Zero-Shot Action Recognition aims to transfer visual-semantic consistency knowledge learned from seen actions to the recognition of unseen actions. However, existing methods predominantly focus on learning the visual-semantic consistency in seen domain, neglecting that distribution discrepancies between seen and unseen domains lead to biased and non-causal feature learning. To address this issue, we propose an Unbiased Spatial-Temporal Atomic Fusion-based Zero-Shot Action Recognition method (USAF) for alleviating the influence of biased data on cross-modal mapping generalization. Specifically, we conceptualize the inference mechanism of ZSAR as a directed acyclic graph-based causal system to diagnose the data bias of cross domains induced by spatial and temporal confounders, and design the model following its causal feature extraction pathway. We first decomposes video representation space into motion and object atomic spaces, allowing for the extraction of fine-grained causal visual features that are robust against spatial confounders. Furthermore, we propose a Cross-modal Fusion & Matching Mechanism that dynamically fuses motion patterns and object cues from atomic spaces, and performing cross-modal matching against label semantic, thereby capturing temporal saliency distribution discrepancies between the atomic spaces. Notably, to extract local spatiotemporal dependencies for constructing the motion atomic space, we propose a novel Short-term Spatial-Temporal Graph Convolutional Network. Extensive experiments demonstrate that USAF achieves state-of-the-art performance on multiple public action recognition benchmarks, significantly enhancing cross-modal mapping generalization from seen to unseen domains.

Video anomaly detection presents a significant challenge in computer vision, with the aim of distinguishing various anomaly events from numerous normal ones. Weakly supervised video anomaly detection has recently emerged as a promising solution, enabling the detection of anomaly snippets with only video-level annotations. However, knowledge about anomaly annotation remains underutilized, resulting in a gap between visual space and semantic understanding of anomalies, thus failing to capture the clear boundary between anomalies and normalities. Therefore, we propose a weakly supervised paradigm of cross-modal detection and consistency learning, leveraging dual consistency to provide discriminative representations for anomalies at both the semantic-to-target and target-to-snippet levels. Specifically, we introduce a cross-modal detection network, which detects the targets in each frame according to given semantic rules, to derive semantic-consistent visual embeddings. To depict the clear boundary between anomalies and normalities, a cross-domain alignment module is proposed to enhance the discriminative representation of abnormal targets by learning the contextual consistency between the target and snippet embeddings. Our architecture integrates the detection of semantic-consistent targets based on variable semantic rules, ensuring transferable deployment across scenarios and enabling comprehensive identification, localization, and recognition of abnormal events through a “when-where-which” pipeline. The evaluation of our approach is conducted on four widely used public benchmarks: ShanghaiTech, UCSD Ped2, CUHK Avenue, and UBnormal through extensive qualitative and quantitative analyzes. The results demonstrate the remarkable performance of our approach in dealing with the VAD task.

Evidence is accumulating that short-chain fatty acids (SCFAs) produced by the gut microbiota play pivotal roles in host metabolism. They contribute to the metabolic regulation and energy homeostasis of the host not only by preserving intestinal health and serving as energy substrates but also by entering the systemic circulation as signaling molecules, affecting the gut-brain axis and neuroendocrine-immune network. This review critically summarizes the current knowledge regarding the effects of SCFAs in the fine-tuning of the pathogenesis of type 2 diabetes mellitus (T2DM) and insulin resistance, with an emphasis on the complex relationships among diet, microbiota-derived metabolites, T2DM inflammation, glucose metabolism, and the underlying mechanisms involved. We hold an optimistic view that elucidating how diet can influence gut bacterial composition and activity, SCFA production, and metabolic functions in the host will advance our understanding of the mutual interactions of the intestinal microbiota with other metabolically active organs, and may pave the way for harnessing these pathways to develop novel personalized therapeutics for glucometabolic disorders.

In medical image recognition, challenges such as data scarcity, interindividual variability, and noise often lead to shortcut learning, where models fixate on visual cues irrelevant to the underlying pathology. Human gaze data naturally highlights critical regions and, when combined with the expert knowledge of clinicians, markedly improves diagnostic accuracy. However, the inability to precisely emulate expert knowledge creates a significant gap between human and model cognition, limiting the accuracy of deep learning models that rely solely on sparse gaze data. To address this, we introduce the Gaze-and-Machine Dualdriven Attention Fusion Network (GMD-AFNet), designed to enhance feature extraction in medical image classification. GMD-AFNet integrates raw image data, human eye-tracking information, and model-generated attention within a multi-branch architecture. By leveraging diversity loss, it encourages each branch to capture complementary and unique features, avoiding redundant representations. Contrastive loss enhances feature discriminability and generalization by pulling together features of similar samples and pushing apart those of dissimilar ones, thereby improving classification performance. We evaluate GMD-AFNet on the CXR-Eye dataset, which includes eye-tracking data from radiologists diagnosing chest X-rays, providing a realistic depiction of expert attention patterns. Experimental results demonstrate that GMD-AFNet effectively fuses human and machine attention, achieving substantial improvements in classification accuracy over baseline models.

Human pose estimation in low-light conditions is vital for applications such as surveillance and autonomous systems, yet the severe visual distortions hinder both manual annotation and estimation precision. Existing approaches typically rely on additional reference information to mitigate these issues, however, customized data collection equipment poses limitations on their scalability. To alleviate the issue, we construct a Low-Light Images and Poses (LLIP) dataset, which includes only paired low-light images and pose annotations obtained using off-the-shelf motion capture devices. Furthermore, we propose a Multi-grained High-frequency Feature Consistency Learning framework (MHFCL), which does not rely on additional reference information. MHFCL employs a Retinex-inspired restoration stream to recover high-frequency details and integrates them into pose estimation using a multi-grained consistency mechanism. Experiments demonstrate that our approach achieves a new benchmark in low-light pose estimation, while maintaining competitive performance in well-lit conditions.

For black-box attacks, most existing attack methods exhibit weak transferability due to the significant discrepancy between substitute model and victim model. We argue that the model-specific discriminative regions are a key factor causing overfitting to the source model. However, existing model augmentation methods focus on augmentations within a single domain, thereby restricting the diversity of the simulated models. In this paper, we present a novel model augmentation method named CSFD, which combines our proposed channel-wise scaling(CS) and frequency-random dropping(FD) to enhance the diversity of simulated model. Specifically, we first scale the image by channel with CS, which augments them in spatial domain. Then we randomly remove specific frequency patterns of the image with our FD, further augmenting the image in frequency domain by introducing a loss-preserving transformation. This enables us to fully exploit the properties of image in different domains and largely increases the diversity of simulated models. Additionally, we inject random noise perturbations into the sample, effectively exploring the decision boundary within an extended data distribution space. Extensive experiments on ImageNet dataset show that the proposed method has better performance than the existing methods.

Weakly-supervised video anomaly detection aims to develop a robust detector that distinguishes anomalies across diverse scenes using only video-level annotations. However, current centralized approaches are impractical for real-world surveillance applications due to stringent privacy and security regulations. Moreover, existing models rely on task-agnostic feature extractors, making them susceptible to domain shifts caused by domain discrepancies and limiting their detection capabilities to a coarse-grained level. In this work, we propose a privacy-preserving federated weakly-supervised architecture by deploying scenario-specific clients for context-aware detection without sharing local data. To address domain shift issue, we introduce a gated mixture of local-to-global experts model, where local experts adaptively learn source-aligned representations from video tubes, and global experts dynamically integrate local and global representations for enhanced robustness. In addition, a tube attention mechanism is designed to enable fine-grained anomaly localization through attention scores. Edge computing devices are employed as local clients for the training and testing of our method. Comprehensive experiments demonstrate that our model substantially outperforms state-of-the-art methods on the ShanghaiTech and UBnormal benchmark datasets, achieving an AUC of 83.06% on UBnormal dataset.

Weakly Supervised Video Anomaly Detection (WSVAD) plays a crucial role in visual surveillance by effectively distinguishing anomalies from normality with only video-level annotations; nevertheless, due to inherent limitations, including imbalanced data, intra-bag similarity, and snippet entanglement, existing methodologies prone to a bias towards normality. This paper proposes a novel framework, Visual Prompt Exemplars-based WSVAD termed VPE-WSVAD, to enhance discriminative representations of anomalies against normality by incorporating scenario-awareness visual prompt exemplars. The proposed VPE-WSVAD framework comprises three key components. Firstly, the prompt retrieval network generates potential abnormal proposals consistent with visual prompt exemplars at the frame level. Secondly, the proposal filtering mechanism aims to select the proposal with the highest confidence score to represent potential anomalies in the snippet. Finally, prompt likelihood learning is designed to capture the correlations between proposals (w.r.t prompt exemplars) and snippets, generating discriminative representations for each snippet. By incorporating visual prompt exemplars, our method provides more detailed reporting of abnormal events, including when, where, and what. To validate the efficacy of our method, we conducted comprehensive evaluations on three publicly available datasets. The experimental results conclusively demonstrate the superiority of our approach, achieving an impressive frame-level area under the curve (AUC) of 96.88% on the ShanghaiTech dataset and 99.86% on the UCSD Ped2 dataset. Our code will be released in the future.

This article analyzed the mechanism of Huangqi Simiao Decoction(HSD)for the treatment of type 2 diabetes mellitus(T2DM).The component targets of HSD and the related disease targets of T2DM were screened through network pharmacology.The protein-protein interaction(PPI)network of intersecting targets and the drug-component-intersecting target network were constructed to screen the potential active ingredients and targets.Molecular docking was performed using AutoDock Vina software to verify the interaction between potential components and core targets.The serum was tested by ultra performance liquid chromatography-tandem mass spectrometry,and multivariate statistical analyses,such as principal component analysis(PCA)and partial least squares discriminant analysis(PLS-DA),were used to search for the differential metabolites and related metabolic pathways of each group by combining with the MetaboAnalyst database.The same metabolic pathways were analyzed by combining the screened differential metabolites with the intersecting targets screened by network pharmacology.Network pharmacology showed that the nine core components of HSD for the treatment of T2DM were quercetin,kaempferol,stigmasterol,baicalein,β-sitosterol,flavodoxin,canthaxanthin,canthaxanthin,berberine,and berberine,and the five core targets included AKT1,TP53,TNF,IL6,and VEGFA.Molecular docking showed that the core components bound well to the target genes.Metabolomics showed that a total of 112 common differential metabolites were identified,of which 88 metabolites exhibited increased concentration and 24 metabolites decreased concentration after treatment with HSD.Enrichment analysis showed that HSD regulated the body metabolism of patients with T2DM,mainly related to seven metabolic pathways,such as amino acid metabolism and tricarboxylic acid cycle.The joint analysis of metabolomics and network pharmacology showed that both involved histidine metabolism,arginine and proline metabolic pathways.This study suggests that HSD has a good efficacy for T2DM.Based on the combined analysis of metabolomics and network pharmacology,it was found that the mechanism may be that the pharmacodynamic bases of quercetin,kaempferol,and stigmasterol in HSD enhance the effects on histidine metabolism,arginine and proline metabolic pathways by modulating a variety of metabolites,which provides the basis for further prevention and treatment of T2DM.

The recommendation system recommends service providers on platforms according to their trust. However, the diversity of platforms and malicious attacks can affect the trust of service providers and lead to reliable services unavailable for users and partners. To address the issue, in this paper, we introduce the blockchain into service recommendation and propose a blockchain-based service recommendation and trust enhancement model. First, we propose platform heterogeneous factors to improve the accuracy of service recommendations. Second, we deploy the smart contracts on the blockchain, which implements trust aggregation that considers multiple recommendations for the same service provider and ultimately generates reliable recommendation results. Finally, we propose a trust enhancement algorithm in the smart contract, which can solve the trust unreliability problem caused by malicious attacks. Extensive experiments on case studies and real datasets. In comparison, our method outperforms the baseline methods by achieving a 22% reduction in RMSE and an 8% increase in coverage, thereby demonstrating the effectiveness and reliability of our method.

Background Knowledge regarding the health impacts of daily eating frequency (DEF) and nighttime fasting duration (NFD) on mortality is very limited. Objective This study aimed to examine whether DEF and NFD are associated with CVD and all-cause mortality. Methods This was a prospective cohort study of a nationally representative sample from the United States, including 30,464 adults who participated in the National Health and Nutrition Examination Survey 2003–2014. Using 24-h dietary recall, DEF was assessed by the number of eating episodes, and NFD was calculated by the first and last eating time across a day. Death information was obtained from the National Death Index up to 2019. Weighted Cox proportional hazards regression models were used to assess survival relationships of DEF and NFD with mortality. Results During 307,686 person-years of follow-up, 4560 deaths occurred, including 1824 CVD cases. After adjustment for confounders, compared to DEF at 4–6 times, participants whose DEF was less than 3 times had greater CVD [hazard-ratio (HR) = 1.33, 95% confidence-interval (CI): 1.06–1.67] and all-cause (HR = 1.16, 95% CI: 1.01–1.33) mortality risks. Furthermore, compared to NFD of 10 to 11 h, participants whose NFD was shorter than 10 h had HRs of 1.30 (95% CI: 1.08–1.55) for CVD mortality and 1.23 (95% CI: 1.08–1.39) for all-cause mortality. NFD longer than 14 h was also related to CVD mortality (HR = 1.37, 95% CI: 1.12–1.67) and all-cause mortality (HR = 1.36, 95% CI: 1.19–1.54). Similar results for the association of NFD and DEF with heart-specific and stroke-specific mortality were observed. Conclusion This study found that DEF less than 3 times and NFD shorter than 10 h or longer than 14 h were independently associated with greater cardiovascular and all-cause mortality.

Weakly-supervised video anomaly detection is typically formulated as a multi-instance learning problem, assigning anomaly scores to each video snippet by learning to rank with only video-level labels. However, previous approaches that rely on snippet-level embeddings generated by task-agnostic feature extractors inevitably encounter challenges such as intra-bag similarities and frame-level entanglement. As a result, the model may exhibit significant performance degradation, particularly when the target is looming or receding. To address these issues, we present a novel weakly-supervised hierarchical position-scale awareness model that incorporates heterogeneous cross-scale correlation learning to improve detection performance. Specifically, in addition to snippet-level embedding, we employ an object detector (e.g., YOLOv5) for frame-level target detection and perform normal target clustering. By introducing a hierarchical ranking strategy, we gradually disentangle potential anomaly targets from the frame to the snippet level. Subsequently, we design a simple yet efficient position-scale awareness inference method that predicts the spatial positions and scales of looming and receding targets based on the abnormal targets with high confidence in adjacent snippets. Furthermore, we introduce heterogeneous cross-scale correlation learning to acquire the correlation between targets and snippet embeddings, enabling our model to increase attention to anomaly-related targets. Compared to previous approaches that generate only anomaly scores for each snippet, our method can locate anomalous targets, making it more suitable for practical applications. Without bells and whistles, evaluations on commonly-used VAD benchmarks: ShanghaiTech, UCSD-Ped2, Avenue, UCF-Crime and UBnormal datasets show that our method yields competitive and highly promising results compared with existing unsupervised, self-supervised, and weakly-supervised competitors. The code will be made publicly available.

Human motion prediction (HMP) aims to predict future human motions from historical pose sequences. Extensive efforts have adopted the Transformers or Graph Neural Networks (GNNs) to capture the spatio-temporal relationships between poses and thus incorporate the contextual information and complex behavior relationships for motion inference. However, most existing approaches treat the HMP task as a deterministic problem, thus resulting in poor diversity and long tail problems. This study attributes such issues to positional bias within the Transformers and the lack of degrees of freedom within the predictive model. Hence, we propose a novel Multi-degree Tail-aware Attention Network (MTAN) involving a tail-aware attention mechanism and a multi-degree feature representation strategy. Specifically, we introduce a tail-aware attention mechanism to adeptly capture spatio-temporal dependencies that accommodate both head and tail actions. Based on CVAE, the multi-degree feature representation strategy learns to capture temporal diversity by learning the joint distribution of observed and future sequences. Ultimately, we leverage GCN to model spatial dependencies effectively, culminating in a comprehensive spatiotemporal prediction model. We evaluate the effectiveness of our approach using three benchmark datasets, including Human3.6M, AMASS, and 3DPW. The results demonstrate that our approach surpasses state-of-the-art transformer methods, establishing its superiority in HMP.

Numerous studies have shown that well-designed perturbations can easily fool deep neural networks. Existing attacks are mainly conducted on the low-level pixels of RGB images, resulting in noise-like perturbations distributed over the entire image, highly vulnerable and low attack transferability. Furthermore, they delve into the data space with point-wise perturbation, which may neglect the geometric characteristics and fail to study the role and impact of various image components. Compared with RGB images, $$YC_bC_r$$ images can express various image components more intuitively. In this paper, we propose generating semantically preserved adversarial examples by perturbing the frequency band energy corresponding to inconspicuous colors and textures in the $$YC_bC_r$$ color space. Specifically, we first transform clean images from spatial to frequency domain, followed by applying a fusion module to indirectly inject perturbations. Moreover, the low-frequency constraint and luma-chroma optimization strategy are further introduced to ensure visual imperceptibility. Extensive experiments on multiple datasets indicate that our attack retains a high attack success rate while significantly improving visual quality.

In a service ecosystem, the trust of users in services serves as the foundation for maintaining normal interactions among users, service providers, and platforms. However, malicious attacks can tamper with the trust value of these services, making it difficult for users to identify reliable services and undermining the benefits of reliable service providers and platforms. When existing trust management models address the impact of malicious attacks on service reliability, they rarely consider leveraging different attack targets to improve the accuracy of compromised service trust. Therefore, we propose a trust enhancement model based on distributed learning and blockchain in the service ecosystem, which adaptively enhances the trust values of compromised services according to the targets of anomalous attacks. Firstly, we conduct a comprehensive analysis of the targets of malicious attacks using distributed learning. Secondly, we introduced a trust enhancement contract that utilizes different methods to enhance the trust of the service based on various attack targets. Finally, our approach outperforms the baseline method significantly. For different attack targets, we observe a reduction in RMSE by 12.38% and 12.12%, respectively, and an enhancement in coverage by 24.94% and 14.56%, respectively. The experimental results show the reliability and efficacy of our proposed model.

Trust prediction facilitates the day-to-day functionality of diverse web-based applications, such as recommendation systems, market advertising and anomaly detection. However, existing works heavily rely on user-user trust interactions, which result in limited performance as the data sparsity. Previous studies have shown that the trust relationship between users is significantly affected by the category of items that the users interacted. In this paper, we propose a MetaTrust model, which generates redundant user-item interactions as the supplement of user-user trust to alleviate the data sparsity on trust prediction. Specifically, we propose category-aware metapaths, which generate abundant user-item-user interactions based on the common item category that users have interacted with. Further, Long Short Term Memory (LSTM) networks are utilized to mine features of multiple category-aware metapaths and their correlations. In order to filter the user-item-user interactions that are not related to the current task, the real trust relationship between users are embedd in the network with MLP. Finally, a multi-headed attention network is utilized to distinguish which metapath determines trust prediction between the current pair of users. Extensive experiments on three real-world dataset show that our proposed model can effectively achieve significant improvements over other competitive approaches and show the potential interpretability of trust building.

Recently, Action Recognition (AR) is facing the scalability problem, since collecting and annotating data for the ever-growing action categories is exhausting and inappropriate. As an alternative to AR, Zero-Shot Action Recognition (ZSAR) is getting more and more attention in the community, as they could utilize a shared semantic/attribute space to recognize novel categories without annotated data. Different from the AR focuses on learning the correlation between actions, ZSAR needs to consider the correlation of action-action, label-label and action-label at the same time. However, as far as we know, there is no work to provide structural guidance for the framework design of ZSAR according to its task characteristics. In this paper, we demonstrate the rationality of using the Energy-Based Model (EBM) to guide the framework design of ZSAR based on their inference mechanism. Furthermore, under the guidance of EBM, we propose an Energy-based Temporal Summarized Attentive Network (ETSAN) to achieve ZSAR. Specifically, to ensure the effectiveness of cross-modal matching, EBM needs to capture the correlations of input-input, output-output and input-output, based on discriminative and focused input and output space. To this end, we first design the Temporal Summarized Attentive Mechanism (TSAM) to capture the correlation of action-action by constructing discriminative and focused input space. Then, a Label Semantic Adaptive Mechanism (LSAM) is proposed to learn the correlation of label-label by adjusting the semantic structure according to the target task. Finally, we devise an Energy Score Estimation Mechanism (ESEM) to measure the compatibility (i.e. energy score) between video representation and label semantic embedding. With end-to-end training, our framework can capture all three of the correlations mentioned above simultaneously by minimizing the energy score of the correct action-label pair. Experiments on the HMDB51 and UCF101 datasets show that the proposed architecture achieves comparable results among methods based on the spatial-temporal visual feature of sequence-level, which demonstrates the efficiency of the EBM in guiding the framework design of ZSAR.

A diabetic ulcer (DU) is a dreaded and resistant complication of diabetes mellitus with high morbidity. Fu-Huang ointment (FH ointment) is a proven recipe for treating chronic refractory wounds; however, its molecular mechanisms of action are unclear. In this study, we identified 154 bioactive ingredients and their 1127 target genes in FH ointment through the public database. The intersection of these target genes with 151 disease-related targets in DUs resulted in 64 overlapping genes. Overlapping genes were identified in the PPI network and enrichment analyses. The PPI network identified 12 core target genes, whereas Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis indicated that upregulation of the PI3K/Akt signalling pathway was involved in the role of FH ointment in treating diabetic wounds. Molecular docking showed that 22 active compounds in FH ointment could enter the active pocket of PIK3CA. Molecular dynamics was used to prove the binding stability of the active ingredients and protein targets. We found that PIK3CA/Isobutyryl shikonin and PIK3CA/Isovaleryl shikonin combinations had strong binding energies. An in vivo experiment was conducted on PIK3CA, which was the most significant gene.This study comprehensively elucidated the active compounds, potential targets, and molecular mechanism of FH ointment application in treating DUs, and believed that PIK3CA is a promising target for accelerated healing.

Introduction: Oral Banzhilian formula (BZLF) is effective in the clinical treatment of psoriasis. However, the effectiveness and mechanism of different drug delivery routes deserve further study.Methods: First, we established the mouse model of psoriasis using imiquimod (IMQ), and high-performance liquid chromatography (HPLC) was used for the quality control of BZLF. Secondly, Total RNA Sequencing and bioinformatics analysis were used to explore the regulatory mechanism of BZLF in improving psoriatic lesions. Finally, further verification was based on animal experiments.Results: we externally applied BZLF for skin lesions in an imiquimod-induced psoriasis mouse model and found that BZLF alleviated psoriasis-like skin lesions while inhibiting the expression of Ki67 and inflammatory factors (Il17a, Tnf-α, S100a7 and Cxcl1) in skin lesions. Transcriptome sequencing results suggested that BZLF inhibited signalling pathways closely related to psoriatic inflammation, such as the IL-17 signalling pathway, chemokine signalling pathway, TNF signalling pathway, and NF-kappa B signalling pathway, and the protein-protein interaction (PPI) network identified LCN2 as one of the core target genes and screened out its regulated downstream gene MMP9.Discussion: Our findings suggest that the anti-psoriatic mechanism of BZLF involved in downregulating the LCN2/MMP-9 axis.