Weakly supervised point cloud semantic segmentation (WSPCSS) has gained attention for reducing reliance on densely annotated data. However, inefficiencies in utilizing sparse annotations hinder comprehensive understanding of complex scenes. Inspired by masked autoencoder (MAE) techniques in image processing, researchers have adapted these methods to WSPCSS. Yet, current 3D masking strategies often fail to capture intricate geometric properties, resulting in generated to-be-filled content that inaccurately represents the underlying 3D scene structure. To address these limitations, this study proposes a novel stereo-wise masking strategy, which extends 2D plane masking into 3D space to generate coherent and semantically rich masked regions with contextual relevance. These regions serve as high-quality learning targets, enabling the model to better comprehend complex point cloud structures. Experimental results demonstrate that, at a 0.01 % annotation density, the proposed method achieves improvements in mIoU by 1.9 % and 4.72 % on the indoor datasets S3DIS and ScanNet V2, respectively, compared to previous methods. Furthermore, at a 0.1% annotation density on the forest dataset For-Instance, the method exhibits a 0.29 % improvement. These results substantiate the effectiveness and stability of the stereo-wise masking strategy.
With the rapid deployment of edge devices, there has been growing concern over the privacy preservation of clients’ private data. Federated learning (FL) has emerged as a promising solution that enables edge devices for collaborative model training without compromising data privacy, as the local datasets of clients are not shared directly. However, previous works have shown that the performance of conventional FL approaches significantly deteriorates when the data distributions across clients are heterogeneous or non-IID. Integrating personalization into FL addresses these challenges by adapting the global model to local data distributions, but existing personalized FL (PFL) frameworks often fail to generalize when clients’ datasets have cross-domain features. One approach to mitigate this challenge is partial model initialization, where to update clients’ local model only a subset of model parameters are initialized from the global model while retaining local parameters. However, obtaining effective generalization across distinct data distributions remains a challenge. In such instances, meta-learning is preferred as it provides the adaptation of a global model across distinct data distributions and tasks by employing only a small number of data samples. In this paper, we propose MetaPartialFL, a novel framework that integrates meta-learning with partial model initialization to address cross-domain personalization and heterogeneity in FL. The approach partitions model parameters into shared global and client-specific local components, enabling effective generalization across clients while maintaining adaptability to each client’s datasets. Extensive evaluations on cross-domain benchmark datasets, such as Office-Home, Office-31, PACS, and DomainNet, demonstrate that MetaPartialFL outperforms state-of-the-art methods, making it a robust and efficient solution for cross-domain FL applications.
This research investigates the application of Large Language Models (LLMs) for automating the assessment of bicycle insurance claims. The study aims to determine the effectiveness of LLMs in interpreting insurance policy documents and evaluating the validity of claims. The ultimate objective being to enhance consistency and efficiency in claims processing by leveraging LLM capabilities to understand complex legal language. The methodology involves training and testing various models on a bicycle insurance policy and hypothetical claims against the policy. Key achievements include the introduction of an agent-based approach that significantly improves the acceptability of written claims handler responses. The findings indicate that, by using a combination of agents with crafted prompts, LLMs can produce acceptable claims handler responses to claim descriptions.
Short-term traffic speed prediction under limited historical data is crucial for intelligent transportation systems, enabling real-time traffic management and congestion mitigation. However, existing methods relying on long lookback window to forecast struggle to capture scenarios where the current traffic state depends more on recent patterns than on distant historical data. To address this challenge, we propose the Spatiotemporal Visibility Graph Network (STVGN), a novel framework that combines visibility graph theory with Graph Neural Networks (GNN). The core of STVGN is the STVGN-Embedding layer, which is designed to enhance short-term spatiotemporal feature representation. This layer leverages the visibility graph's ability to capture transient patterns and integrates GNNs to model intrinsic relationships within the visibility graph derived from time series data. In Combination with a transformer architecture, STVGN-Embedding extracts complex spatiotemporal features, while the transformer uncovers inherent relationships between temporal and spatial dimensions. To the best of our knowledge, this is the first study to introduce visibility graph theory as an embedding layer within a transformer framework. Experiments on three benchmark datasets, covering urban and freeway traffic scenarios, demonstrate STVGN's effectiveness, achieving state-of-the-art performance with improvements of 4.8%-16.8% in RMSE and 8.4%-13.3% in MAE over existing SOTA methods. These results highlight the potential of visibility graph-based embeddings to address challenges posed by limited historical data and to capture intricate traffic patterns.
Industrial Anomaly Detection (AD) encounters a significant cold-start challenge due to the requirement of a large number of labeled normal samples, which are often difficult to obtain in new production lines. Although Zero-Shot Anomaly Detection (ZSAD) and Few-Shot Anomaly Detection (FSAD) have been proposed as potential solutions, existing methods suffer from limitations in generalization ability and are prone to contamination by anomalies in few-shot scenarios. To address these issues, we propose DiCLIP, a unified framework that integrates ZSAD and FSAD through three key innovations. First, Versatile Combination Prompts Learning combines static, dynamic, and anomaly-sensitive prompts to leverage textual anomaly cues together with image features for accurate anomaly localization in images. Second, the Anomaly-Aware Memory Bank utilizes ZSAD priors to filter contaminated features, enabling anomaly detection based on a small number of anomaly samples. Third, Adaptive Threshold Optimization integrates semantic alignment from ZSAD with feature matching from FSAD to release the constraint of a uniform threshold for test images, thereby achieving higher-precision segmentation and localization performance. Extensive experiments on the standard MVTec and VisA benchmark datasets demonstrate the superior performance of DiCLIP, highlighting its effectiveness and practical value for industrial deployment.
With the rapid development of generative models, it has become increasingly challenging for humans to distinguish between AI-generated and real images, which may lead to the rampant spread of AI-generated misinformation. Therefore, detecting AI-generated images is a critical task to address this issue. In this work, we propose a two-stage method for Real/Fake detection and local detection of manipulated images. For Real/Fake detection, we are the first to propose rearranging the spatial features of large-scale pre-trained multimodal model encoders to achieve this goal. To this end, we introduce a linear alignment module in the first stage to fine-tune the pre-trained model. The fine-tuned model produces spatial distributions with clearer decision boundaries compared to other methods, demonstrating state-of-the-art performance in Real/Fake detection, achieving an average accuracy of 89.8%. Due to the lack of high-quality local detection data, we have created a dataset comprising real images, manipulated images, and corresponding masks (used to identify the manipulated regions) for research on local detection. For local detection, we innovatively introduce the rearranged spatial features from the first stage as auxiliary conditions into the second stage of local detection. We leverage the rearranged spatial features for local detection by employing cross-attention in the latent space as the backbone. Compared to other methods, our method can more accurately localize the manipulated regions in the source image, achieving state-of-the-art performance with an IoU of 61.5% and an F1 score of 71.2%.
This paper proposes an enhanced YOLOv7 algorithm specifically designed for small object detection on UAV platforms. The key improvements include the introduction of an optimized P2 detection layer, replacing the ineffective P5 layer for small object detection, and streamlining the network structure to reduce the number of parameters and computational cost. We introduce the Partial Agent Self-Attention (PASA) module to improve the aggregation of non-local information. Additionally, by incorporating depthwise separable convolutions and optimizing the structure, we further enhance the multi-scale feature extraction capability and computational efficiency of the Spatial Pyramid Pooling Cross-Stage Partial Connection (SPPCSPC) module. Furthermore, we propose a new feature extraction module, Faster Implementation of Enhanced CSP Bottleneck through Multiple Convolutions (FECM), which is implemented through multiple convolutions. We conduct experimental validation on the public Visdrone2019 and DroneVision-6 datasets. The experimental results show that the proposed method improves both precision and recall while reducing the number of parameters by 72%, providing a powerful solution for small object detection in complex UAV environments.
With the rising usage of artificial intelligence systems, social concerns around them and the need for regulation is also increasing, including requirements for data privacy. Federated Learning addresses data privacy concerns in distributed machine learning by training models collaboratively without centralizing data. However, Federated Learning is susceptible to Byzantine attacks, where malicious nodes submit corrupted updates. Krum, a robust aggregation algorithm, has been widely adopted as a defense mechanism. However, recent studies have shown that Krum’s performance degrades significantly in high-dimensional settings. This work proposes Layerwise Krum, a novel aggregation method that enhances Krum’s robustness in high-dimensional spaces while maintaining computational efficiency. We provide theoretical analysis of the improved robustness of Layerwise Krum compared to standard Krum. Furthermore, we empirically evaluate Layerwise Krum on various image classification datasets under diverse data distributions and Byzantine attack scenarios, consistently demonstrating superior performance compared to the original Krum operator.
Recently, the direct application of large language models (LLMs) for in-context learning (ICL) in the field of information extraction (IE) has demonstrated impressive performance. However, ICL still lags behind the supervised fine-tuning approaches. Although current works attempt to apply ICL to IE tasks, they primarily emphasize positive ICL demonstrations without fully exploiting the potential value of negative instances. Incorporating negative instances introduces additional contextual variations and enhances the model’s sensitivity to subtle semantic differences. In this paper, we propose ReNIE, a semantic retrieval and negative instances augmented IE framework. Our framework introduces Semantic-Level Retrieval (SLR) and a Robustness Enhancement Strategy (RES) to enhance the model’s generation process. SLR retrieves demonstrations with the highest semantic similarity to serve as the model’s in-context prompt, guiding the model to generate predictions that align with task characteristics. RES comprises two components: the boundary disturbance method and the synonym perturbation method. The boundary disturbance method generates negative samples with boundary errors, guiding the model to avoid potential boundary-related mistakes, while the synonym perturbation method creates negative samples with synonym variations, providing additional contextual information to enhance the model’s understanding of predefined IE categories. To validate our proposed method, we conduct experiments on 13 Named Entity Recognition (NER) datasets, 2 Relation Extraction (RE) datasets, 2 Event Extraction (EE) datasets, and 3 Event Argument Extraction (EAE) datasets. Experimental results demonstrate that our method achieves comparable performance across the majority of datasets in unseen domains.
Hate Speech Detection (HSD) plays a crucial role in ensuring respectful online communication and preventing the spread of harmful content. However, existing studies on HSD often focus on the overall intent of the target message while overlooking the fine-grained details in the message. We propose a RatIonale-guided multi-taSk lEarning framework (RISE), which frames HSD as the main task and Human Rationales Tagging (HRT) as the auxiliary task. This enables simultaneous message-level understanding and the identification of expressions that trigger human judgments of hate speech, mimicking human cognitive processing to capture fine-grained information and enhance HSD performance. Furthermore, we extend rationale annotations from binary labels to BIO tagging, capturing the positional roles of tokens within rationales. Additionally, we integrate an emoji semantics interpretation module that interprets emoji meanings, enriching contextual information. Extensive experiments demonstrate that RISE outperforms state-of-the-art models, with each component contributing significantly to improved performance.
Long-tailed recognition tasks face the critical challenge of extreme data imbalance, where overrepresented head categories dominate, while tail categories suffer from severe underrepresentation. This imbalance hinders the model’s ability to generalize to tail categories, a key bottleneck in real-world applications. Existing approaches, such as resampling strategies or loss function adjustments, often fail to balance the trade-off between head and tail categories. To tackle these challenges, we propose SELA-Net (Skill-specialized Experts with Bootstrap Latent Adaptation Network), a novel framework that synergizes a multi-expert design with adaptive layer-wise feature alignment to enhance representation learning for underrepresented categories. SELA-Net integrates two core modules: Skill-specialized Experts (SE) and Bootstrap Latent Consistency (BLC). The SE module leverages logit adjustment and Mixup techniques to optimize category-specific feature learning, achieving fine-grained representations for head, medium, and tail categories. Meanwhile, the BLC module enforces adaptive layer-wise consistency across feature representations, effectively enhancing the model’s generalization, particularly for tail categories. Experimental results on four widely used long-tailed datasets, including CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist2018, demonstrate that SELA-Net consistently outperforms state-of-the-art methods, achieving significant improvements in tail category recognition. Notably, SELA-Net excels in diverse test distributions and ablation studies, validating its robustness and adaptability in addressing long-tailed challenges. The code for this work is publicly available at: https://github.com/MuZiYuHui/SELA-Net.
Federated learning mitigates data leakage risks while maintaining training efficiency via gradient sharing. Nonetheless, previous studies have demonstrated persistent privacy vulnerabilities because attackers can reconstruct training data from shared gradients. Existing reconstruction methods assume that attackers can access all model parameters; however, sensitive parameters (such as user embeddings in federated recommendation systems) often remain private. Limited access results in inaccurate reconstructions. Using federated recommendation systems as a case study, we identify insufficient attack constraints as the origin of reconstruction failures. To address this limitation, we propose the MGradInv method, which leverages gradients from multiple training steps as additional reconstruction constraints. The experimental results demonstrate that this approach prevents convergence to local optima and reduces reconstruction errors by establishing sufficient constraints. We investigated two key factors affecting MGradInv ' s performance: target model convergence and gradient intervals. Results indicate that attacks are most effective during the early training stages but deteriorate as the model converges.The effects of MGradInv are clear even with gradient intervals of up to 230 steps. Our code and data are available here.
Reinforcement learning (RL) has gained significant attention due to its end-to-end learning capabilities and model-free nature. In the realm of vehicle engineering, active suspensions are crucial for enhancing both comfort and safety. The Deep Deterministic Policy Gradient (DDPG) algorithm, characterized by its stability and efficiency, offers valuable guidance for addressing the nonlinear issues in suspension systems. However, its learning efficiency may be compromised when dealing with high-dimensional state or action spaces. Therefore, this research primarily focuses on improving the DDPG algorithm by implementing enhancements by improved reward modules and integrating Long Short-Term Memory (LSTM), in order to maximize its control performance. By conducting five sets of experiment, the results demonstrate that the DDPG algorithm, after being integrated with various methods, exhibits superior control performance.
The increasing demand for reliable AI models is also evident in Hierarchical Energy Management Systems (HEMSs). Especially in complex energy systems like Renewable Energy Communities (RECs), efficient cost minimization is a strong incentive for potential users to participate. Nevertheless, trust in AI stems not only from its performance but also from AI Explainability (XAI). In this regard, the new Kolmogorov-Arnold Network (KAN) XAI paradigm offers significant advantages over Multi-Layer Perceptrons (MLPs). In this work, a KAN is optimized by a Genetic Algorithm (GA) to serve as an inference engine in a realistic REC HEMS. The main goal is to validate KANs in that application domain. Secondly, a custom coefficient space quantization is proposed to enable efficient KAN-GA encoding. Specifically, the generic KAN model is encoded as a GA individual and optimized to minimize the operational cost of the REC. Then, an explainable AI model is extracted from the original KAN by fitting its connection splines with simpler function forms and by applying the Kolmogorov-Arnold Theorem to get the output. The results show a high precision of the KAN models with a low computational cost. In addition, the explainable model performance is very close to that of the original KAN in terms of precision, while being significantly better in terms of computational efficiency, with about 10% time-saving. Therefore, developing fast and explainable KAN-based models is worthwhile in that application field.
Generalization and explainability are two key challenges in synthetic audio detection. Effective detectors should not only reliably classify unseen data from unknown synthesis algorithms, but also provide insight into their decision-making process and explain why a given input was classified as real or fake. To promote generalization we use the Person-of-Interest approach, which allows us to detect synthetic audio using a model trained only on real data, provided that some pristine audio of the putative speaker is provided. To support explainability, we instead use an encoder-decoder backbone such that the bottleneck features ensure syntactic and semantic fidelity to the input, as well as enable reliable decisions. Experiments show that our approach outperforms both state-of-the-art models based on supervised learning and methods based on speaker verification.
In urban planning, traffic flow prediction, a core component of Intelligent Transportation Systems, has made significant progress with the development of deep learning. The key problem of traffic flow prediction lies in capturing the complex spatio-temporal correlations in traffic flow. In recent years, more and more research has tended to apply Transformer-based models to solve this problem. However, Transformer-based models have two major limitations for traffic flow prediction: i) Most methods only focus on extracting data features within the attention head, while ignoring the correlation between these heads, making it difficult to integrate the multi-features of traffic data; ii) Most methods do not recognize the unique impact of nodes that serve as pivotal traffic hubs in the traffic networks, which cause the Transformer to excessively focus on the influence of non-pivotal nodes. In this study, we propose a novel Transformer-based model, the Multi-Feature Fusion Transformer (MF2former), aimed at addressing the above limitations of traffic flow prediction. MF2former incorporates the Augmented Synergistic Transformer Module to achieve a comprehensive multi-feature fusion of traffic data by enhancing the information capacity within self-attention heads and performing information fusion between these heads. Additionally, our model incorporates the Pivotal Node Module, which extracts pivotal nodes from all nodes and masks the global receptive field of the Transformer to enhance its focus on pivotal nodes in the traffic network. Our model is evaluated using two real-world traffic datasets, demonstrating superior performance compared to existing methods. This study provides a stable framework for accurate traffic flow prediction, offering valuable insights for urban planners and commuters.
