimental results show that the JCLA model proposed in this paper is better than the traditional deep learning model or CNN_LSTM combined model for detecting malicious URLs.

Malaria is a kind of disease that greatly threatens human health. Nearly half of the world’s population is at risk of malaria. Anti-malarial drugs which are sought, developed and synthesized keep malaria under control, having received increasing attention in drug discovery field. Machine learning techniques have been used widely in drug research and development. On the basis of semi-supervised machine learning for molecular descriptions, this research develops a multilayer deep belief network (DBN) that can be used to identify whether compounds have the anti-malarial activity. Firstly, the influence of feature dimensions on predicting accuracy is discussed. Furthermore, the proposed model is applied to contrast shallow machine learning and supervised machine learning with the similar deep architecture. The research results show that the proposed model can predict anti-malarial activity accurately. The stable performance on the evaluation metrics confirms the practicability of our model. The proposed DBN model performs better than other shallow supervised models and deep supervised models. Moreover, it could be applied to reduce the cost and the time of drug discovery.

As a core subtask in anaphora resolution, anaphoricity determination has aroused the interest of researchers. However, in recent work, the influence caused by the deep semantic information and the context of the coreference elements have not been taken into account. In this paper, by combining the semantic feature of Uygur, we established a Convolutional Neural Network & Long Short-Term Memory (CNN_LSTM) model in determining the anaphoricity of Uygur pronoun. Firstly, the deep negative semantic feature representation is extracted via word2vec. Secondly, the shallow explicit feature representation of coreference elements is extracted by our system. Afterwards, two kinds of features are combined to recognize whether coreference element is referential or not. The results showed that the method we used can distinguish coreference element accurately, the ACC[Formula: see text] score is 90.18% and the ACC[Formula: see text] score is 89.93%, which are higher than ANN (Artificial Neural Network) and SVM (Support Vector Machine) respectively.

The cytochrome P450 (CYP) superfamily, exists in the human liver, is responsible for more than 90% of the metabolism of clinical drugs. So it is necessary to adopt a new kind of computer simulation methods that can predict the rejection capability of compounds for a concrete CYPs isoform. In this work, a model is presented for classification of CYP450 1A2 inhibitors and noninhibitors based on a multi-tiered deep belief network (DBN) on a large dataset. The dataset composed of more than 13,000 heterogeneous compounds was acquired from PubChem. Firstly, 139 2D and 53 3D descriptors are calculated and preprocessed. Then, the unsupervised learning method is used to train DBN model to automatically extract multiple levels of distributed representation from the descriptors of training set. Finally, by using testing set and external validation set, we evaluate the classified performance of DBN for the inhibition of CYP1A2. Meanwhile, the proposed model is compared with shallow machine learning models (support vector machine (SVM) and artificial neural network (ANN)). We also discussed the performance of DBN by comparing it with different features combination. The experimental results showed that DBN has a better prediction ability compared with SVM and ANN. And these models combined with the features of 2D and 3D obtain the best forecast accuracy.

Traditional machine learning methods for water body extraction need complex spectral analysis and feature selection which rely on wealth of prior knowledge. They are time-consuming and hard to satisfy our request for accuracy, automation level and a wide range of application. We present a novel deep learning framework for water body extraction from Landsat imagery considering both its spectral and spatial information. The framework is a hybrid of convolutional neural networks (CNN) and logistic regression (LR) classifier. CNN, one of the deep learning methods, has acquired great achievements on various visual-related tasks. CNN can hierarchically extract deep features from raw images directly, and distill the spectral–spatial regularities of input data, thus improving the classification performance. Experimental results based on three Landsat imagery datasets show that our proposed model achieves better performance than support vector machine (SVM) and artificial neural network (ANN).