The pedestrian trajectory prediction is a crucial research topic in artificial intelligence application scenarios like autopilot and robotics. In these kinds of scenarios, the autopilot vehicle or robot should have a cautious interaction with human to avoid accident. Over the past decade, researchers have continuously proposed high-performance pedestrian trajectory prediction methods by leveraging the powerful tool of artificial intelligence. In particular, the spatial-temporal features based methods have been successfully applied. However, one potential issue with spatial-temporal features has been overlooked. Due to the sensitivity inherent in pedestrian dataset collection, the diversity of spatial features is far less than that of temporal features. Therefore, most spatial-temporal features based methods are tend to overfit to scenarios features, results in an unstable results across different scenarios. In our work, a Kinematic Temporal Conditional Variational Autoencoder (KT-VAE) that emphasizes the importance of temporal features along with a reliable spatial post-processing method is proposed. In KT-VAE, the spatial features are compressed instead of the temporal features to ensure that the model focuses more on the temporal continuity of pedestrian kinematic. This approach enables the VAE to better capture the temporal continuity and dynamic characteristics of pedestrian motion, while avoiding scenario overfitting that can result from insufficient spatial features. Through experiments, the KT-VAE maintains stability across different scenarios in cross-validation and demonstrates competitive performance in practical applications.
The growing demand for high-fidelity, large-scale molecular dynamics (MD) simulations has posed significant performance challenges for traditional MD applications, especially on memory-constrained heterogeneous platforms. To address these challenges on the SW26010-Pro processor, we propose a parallel optimization framework for miniMD. Introducing a full-neighbor force model to eliminate atomic data dependencies and an integrated segmented prefetching and caching mechanism to enhance memory bandwidth utilization. Meanwhile, we employ shared-bin neighbor list reconstruction for redundant access reduction, coupled with an asynchronous master-slave parallel strategy to fully exploit the heterogeneous resources of the SW26010-Pro. Experimental evaluations demonstrate that our framework achieves up to 12.7x speedup on a single node and sustains 84% parallel efficiency when scaling to 399,360 cores in billion-atom simulations. The proposed techniques are broadly applicable to MD workloads with irregular memory access patterns and limited cache resources, offering a general solution path for high-performance simulations on heterogeneous architectures.
Limited EEG data and subject variability pose significant challenges to the generalization of EEG-based emotion recognition. Most existing approaches augment EEG data using deterministic methods, often neglecting to ensure both diversity and fidelity in the generated samples. This oversight leads to insufficient domain diversity and emotional semantic information for a generalized model independent of individuals. This paper proposes an Information-Aware Graph Augmentation (InfoGA) framework for generalized EEG emotion recognition. The graph uncertainty augmentation module augments both the connectivity and features of EEG graphs by modeling statistical uncertainty, enabling the model to simulate domain shifts and improve generalizability against subject variability. Additionally, two information-aware constraints are introduced to ensure diversity and fidelity in the augmented EEG graphs. The graph diversity constraint enriches the emotional knowledge of the augmented graphs, while the graph fidelity constraint preserves their emotional semantic fidelity by integrating consistency learning with supervised learning. Extensive experiments on three public EEG emotion datasets, i.e., SEED, SEED-IV, and SEED-V, demonstrate that InfoGA achieves superior generalizability compared to baseline methods.
Visual information processing is an important part of industrial intelligence. In industrial settings, eliminating camera lens fog-induced blur poses a formidable challenge. Traditional dehazing methods based on imaging principles are difficult to meet the needs of industrial scenes. In recent years, methods based on complex deep networks, such as transformers, have shown better dehazing performance. However, their practicability is restricted by low adaptability to industrial scenarios. It is necessary to realize an effective and lightweight model that can be applied in industrial sites. For such a purpose, an effective image dehazing model named MFDehaz-Net is proposed. MFDehaz-Net has two specially designed components, i.e., Multi-scale Fusion Block and Point-Depth wise Block, which help it achieve a deep fusion of image features of different scales to obtain better global understanding. Besides, to reduce the damage to the original color of the image caused by image dehazing, MFDehaz-Net integrates the supervision signal in the frequency domain with a specially designed loss function. Experimental results demonstrate that MFDehaz-Net outperforms SOTA models in terms of dehazing ability with much shorter inference time.
While current machine-learning-based AI techniques have been spectacularly successful, their present applications still leave many important open questions - for example, how to make their results more reliable or, at least, how to gauge how reliable is each AI recommendation. In this paper, we argue that to fully answer these questions, we need to go beyond the current AI techniques, and that in this development, systems-, human-, and cybernetics-based ideas not only naturally appear, they seem to provide a way to the desired answers.
We propose a reinforcement-aware stochastic control framework for real-time reward optimization in sharing economy platforms. Unlike traditional static incentive schemes, our model dynamically allocates incentives by integrating belief-driven user behavior modeling, Nash equilibrium assumptions, and constrained utility maximization. The core framework unifies dynamic programming with reinforcement learning (RL) approximations to handle partial observability and large-scale deployment. A novel pricing-based calibration method is introduced to quantify the marginal value of a successful transaction, enabling budget-aligned incentive strategies. We further address theoretical assumptions, computational complexity, and practical implementation, providing a scalable path toward intelligent reward systems for real-world digital platforms.
Accurate pantograph-catenary attitude detection is essential for high-speed railway safety, yet existing methods face significant challenges when dealing with sparse keypoints, small target regions, and complex operational environments. To address these limitations, we propose PantoPoseNet, a novel two-stage framework designed for real-time pantographcatenary keypoint detection. Our approach introduces three key innovations in the first stage: (1) integration of VanillaNet with SIAF activation as the backbone network, achieving a 55.6% reduction in model parameters while preserving detection accuracy; (2) replacement of the conventional RepC3 module with CSP modules to enhance multiscale feature fusion capabilities; and (3) implementation of a hybrid loss function that combines GIoU and NWD metrics, specifically engineered to mitigate gradient vanishing issues inherent in small keypoint region detection. The second stage employs a specialized VanillaNet-KP network that processes 32x32 pixel regions to achieve precise keypoint localization. Comprehensive experiments conducted across various railway operational scenarios demonstrate that PantoPoseNet achieves superior performance with 98.2% mAP for keypoint region detection and 93.6% PCK for overall keypoint localization, while maintaining real-time processing at 25.94 FPS. These results significantly outperform current state-of-the-art methods, indicating strong potential for practical deployment in pantograph-catenary monitoring systems.
Accurate prediction of intraoperative hazardous events and generation of effective intervention plans are critical to surgical safety, but face multiple challenges of real-time, accuracy, and interpretability. Large-scale language models have potential, but their high cost and potential ‘illusion’ problems limit their application in real-time clinical environments. Traditional multitask learning models are efficient but knowledge-constrained, making it difficult to capture complex reasoning processes. To bridge this gap, this paper proposes a multi-objective distillation knowledge enhancement model-KEMO, which innovatively adopts a multi-objective chain-of-thought distillation framework to not only mimic the prediction results of the instructor’s LLM, but also explicitly migrate its structured reasoning process to the lightweight student model, which improves the answerability of the model by synergistically optimising the three objectives of event prediction, reasoning alignment and scenario generation. Interpretability. Meanwhile, combined with the Knowledge Graph-based Retrieval Augmented Generation mechanism, validated medical knowledge is dynamically injected to enhance the accuracy and reliability of decision-making and reduce model illusion. The experimental results show that the KEMO model significantly outperforms traditional models of the same magnitude in intraoperative hazardous event prediction and prognostic proposal generation, and achieves a performance comparable to that of a large faculty model.The KEMO model effectively bridges the gap between the large language model and the actual clinical application, and facilitates the transformation of the large model knowledge to the actual clinical deployment.
Current Multi-Modal Large language Models (MMLMs) primarily rely on instance-level feature statistics for cross-modal alignment. However, they commonly suffer three inherent limitations including vulnerability to outlier perturbations, neglect of inter-feature covariance structures, and local optimum trapping. These limitations stem from a critical oversight—existing approaches disregard the global statistical structure of multi-modal data, treating cross-modal alignment as isolated feature-level alignment rather than systematic distribution-level alignment. To address these issues, this paper proposes Layer-wise Covariance Alignment (LCA), which first leverages distribution-level alignment for cross-modal alignment. The effectiveness of LCA is validated through the use of parameter-efficient Low-Rank Adaptation (LoRA) on CLIP architectures. Experimental validation across eight benchmarks demonstrates state-of-the-art performance, confirming the critical role of distribution-level alignment in overcoming sample-level optimization constraints for cross-modal learning.
Sparkplug [1] is an emergent open-source software specification for Industrial Internet of Things (IIoT) systems, designed to favor data integration and device interoperability in an MQTT infrastructure. Although the security issues of IIoT systems can have relevant safety implications, Sparkplug only provides basic security features and essential, coarse-grained access control (AC) mechanisms. Effective AC solutions for Sparkplug-based IIoT systems still need to be designed, and, due to the Sparkplug’s increasing popularity and its recent definition as an ISO Standard [1], this has become a crucial need. To fill this void, this paper proposes an approach to efficiently enforcing fine-grained AC in Sparkplug-based IIoT systems. In particular, we define a fine-grained discretionary AC model and a related reference monitor implementing an efficient enforcement mechanism. Early performance evaluations show a reasonably low time overhead.
Smart grocery shopping presents a complex decision-making challenge involving dynamic pricing, real-time discounts, budget flexibility, and trip cost optimization. This paper proposes a mathematical framework that models shopper behavior through a stochastic partial differential equation (PDE), capturing utility evolution under uncertainty. The model integrates shopper preferences, inferred discounts from nearby consumers, and stochastic price fluctuations to formulate a utility maximization problem subject to soft budget constraints and trip cost penalties. A novel assignment matrix ensures each item is purchased from exactly one store, enabling the derivation of a closed-form solution for optimal item quantities and store selections. The Lagrangian formulation enforces budget flexibility while a penalty function captures psychological aversion to overspending. Designed for efficient implementation on mobile devices, the model is suitable for deployment in real-time consumer applications. This framework offers a principled, interpretable, and deployable tool for intelligent grocery shopping in data-rich environments.
Path planning is an essential task for the mission execution of unmanned surface vehicle (USV). However, existing advanced techniques based on deep reinforcement learning (DRL) often suffer from low learning efficiency and insufficient environmental perception from single-sensor configurations. To address these issues, this paper proposes a hybrid framework named A*D3QN, which integrates the heuristic efficiency of the A* algorithm with the adaptive decision-making of a dueling double deep Q-network (D3QN), associated with the multimodal data fusion for precise environment modeling. The proposed A*D3QN incorporates prior knowledge from the global paths generated by A* to initialize and guide the D3QN learning process. The prior knowledge, formatted as RL transition tuples, is used in the reward function design and the N-step prioritized experience replay, which significantly accelerates overall learning efficiency. Moreover, an improved D3QN architecture is designed to dynamically fuse visual data and navigation states via a cross-entropy attention mechanism, enabling multimodal perception in partially unknown environments. Extensive experiments across three scenarios with varying obstacle densities demonstrate that A*D3QN significantly outperforms state-of-the-art DRL baselines. Ablation studies further validate the necessity of each component.
This paper proposes a robust composite control strategy for constrained continuous-time nonlinear systems by integrating sliding mode control (SMC) and model predictive control (MPC). SMC enhances disturbance rejection, while MPC handles constraints by solving an optimal control problem (OCP) based on the SMC input. The resulting control input ensures both robustness and constraint satisfaction. To improve computational efficiency, the OCP is solved only at sampling instants. Recursive feasibility and closed-loop stability are rigorously analyzed, and the method’s effectiveness is demonstrated on a cart-damper-spring system.
In recent years, GNN-based fake news detection models integrating news content, user characteristics, and propagation structure have gained substantial attention, yet they often face the potential homogenization issues in GNNs, limiting performance in detection. Despite numerous studies focusing on sophisticated models to tackle this issue, many have overlooked the unique structural characteristics of propagation trees. Here, we propose a structure-oriented model named DaFAN, which leverages a dual-attention mechanism to not only address the homogenization issue in message passing but also be able to boost the distinction between true and fake news. In specific, we design a novel Dual Attention Module with the multi-head graph attention mechanism to fuse the multi-modal features by utilizing the inherent characteristics of news propagation trees, and introduce a lightweight feature augmentation module compatible with various GNNs to retain the initial features and optimize the feature selection. Experiments on real datasets demonstrate that our DaFAN model outperforms the state-of-the-art models. Furthermore, the feature augmentation module has notably bolstered our model’s transferability across languages and datasets, fine-tuning on 10% of the target data can significantly surpass the supervised training from scratch.
Arc fault diagnosis constitutes a critical challenge in fault diagnosis, focusing on rapid and precise identification of arc-induced safety risks in power systems. Previous methodologies hard to strike a balance between real-time performance and detection accuracy in arc fault diagnosis systems. To address this challenge, we propose a linear-time Mamba-based model enhanced with a Spatial Awareness Module(SAM), achieving real-time DC arc fault diagnosis. Specifically, our approach leverages a state-space model (SSM) framework and employs a hardware-aware parallel algorithm for efficiency. To further improve accuracy while maintaining the computational efficiency of the base model, we integrate a spatial awareness module to capture global features, enabling precise fault diagnosis. Experimental results demonstrate that our method achieves 96.72% accuracy with a 1.87 ms response time, making it highly suitable for industrial applications where rapid and reliable arc fault diagnosis is critical. This advancement holds significant promise for enhancing safety in industrial electrical systems.
Deploying AI services on battery-powered mobility platforms such as autonomous vehicles, mobile robots, and large scale IoT sensor networks requires determining the most suitable execution environment for each workload across the cloud, edge, and device computing options. Because every placement option imposes different trade-offs among Accuracy, Latency, and Energy efficiency (ALE), stakeholders face a difficult, mission-critical decision that existing studies seldom address in a holistic, mission-aware manner. To fill this gap, we introduce the Mission-driven ALE (MALE) evaluation method. MALE couples ALE metrics with explicit mission objectives by allowing analysts to apply customizable weights to each criterion. The evaluation results are aggregated and visualized as heatmaps, helping transform a previously heuristic and opaque placement decision (black-box) into a more transparent and interpretable process (white-box). We examine the applicability of MALE through three representative case studies: Autonomous Vehicles, Real-Time Robotics, and IoT Sensor Networks, each reflecting distinct ALE priorities. By supplying a structured, mission-aware decision-support method, MALE strengthens stakeholder confidence and accelerates the optimization of AI service placement across the cloud-edge-device continuum, offering a practical foundation for future validation in real-world deployments.
Aiming at the problem that the traditional YOLOv7 network structure is difficult to accurately and efficiently detect the growth status of strawberries, an improved YOLOv7 network model is proposed to more efficiently and accurately identify the growth status of strawberries. Firstly, by introducing the improved Transformer module self-attention mechanism, Bottleneck Transformers (BotNet) and CBAM to obtain better feature extraction capabilities, a new model is proposed, which makes the model have a more efficient and accurate recognition rate for the growth status of strawberries compared with the traditional model. Secondly, by introducing the Normalized Wasserstein Distance (NWD) loss function, the original IoU algorithm of YOLOv7 is improved to solve the classification and regression problems of the model, and at the same time, the model’s detection ability for tiny objects is further improved, so that the model’s detection accuracy for each category is further improved. Combined with the above improvements, an improved YOLOv7 model (BotNet-YOLOv7) combining Bot-Net and YOLOv7 is proposed. The results show that the average recognition accuracy (mAP0.5), Precision, and Recall reached 96.4%, 92.2%, and 90.9%, respectively, which are 4.3%, 5.9%, and 4.8% higher than YOLOv7, and the memory usage has been greatly improved, which is 24% higher than the traditional model, providing better technical support for strawberry growth status detection.
To address the limitations of existing reversible data hiding (RDH) methods, including low redundancy space utilization, static embedding strategies, and insufficient image adaptability, this paper presents an adaptive dynamic capacity allocation reversible data hiding scheme based on interpolation and integer wavelet transform (ADWC-RDH). By implementing a spatial-frequency domain collaborative optimization mechanism, the proposed method generates multi-dimensional redundancy spaces through combining nearest neighbor interpolation with integer wavelet transform. We design dynamic capacity allocation and low-bit priority embedding strategies to achieve adaptive adjustment of embedded bitrates. Experimental results demonstrate that the proposed method achieves significant improvements in average embedding rate across test images, along with excellent Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) performance. ADWC-RDH achieves substantial enhancements in embedding capacity and visual quality for complex texture images through dual-domain collaborative optimization and dynamic capacity allocation, thereby demonstrating superior universality and robustness.
Integrating artificial intelligence into educational technology presents great opportunities for automated educational systems. These systems could relieve teacher resources and support underperforming students. However, creating systems that are adaptive, scalable, and factually correct is resource intensive. Furthermore, there are many technologies that are prevalent, but lack systematic ways to integrate them into existing educational technologies. Building on reinforcement learning and large language models (LLMs), this paper introduces a multi-agent framework for adding both a reinforcement learning-based tutor and an LLM-driven peer to educational systems. The integrated architecture is unified with a central ontology, acting as a symbolic knowledge base and facilitating data transformation. We also detail a novel windowed experience sharing method for improving reinforcement learning training efficiency when dealing with similar environments and low-data situations. We present our architecture and simulated results to verify the reinforcement learning algorithm as an adaptive tutor, as well as the integration of an LLM-driven peer and educational outcomes from this integration.
