Keywords: Evolutionary Computation, Human Factors, Multi-User Interaction
Abstract: In recent years, generative AI has been attracting attention. Generative AI acquires its generation methods through automatic learning of past data. ChatGPT is a prime example of generative AI that responds to requests from humans using natural language. Meanwhile, people on the web have traditionally gathered their knowledge, experience, and creativity to solve a variety of problems. Furthermore, in the field of evolutionary computation, there is human-based evolutionary computation (HBEC), in which human groups solve problems using evolutionary computation procedures. Now that generative AI has emerged, it is necessary to identify problems for which problem-solving methods based solely on the power of human groups as HBEC are effective. Therefore, in this study, we experimentally compare the problem-solving performance of HBEC with that of ChatGPT for two types of problems: problems where the web is overflowing with information that can provide clues for solving the problems, and problems where there are few such problems on the web. The experimental conditions here, such as the fact that the subjects were only university students and the number of subjects was very small, do not allow for the drawing of general conclusions. However, the following tendency was observed from the experiment: ChatGPT is more suitable for the former type of problem and HBEC system is more suitable for the latter type of problem. This study is the first to compare HBEC with generative AI, and its greatest contribution is to initiate such comparative research.

Keywords: Evolutionary Computation, Metaheuristic Algorithms, Optimization and Self-Organization Approaches
Abstract: This paper proposes a new type of black-box optimization problem called the pattern-first problem (PFP). While variable values are directly used to calculate the fitness value in conventional problems, the instances of PFP require the correct pattern created by the variable values to proceed to the calculation of the fitness value. For example, a fitness value can be calculated only if the values of the first and second pair of variables are the same as the values of the third and fourth pair of variables. Also, this paper proposes the bird song grammar evolutionary algorithm II (BSGEA-II). BSGEA-II generates solution candidates by pairing male individuals, who create solution patterns according to a grammar, with female individuals, who determine the variable values that match the patterns. BSGEA-II is expected to efficiently solve PFP by simultaneously searching for patterns and variable values that match the patterns. Evaluation experiments show that BSGEA-II outperforms conventional methods that directly evolve only the values of variables when solving simple instances of PFP that are ideal for BSGEA-II, combining patterns that BSGEA-II can generate in principal with OneMax problem.

Keywords: Evolutionary Computation, Computational Intelligence, Heuristic Algorithms
Abstract: For multimodal optimization problems (MMOPs), existing algorithms often lack sufficient adaptability to adjust their search behavior in response to the evolving states of individual solutions. To overcome this limitation, this paper proposes a hybrid search framework guided by fitness landscape characterization for MMOPs. Two indicators, namely local fitness advantage (LFA) and neighborhood median–mean offset (NMMO), are introduced to characterize the local states of individuals, based on which heterogeneous evolutionary strategies are adaptively assigned. Extensive experiments conducted on the CEC2013 multimodal benchmark suite demonstrate that the proposed approach outperforms representative niching methods in terms of peak ratio (PR) and success rate (SR), while also exhibiting superior convergence performance. Overall, the proposed framework offers an effective and adaptive solution for addressing MMOPs.

Keywords: Evolutionary Computation, Computational Intelligence, Heuristic Algorithms
Abstract: Small-Scale Learning Particle Swarm Optimization (SSLPSO) has demonstrated promising performance on large-scale optimization problems (LSOPs). However, its effectiveness is often hindered by several factors, including the degradation of distance-based discrimination in high-dimensional spaces, the lack of adaptive population-size control, and sensitivity to the initial population distribution. To overcome these limitations, this paper proposes an enhanced SSLPSO approach that incorporates three key improvements. First, a composite distance measure is introduced by integrating global, local, and informative-dimensional distances to improve discrimination capability in high-dimensional search spaces. Second, a fitness-based population-size control strategy dynamically adjusts the search effort based on the swarm's evolutionary state. Third, a multi-hypercube initialization scheme is employed to ensure a well-dispersed initial population, thereby enhancing exploration in high-dimensional spaces. The proposed method is evaluated on the CEC2013 LSGO benchmark suite and compared with the original SSLPSO and several state-of-the-art algorithms, including HCLPSO, MuDSDE, SDLSO, and IGKT. Experimental results confirm that the proposed enhancements significantly improve the robustness and efficiency of SSLPSO for large-scale optimization scenarios.

Keywords: Intelligent Transportation Systems, Distributed Intelligent Systems, Energy Systems
Abstract: Cloud-hosted large multimodal models (LMMs) can provide strong open-vocabulary perception for Vehicle-to-Everything (V2X) systems, but naively transmitting full-resolution frames from edge to cloud causes severe communication overhead and high cloud-side prefill latency. We present CABLE, a closed-loop adaptive bandwidth-efficient LMM-based encoding framework for edge-cloud perception. CABLE propagates the previous cloud segmentation mask on the edge using ego-motion compensation, refines it with residual-motion cues, and consolidates disconnected regions via a corridor envelope to form a robust region of interest (ROI). Only ROI-masked images are uploaded, while the cloud segmentation output is fed back as the prior for the next frame, forming a mask-to-ROI-to-LMM closed loop. Experiments on five datasets (nuScenes, WOD-ZB, Waymo, KITTI, and CADC) show consistent communication savings and perception preservation, achieving 73--87% bandwidth reduction with 5--8times estimated LMM prefill speedup while preserving downstream detection performance relative to full-frame inference.

Keywords: Neural Networks and their Applications, Decision Support Systems, Human Performance Modeling
Abstract: Assessing air traffic controller workload is essential for airspace monitoring and decision support. However, direct workload measurements, while possible through human-centered studies, are difficult to obtain at scale in operational settings. This paper proposes a trajectory-deviationbased framework to construct a sector-hour workload proxy and predict its temporal evolution. The framework derives interpretable indicators from surveillance trajectories relative to planned routes, including cross-track deviation, deviation magnitude and duration, altitude level changes, and ground speed changes, which are aggregated and combined into a composite workload proxy. The resulting proxy time series is used to train and evaluate learning-based temporal models that forecast sector-hour proxy values from historical patterns and contextual features. Experiments on the Swedish Civil Air Traffic Control (SCAT) dataset show that learning-based models consistently outperform the baseline, with gradient boosting and long short-term memory (LSTM) models achieving comparable accuracy. These findings show that trajectory- and flight-plan-derived workload proxies are both interpretable and predictable, supporting scalable and reproducible workload analysis and short-horizon forecasting.

Keywords: Fault Monitoring and Diagnosis, Neural Networks and their Applications, Cyber-physical systems
Abstract: Multirotor unmanned aerial vehicles (UAVs) experience degraded control authority under partial actuator faults, jeopardizing mission reliability and safety. Data-driven actuator-fault-diagnostic methods require large labeled datasets and offer limited physical interpretability, whereas model-based methods demand complex parameter tuning and incur substantial computational overhead. This paper proposes a physics-informed neural network (PINN)-based Kolmogorov-Arnold Networks (PIKAN) for actuator fault diagnosis in quadrotor aerial robots, pairing a KAN backbone with a discrete Newton-Euler physics loss to enforce dynamically consistent predictions of per-actuator thrust factors. The KAN backbone captures sharp actuator-fault signatures through spline-parameterized edge functions, while the physics loss regularizes severity predictions across unseen actuator-fault magnitudes. PIKAN is evaluated against three ablation baselines (MLP, KAN, PINN) that remove one or both components, across a range of training-set fractions and actuator-fault magnitudes. At unseen actuator-fault magnitudes, PIKAN reduces fault-only severity mean absolute error by up to 62.5% relative to a data-only KAN baseline, with 51 to 63% reductions sustained as training data shrinks. The ablation shows that the KAN backbone enables precise modeling of sharp actuator-fault transitions while the physics loss delivers cross-magnitude generalization.

Keywords: Swarm Intelligence, Game Theory and Multi-Agent Systems, Reinforcement Learning
Abstract: High-level task language offers an intuitive interface for supervising large robot swarms, where collective behavior must emerge from local interactions under limited information. However, most swarm controllers rely on fixed interaction rules, while language-guided robotic systems are typically designed for single-robot execution or centralized coordination. This creates a key challenge: how can structured task commands be translated into decentralized local interaction laws that preserve safety, maintain coherence, and adapt to changing conditions?This paper formulates language-conditioned decentralized swarm control and proposes a semantic constraint propagation framework. Each command is mapped to a structured semantic representation and propagated into adaptive weights over a set of interpretable swarm primitives, including separation, cohesion, alignment, goal-attraction, and obstacle-avoidance. Rather than generating low-level actions directly, the controller modulates the relative influence of these primitives, allowing coordinated swarm behavior to emerge from local interactions. A safety-aware projection layer enforces minimum-spacing and exclusion constraints, ensuring robustness during dynamic execution.

A 2-D simulation environment is developed to evaluate the framework under balancing, region exclusion, redistribution, and command-switching scenarios. Preliminary implementation results demonstrate end-to-end feasibility, stable safety behavior, and semantically consistent modulation of local interaction dynamics. In particular, the framework enables structured adaptation during command transitions, where different regions of the swarm respond differently under the same global instruction. The proposed approach provides a scalable and interpretable mechanism for translating high-level task intent into decentralized swarm coordination.

Keywords: Autonomous Vehicle, AI and Applications, Control of Uncertain Systems
Abstract: Real-time quadrotor control is often degraded by regime-dependent modeling mismatch from aerodynamic effects, payload variation, actuator dynamics, and aggressive maneuvering. Learned residual models can improve prediction accuracy, but injecting them directly into the control loop risks unreliable behavior in shifted regimes. This paper proposes a stability-governed regime-aware neural ODE framework that augments a nominal quadrotor model with a learned residual acting only in the acceleration channels, preserving the nominal kinematic structure. A stability-aware training objective and a Lyapunov-compatible admissibility condition jointly ensure the learned correction is admitted, attenuated, or rejected online according to uncertainty surrogates, retaining the baseline stabilizing action when the residual becomes unreliable. Real-time experiments on a Pixhawk-based quadrotor across hover, aggressive flight, payload variation, and disturbed-air regimes demonstrate improved prediction and reduced tracking error over both a nominal baseline and an unguided augmentation. Ablation confirms that regime awareness, stability-aware training, and governed deployment each contribute materially.

Keywords: Fault Monitoring and Diagnosis, Autonomous Vehicle, Systems Safety and Security
Abstract: In closed-loop autonomous systems, distinguishing physical faults from cyberattacks is essential because the two mechanisms call for different mitigation actions. This attribution problem becomes difficult when feedback suppresses or reshapes distinct anomaly signatures, creating similar output-level responses between fault and attack cases. Under output-response matching, conventional residual methods can detect abnormal operation but provide limited evidence for attributing the underlying mechanism. This paper proposes a rule-based attribution framework that uses temporal causal structure rather than a trained discriminative classifier. Instead of comparing signal-level deviations, the proposed method uses PCMCI+ to infer a temporal causal graph over a domain-knowledge-guided set of sensing, state, and actuation variables, and quantifies node dominance through the Hub Identity Index. Evaluation in a QBall-2 quadrotor unmanned aerial vehicle hover setting with output-response-matched anomalies shows that the matched IMU abrupt fault and FDI attack induce different dominant hubs. The IMU abrupt fault makes the physical pitch state dominant, whereas the FDI attack makes the controller-side pitch measurement dominant. By shifting the attribution evidence from matched output responses to structural causal dominance, the proposed framework provides a transparent and interpretable rule for fault-attack attribution in closed-loop autonomous systems.

Keywords: Machine Learning, Medical Informatics, Application of Artificial Intelligence
Abstract: Traditional survival models in environmental epidemiology often struggle to capture non-linear exposure–response relationships, time-varying clinical trajectories, and spatial confounding. To address these limitations, we propose an explainable artificial intelligence framework for dynamic survival modelling and apply it to a highly exposed industrial cohort in Taranto, Italy. Using a two-stage landmark-based pipeline, we compare Random Survival Forest and XGBoost across multiple prediction horizons while incorporating time-dependent clinical covariates. XGBoost achieved the best predictive performance, with a bootstrap-validated concordance index above 0.93. Feature selection identified PM2.5, PM10, and NO2 among the most influential predictors of long-term mortality. SHAP analysis, supported by piecewise Cox modelling, revealed non-linear exposure–response patterns with a changepoint around 11.8 μg/m³, beyond which predicted mortality risk increases more steeply. Spatial propensity score matching indicated a survival disadvantage in highly exposed districts after adjustment for major observed confounders. Finally, under an in-distribution counterfactual scenario reducing exposures to the 10th percentile of the local distribution, Restricted Mean Survival Time analysis estimated an average gain of approximately 49 life-days per patient over 5 years. Overall, the proposed framework provides an interpretable and policy-relevant approach for environmental health risk assessment.

Keywords: Assistive Technology, Human-Machine Interaction, Deep Learning
Abstract: Unintentional falls pose a severe threat to vulnerable populations, particularly the elderly and visually impaired. While traditional assistive systems focus on post-impact fall detection, there is a critical need for proactive risk anticipation to enable timely intervention. Existing skeleton-based approaches often treat human motion as independent vector sequences, ignoring the spatial connectivity crucial for identifying biomechanical instability. In this paper, we propose Adaptive Risk-GCN, a privacy-preserving and topology-aware framework designed to anticipate falls before ground impact. By integrating a learnable adjacency matrix into a Spatial-Temporal Graph Convolutional Network (ST-GCN), our model dynamically captures non-local joint compensations during the loss of balance. Furthermore, we introduce a granular three-stage risk assessment strategy to distinguish subtle pre-fall precursors from complex Activities of Daily Living (ADLs). Extensive experiments on the UR Fall Detection Dataset demonstrate that the proposed Adaptive Risk-GCN significantly outperforms traditional sequence models (e.g., LSTM), achieving a highly competitive overall accuracy of 88.0% while strictly minimizing false alarms. Crucially, the system secures an average Time-to-Fall (TTF) lead time of 500 to 1000 ms, proving its viability for triggering real-time proactive interventions, such as wearable airbags, in practical human assistive environments.

Keywords: Biometrics and Applications, Human Enhancements, Cyber-physical systems
Abstract: Gait dysfunction is a major motor symptom of Parkinson's disease (PD) and an important target for neuromodulation-based rehabilitation. However, conventional clinical scales and simple gait metrics often fail to capture coordinated, system-level adaptations that emerge from the tightly integrated interactions across cortical, muscular, and kinematic domains. This limitation is particularly pronounced for transcranial direct current stimulation (tDCS), whose reported effects in PD vary widely across individuals and whose underlying neuro-motor mechanisms remain unresolved when evaluated through single-domain outcomes. This study proposes a multimodal framework to characterize neuro-musculo-kinematic changes induced by tDCS. Simultaneous EEG, EMG, and motion capture data were acquired from individuals with Parkinson's disease, from which 48 cross-domain features were systematically extracted. Compact multimodal feature subsets associated with tDCS and sham stimulation were identified through a Backward Beam Search with cross-validated feature selection. TreeSHAP analysis localized the dominant changes to the unaffected limb, including elevated tibialis anterior median frequency, temporally focused gastrocnemius activation, and increased cadence, while EEG features contributed only within the combined muscular–kinematic context. The derived nine-feature subset achieved F1 = 0.793 and AUC = 0.844 with statistical significance (p = 0.031) under exhaustive subject-level permutation testing. This AUC was unattainable by any single-modality configuration (max 0.737). These results indicate that tDCS elicits a structured cross-modal redistribution of neuromuscular control, offering a framework for mechanism-informed neuromodulation.

Keywords: Human-Machine Interaction, Human Enhancements, Assistive Technology
Abstract: Tonic vibration reflex (TVR) and kinesthetic illusion (KI) induced by vibration stimulation to muscle or muscle tendon might be applicable to neurorehabilitation for paralyzed persons such as stroke survivors. It is found that the amount of TVR and KI is affected by changing skin surface temperature of the stimulated muscle although they are basically affected by the vibration parameters. The amount of TVR and KI increased as the skin surface temperature of the vibrated muscle is cooled and decreased as the skin surface temperature of the vibrated muscle is warmed. In this paper, the effective local temperature stimulation area is experimentally investigated. In the experiment, the amount of induced TVR and KI were evaluated on three conditions of thermal stimulation area (i.e., small area, medium area, and wide area). Furthermore, three-level temperature (i.e., cool, normal, warm) is prepared for each area. Therefore, seven conditions of temperature settings are prepared in total including the condition of no stimulation in the experiments. The experimental results showed that the wider the temperature setting area, the greater the effect of the temperature change to the amount of induced TVR and KI. The experimental results also showed that the effect of temperature stimulation can be observed even in the small thermal stimulation area.

Keywords: Computational Intelligence, Systems Safety and Security, Autonomous Vehicle
Abstract: Mobile robots operating in human-centered environments are required to ensure safety while flexibly adapting their behavior. However, real-world environments contain small-scale obstacles such as cables and power strips, which are often not properly represented due to insufficient point cloud density and sensing limitations. In particular, topological map-based methods depend on the distribution of observed data, and thus may fail to capture local geometric structures, leading to degraded path generation.

In this study, we propose a path planning method that locally refines environment representation based on uncertainty. Specifically, regions with insufficient information are detected from discrepancies between a topological map constructed from LiDAR point clouds and the observed data, and are defined as attention nodes. By selectively updating the environment representation using these nodes, the proposed method integrates perception and action in path generation.

Experimental results in indoor environments with small-scale obstacles demonstrate that attention nodes are appropriately generated around obstacles, and that updating the topological map improves path generation. In particular, the proposed method enables avoidance of small obstacles and replanning behavior.

These results suggest that the proposed approach achieves safe and adaptive navigation even under incomplete environment representations.

Keywords: Decision Support Systems, Cyber-physical systems, Wearable Computing
Abstract: One in four musculoskeletal (MSK) rehabilitation patients experiences clinically meaningful decline that goes undetected between episodic assessments. This paper introduces the Digital Smoke Alarm, a proof-of-concept wearable decision support system that converts 14 days of passive Fitbit data into a binary clinical risk flag, a window we term the Cognitive Assistive Layer. Six theory-grounded behavioural and physiological features were engineered from activity and resting heart rate data for 60 patients (15 decliners, 25%). An interpretable, penalised logistic regression model was developed; all feature associations are exploratory and require prospective replication. Clinical value is demonstrated through decision curve analysis, which confirms positive net benefit over both treat-all and treat-none alternatives, and through an operational profile that reduces clinician monitoring burden by 73% while detecting 47% of eventual decliners (missing 53%, a trade-off adjudicated by the clinician at each stage). Subgroup analyses show that activity-based sensing detected physical decline with an AUC of 0.77, compared to 0.64 for psychological decline. All findings are proof-of-concept and require prospective validation before clinical use.

Keywords: Autonomous Vehicle, Cognitive Computing, Systems Safety and Security
Abstract: As autonomous vehicles progress toward higher levels of autonomy, ensuring their safety has become a central concern in both research and development. Detecting and recovering from hazardous scenarios are critical components in the design and deployment of autonomous driving systems. Current detection mechanisms primarily rely on sensor performance and the accuracy of learning algorithms, which are often deep learning–based and require large amounts of training data. As a result, these methods may struggle to effectively learn from and prevent rare real-world accidents, which occur far less frequently than the large volumes of data required by typical machine learning approaches.

This paper introduces a novel mechanism for detecting and preventing accident-prone scenarios, drawing inspiration from human error theories rooted in cognitive psychology. Based on two principles of how humans learn from critical mistakes, the proposed mechanism efficiently detects and prevents accident-prone situations by modeling the intervention behaviors of human drivers or supervisors. A case study using Autoware simulation was conducted to assess the feasibility and benefits of the approach in a real accident-prone scenario. The results demonstrate the feasibility of the proposed approach for detecting and mitigating accident-prone situations after only two human interventions. We recommend the proposed mechanism as an effective complement to enhance learning efficiency in critical situations for Level 2 and 3 autonomous vehicles. Future work will focus on developing additional scenarios and extending the approach to non-critical situations, such as personalized cruise control.

Keywords: Human Factors, Human-Computer Interaction, Brain-Computer Interfaces
Abstract: Eliciting explainable AI (XAI) requirements from stroke survivors presents a methodological challenge with direct implications for the design of trustworthy brain-computer interfaces for rehabilitation. How can patients and caregivers articulate preferences about algorithmic transparency when they lack conceptual frameworks for explainability, and when standard elicitation approaches are structurally inadequate for users with acquired communication disorders? We present a video-based scaffolding protocol for XAI requirements elicitation, developed and piloted in a rehabilitation context. In a formative study with three stroke survivors (two with moderate-to-severe aphasia) and three caregivers, facilitators employed four scaffolding approaches alongside the videos: 1) analogical bridging mapping AI states to familiar systems, 2) projective personas depersonalising sensitive topics, 3) binary forcing reducing cognitive load, and 4) extended response time. These approaches successfully surfaced heterogeneous, sometimes conflicting XAI needs across participants. Reflexive analysis additionally revealed three systematic facilitation biases, namely, normative bias, hypothesis confirmation bias, and presence effect, where scaffolding inadvertently shaped responses. We present these as protocol risk guidelines for practitioners. Together, the protocol and guidelines constitute a reusable methodological contribution for eliciting patient-facing XAI requirements in rehabilitation, arguing that such elicitation is a necessary prerequisite for trustworthy human-machine systems design, not an optional preliminary.

Keywords: Augmented Cognition, Computer Interfaces, Human-Machine Cooperation and Systems
Abstract: Cooperation and defection in social dilemmas are often studied using outcome-locked neural responses, yet converging evidence suggests that endogenous brain states pre- ceding a choice may already constrain forthcoming decisions. Here, we test whether upcoming cooperation versus defection in an iterated triadic Prisoner’s Dilemma can be predicted from pre-decision EEG alone. EEG was recorded from 33 participants (11 groups), epoched to decision onset, and features were extracted only from the pre-decision interval (−3 to 0 s). We constructed a 120-dimensional feature set spanning spectral power and asymmetry, slow potentials/readiness-like indices, fronto-parietal phase coupling, entropy/complexity, envelope variability, and phase–amplitude coupling. Classification was performed using an XGBoost model with fold-wise SHAP feature selection (Top-15), evaluated using pooled-trial stratified 5-fold cross-validation across participants (i.e., cross-trial classification across subjects). To address class imbalance (cooperation:defection ≈ 4:1), resampling was applied to the training split only, while the held-out test split retained the original class distribution. The proposed approach achieved 0.8097±0.0172 accuracy and 0.7933±0.0183 balanced accuracy (F1 = 0.6172±0.0262) across folds. SHAP-based interpretation indicated that cross-frequency coupling, fronto-central oscillatory features, and slow/complexity-related state metrics contributed to prediction. These results suggest that pre-decision EEG states contain discriminative information associated with upcoming social choices in a multi-agent dilemma, supporting the feasibility of pre-decision decoding in a pooled-trial setting.

Keywords: Cognitive Situation Management, Decision Support Systems, Environmental Sensing
Abstract: Air-pollution-related mortality is driven by the interplay of environmental exposure, demographic vulnerability, clinical conditions, and spatial context. However, these factors are often analyzed through separate pipelines, limiting integrated and interpretable assessments of environmental health risk and reducing stakeholders’ situation awareness. To address this limitation, this paper proposes a Situation-Aware Cyber-Physical-Social System (CPSS) for explainable environmental mortality analysis, aimed at supporting improved human–machine interaction and human-centered decision-making in complex environmental health scenarios. The framework integrates heterogeneous environmental, demographic, and clinical data through sensing and perception stages, producing refined features for downstream analysis. In the comprehension phase, it combines survival modeling, spatial balancing, and explainability techniques to assess the impact of pollutant exposure on mortality, providing interpretable insights and visual analytics to support user understanding. In the projection phase, counterfactual analysis enables interactive exploration of what-if scenarios for informed decision-making.

The methodology supports interpretable territorial risk assessment and is preliminarily evaluated within the MISTRAL Nexus platform in Taranto, Italy, as an explainable and interactive decision support environment for policymakers and domain experts.

Keywords: Image Processing and Pattern Recognition, Deep Learning
Abstract: Abstract—Unpaired underwater image enhancement is challenging because degraded and clear images are sampled from independent domains without pixel-level correspondence. In real scenes, wavelength-dependent attenuation, scattering, and illumination variation further cause color distortion, contrast degradation, and structural loss. To address these issues, we propose PhysDec-SB, a unified framework for unpair -ed underwater image enhancement. PhysDec-SB contains an offline physics-guided pseudo-label construction stage and a joint pseudo-label-guided Schrödinger Bridge (SB) restoration stage. Specifically, adaptive dark channel prior (DCP) statistics are extracted from degraded images to build physical descriptors and generate offline image-level pseudo-labels. These labels provide weak degradation supervision and organize the degradation residual space rather than directly supervising the output. During restoration, a lightweight NAFNet-based shared generator performs barycenter–residual latent decoupling and is repeatedly applied across bridge states to progressively transport degraded samples toward the clear-image distribution. In this way, physical priors guide degradation-aware representation learning, while SB-based transport improves unpaired domain alignment. Extensive experiments demonstrate the effectiveness of PhysDec-SB on real-world underwater datasets.

Keywords: Decision Support Systems, Human-Machine Cooperation and Systems, Human-Machine Interaction
Abstract: Cryptocurrency airdrops (free token giveaways to attract early users) are routinely exploited by strategic actors who create many wallets to claim rewards repeatedly. For operators, the challenge is not only identifying suspicious accounts, but configuring rules they can understand, audit, justify, and adapt across campaigns. Using 2.5 million address-project records from 16 Ethereum projects, we analyze every possible combination of five human-understandable behavioral indicators and quantify each indicator's contribution under an internal-label utility. Three findings characterize this rule-design space: (1) under the selected internal-label utility, fund-flow indicators receive the largest Shapley contribution shares; (2) using HasciDB's five-indicator OR label as the reference, the HF+RF+MA subset retains 78% of positive address-project records; and (3) externally listed address groups exhibit different indicator-trigger profiles in their available HasciDB histories. The external comparison measures flagged-list agreement rather than independent validation of detection accuracy. The result is a compact, auditable summary for examining and documenting candidate screening configurations, not a validated universal defense.

Keywords: Discrete Event Systems, System Modeling and Control, Manufacturing Automation and Systems
Abstract: In automated manufacturing systems (AMS), the occurrence of partial deadlocks can cause significant operational disruptions. Traditional detection methods based on the reachability graph (RG) of their Petri net models suffer from the state-space explosion problem. Conversely, structure-based methods, while avoiding RG construction, either fail to detect all deadlocks or cannot distinguish which ones are reachable from the initial marking of Petri nets. This paper presents an innovative method for detecting reachable partial deadlocks in automated manufacturing systems modeled with Petri nets, without constructing their RG. Experimental results on large-scale models with hundreds of millions of reachable markings show that the proposed method successfully detects partial deadlocks, whereas RG-based methods fail due to memory overflow, demonstrating its scalability and practical advantages.

Keywords: Intelligent Transportation Systems, Connected and Automated Vehicles, AI and Applications
Abstract: This study explores the traffic management optimization problem in the critical segment of highway with an on-ramp by using deep reinforcement learning. Leveraging the traffic simulator and the Actor-Critic framework with the proximal policy optimization algorithm, we design an agent capable of regulating the maximum speed on the highway and the traffic light status on the ramp. The main objective is to avoid collisions, minimize traffic congestion on the ramp, and abrupt decelerations while maximizing traffic flow. To deepen the analysis and make the study more representative of future mobility challenges, two different scenarios are outlined and compared. The first scenario simulates traffic composed exclusively of traditional vehicles, divided into normal and aggressive (drivers who do not respect road safety rules). The second scenario introduces additional autonomous driving vehicles. Through comparison of the results obtained from these two contexts, it demonstrates how different compositions of traffic vehicles influence adaptive control strategies and overall system performance.

Keywords: Discrete Event Systems, AI and Applications
Abstract: This paper investigates the problem of opacity enforcement in discrete event systems under limited and cost-sensitive observation capabilities. Unlike existing reinforcement learning (RL)-based approaches that assume a fixed observation structure, we formulate a dynamic scenario where the supervisor can adaptively adjust its observation capability among multiple levels, each associated with different sensing costs. In addition, switching between observation levels incurs excessive reconfiguration penalties. To address this problem, we propose a unified RL-based framework that jointly optimizes both control action and observation level selection. A cost-aware reward function is designed to guarantee opacity enforcement, while minimizing overall sensing and switching cost. Experimental results demonstrate that the proposed method can effectively maintain opacity while significantly reducing sensing cost compared to the existing approaches. Moreover, the learned policy achieves a favorable trade-off between performance and resource usage by adaptively switching observation levels only when necessary.

Keywords: Discrete Event Systems, Intelligent Transportation Systems, Decision Support Systems
Abstract: This paper presents a framework for designing on-demand Urban Air Mobility (UAM) networks, from demand estimation and vertiport siting to operational evaluation. Demand is built from commuter and passenger activity data using eligibility filters and converted into spatial demand points. Vertiport candidates are obtained by clustering and screened to satisfy range and minimum station-spacing constraints. Network performance is evaluated using a discrete-event simulation (DES) with on-demand dispatch and deadhead relocation. The DES is coupled with a point-mass eVTOL performance model that uses an energy-efficient cruise altitude and accounts for fixed service times. Day-to-day demand variability is captured through Monte Carlo origin--destination generation with randomized initial vehicle placement. Candidate networks are ranked using a weighted criterion based on completion time, service regularity, deadhead burden, and battery swaps. A case study is carried out on the greater Los Angeles region and illustrates station-fleet trade-off. A slack-based analysis quantifies the non-flight time budgets needed to meet door-to-door travel-time savings targets.

Keywords: System Modeling and Control, Systems Safety and Security
Abstract: Cellular communication networks must remain operational during disasters and large-scale failures; to address these, emergency situations, many existing contingency management approaches rely on conceptual or architectural models that lack rigorous verification of timing behavior, concurrency, and system correctness. Although such models provide valuable design insight, they are often insufficient to guarantee dependable operation under dynamic runtime conditions. Building on the DERIVE (Digital Engineering for Resilience with Integration of Verifiable Evidence) framework for contingency planning in cellular networks using Object-Process Methodology (OPM), this paper integrates model-based systems engineering with formal verification. A rule-based transformation pipeline is proposed to translate an OPCloud-based OPM conceptual model into a verifiable Timed Automata (TA) model in UPPAAL. Structural and behavioral elements of the OPCloud System Diagram (including objects, processes, states, attributes, and links) are systematically mapped to UPPAAL templates, locations, synchronization channels, guards, update functions, and timing constraints through structured transformation rules. The resulting model is verified against key behavioral and temporal properties, including deadlock freedom, reachability of contingency detection states, ensured recovery to functioning operation, service continuity under adverse conditions, and correct coordination among concurrent components. The results demonstrate that the transformed TA model preserves the semantics of the original OPM design while enabling rigorous verification of timing-dependent and concurrent behaviors. This work establishes a foundation for integrating conceptual modeling with formal verification and for future automated OPM-to-TA transformation and scalable verification of resilient cellular network architectures.

Keywords: Intelligent Power and Energy Systems, AI and Applications, Energy Systems
Abstract: リチウムイオンバッテリーの正確な健康状態(SOH)推定は、ますます要求の高いエネルギー貯蔵用途において安全性と信頼性を確保するために不可欠です。物理学に配慮したニューラルネットワーク(PINN)は、劣化メカニズムをディープラーニングと統合しますが、従来のアーキテクチャではデータ忠実度と物理的一貫性の損失との間に勾配の衝突が生じることが多いです。この競争は最適化を不安定化させ、LSTM、GRU、Mambaなどの最先端(SOTA)データ駆動モデルに比べてPINNのパフォーマンスを低下させる可能性があります。これらの課題に対処するため、私たちはPHOENIX(Physics-Oriented Expert Network for Intelligent Execution)というフレームワークを提案します。これはPINNとMixture-of-Experts(MoE)の強みを組

Keywords: Information Systems for Design and Marketing, Machine Learning, Application of Artificial Intelligence
Abstract: Counterfactual explanation (CE) is widely used to enhance the interpretability of machine learning models and support data-driven decision-making based on model predictions. However, existing CE methods typically require two exogenously specified inputs: a desired output value (target) and a distance function that quantifies changes in explanatory variables. In regression settings, neither the validity of target specification nor the practical interpretation of the distance metric has been sufficiently addressed. Furthermore, most existing CE methods focus on altering predictions rather than optimizing a decision objective, even though real-world decision-making often requires explicit objective maximization. To address these limitations, we formulate CE as a profit maximization problem in management and marketing contexts and propose a framework termed profit-based counterfactual explanation (PBCE). PBCE eliminates the need for exogenous target specification by directly maximizing profit as the primary optimization objective. Concurrently, the distance term is reinterpreted as the cost of modifying product attributes, providing a clear and economically grounded interpretation.

Keywords: AIoT, Application of Artificial Intelligence, Intelligent Transportation Systems
Abstract: The rapid advancement of artificial intelligence (AI) and computer vision technologies has created new opportunities to improve road safety, particularly in mitigating wildlife-vehicle collisions. This study proposes YOLO-RAD (Road Animal Detection), an enhanced real-time thermal-based wildlife detection framework built upon the YOLOv12 architecture. The proposed model integrates several attention-based modules, including Cross Scale Multi Attention (CSMA) for cross-scale feature interaction, Multi-head Window Attention (MWA) for local spatial dependency modeling, Adaptive Channel Feature Attention (ACFA) for enhancing appearance-aware feature representation, and Large Kernel Pyramid Attention (LKPA) to capture broader contextual information in thermal imagery. These modules collectively improve the model’s ability to detect animals under challenging conditions such as low illumination, complex backgrounds, and partial occlusions. Extensive experiments were conducted using a thermal wildlife dataset collected from roadway environments as well as the Arribada dataset to evaluate detection performance and generalization capability. The results show that YOLO-RAD achieves 98.23% of precision, 96.18% of recall, 97.52% of F1-score, and 96.82% of mAP50, outperforming several advanced YOLO-series models while maintaining real-time inference speed of 82 FPS. These results demonstrate that YOLO-RAD provides an effective and practical solution for real-time wildlife detection systems, contributing to improve road safety and wildlife conservation.

Keywords: AI and Applications, Generative AI, Optimization and Self-Organization Approaches
Abstract: This paper investigates the use of Large Language Models (LLMs) for solving mathematical programs. Modern optimization techniques are incredibly powerful in applications ranging from engineering, science, business, and everyday life. However, until recently, utilizing modern optimization frameworks required significant technical knowledge, and this has limited their impact. The development of LLMs raises the possibility of performing the translation of real world problems into the language of mathematical optimization even if the user is unfamiliar with optimization as a technical area. To leverage this possibility, we explored the ability of various LLMs to solve linear programs (LPs) and integer linear programs (ILPs)both when phrased technically and non-technically. Our results show that non-technical prompts are up to 5.13% less accurate across studied LLMs and problems, implying that technical background knowledge is useful even when prompting LLMs.

Keywords: Reinforcement Learning, Deep Learning, Human Factors
Abstract: In the Reinforcement Learning from Human Feedback (RLHF) pipeline, traditional single-scalar reward models are often criticized for their susceptibility to reward hacking and lack of interpretability. Multi-objective reward modeling has emerged as a robust framework to address the limitations of single-scalar reward models. However, multiple objective reward modeling demands fine-grained, attribute-specific labels, making the annotation process substantially resource-intensive. In this paper, we propose the method of Implicit Objective Extraction (IOE), an annotation-free multiple-objective-inspired reward modeling approach that discovers implicit features across decorrelated latent dimensions leveraging only standard preference data. Our method preserves the strengths exhibited by multi-objective reward models and accomplishes comparable performance to corresponding multi- objective models on RewardBench.

Keywords: AI and Applications, Machine Learning, Generative AI
Abstract: To investigate which linguistic phenomena machine translation stumbles upon, we examine a set of 29 text-level features motivated by prior translation studies and psycholinguistics, and analyze their correlation with machine translation quality on the WMT benchmark. Our analysis identifies a subset of features that are strongly associated with translation performance. Based on these features, we construct a difficulty metric and use it to perform difficulty-aware sentence sampling from the WMT dataset. Experimental results demonstrate that such difficulty-based sampling consistently benefits LLM-based translation. Furthermore, we incorporate explicitly identified translation challenges into prompt design and propose difficulty-aware prompting frameworks. Our experiments reveal that difficulty-aware prompting achieves translation quality comparable to that of reasoning-augmented models, while significantly reducing model response time. This suggests that explicitly guiding models to focus on translation difficulty can offer an efficient alternative to more computationally expensive reasoning strategies.

Keywords: Knowledge Acquisition, AI and Applications, Neural Networks and their Applications
Abstract: We empirically study the effectiveness of sets of Boolean function metrics to estimate the size of the minimum circuit capable of computing any given Boolean function. This prediction problem is essential to addressing paramount open questions in computational complexity such as P vs NC or P vs NP. In a previous work we observed that a simple neural network reached a high accuracy (over 97%) in the task of identifying the 5-to-1 Boolean functions that require the smallest circuits to be computed. In this paper we empirically study the accuracy reached in the task of minimum circuit size prediction by combining, in several ways including a neural network, up to 11 Boolean function complexity metrics. A larger and more fine-grained dataset, optimal5, is used in this case. The goal of this new approach is to strive towards better explainability: by inputting into the neural network some known and understandable metrics rather than the raw truth table itself, we investigate to what extent a combination of well-known and novel human-designed interpretations of the problem (i.e., the metrics) can accurately predict when, as well as simultaneously explain why, a Boolean function can be computed with small circuits. In addition, information-based techniques are used to analyze the performance of the metrics and their combinations.

Keywords: Representation Learning, Deep Learning, Application of Artificial Intelligence
Abstract: Parking spot classification is a fundamental task in intelligent transportation systems, yet most deep learning approaches rely on large amounts of annotated data and exhibit limited generalization across heterogeneous environments. To address these limitations, we investigate a self-taught learning framework based on unsupervised representation learning with convolutional autoencoders. The proposed approach learns transferable visual representations from unlabeled data and reuses the learned encoders as fixed feature extractors for supervised classification with limited annotated samples in the target domain. To further enhance robustness and mitigate architectural bias, an ensemble of heterogeneous autoencoders is employed, with independent classifier heads and prediction fusion at inference time. Experiments conducted on the PKLot and CNRPark benchmarks under cross-dataset evaluation protocols show that the proposed ensemble-based strategy substantially reduces annotation requirements while improving robustness under significant domain shifts, achieving accuracies between 93% and 96% in data-constrained scenarios.

Keywords: AI and Applications, Deep Learning, Application of Artificial Intelligence
Abstract: In today's information-overloaded society, recommendation systems are indispensable for effectively discovering relevant information from massive datasets. Probabilistic Matrix Factorization (PMF) is a highly successful collaborative filtering technique, but it suffers from data sparsity, which is often mitigated by incorporating auxiliary information. This paper proposes a novel method named MEF-PMF (Multimodal Embedding Fusion for PMF) that comprehensively integrates non-linear embedding vectors extracted from three distinct modalities—text, images, and social networks—into the PMF framework. While state-of-the-art models like TSIMF utilize multi-modal data, they rely on simple linear combinations for social network features, limiting their ability to capture complex user dynamics. To address this, MEF-PMF leverages Convolutional Neural Networks (CNNs) to extract contextual features from text and contour features from images, and node2vec to generate non-linear graph embeddings from social networks. Extensive experiments conducted on Amazon (Clothes and Accessories) and Rakuten datasets demonstrate that MEF-PMF consistently outperforms existing state-of-the-art models. Notably, our evaluation reveals that the accuracy improvement is particularly pronounced for users with extensive social connectivity, underscoring the effectiveness of neural-network-based social relationship modeling.

Keywords: AI and Applications, Deep Learning, Wearable Computing
Abstract: Privacy-preserving Human Activity Recognition (HAR) is vital for sensitive environments like home care. While LiDAR and IMUs provide non-intrusive alternatives to cameras, fusing their heterogeneous spatiotemporal characteristics often causes modality entanglement, where models over-rely on dominant geometric cues. We propose the Tri-Stream Privacy-Preserving Transformer (TPPT), synergizing LiDAR, IMU, and skeletal data via a novel Hybrid Early-Late Fusion strategy. By introducing Factorized Spatiotemporal Encoding, TPPT decouples intra-frame structural analysis from inter-frame temporal modeling to reduce computational complexity. Experiments on the LIPD-Babel-v2 and HMPEAR datasets demonstrate TPPT outperforms state-of-the-art multi-modal baselines. Crucially, TPPT features dynamic ``modality switching,'' autonomously up-weighting inertial features during geometric occlusion to maintain robustness even under critical LiDAR sparsity.

Keywords: Generative AI
Abstract: Artificial intelligence (AI) technology is transforming the way financial market participants obtain information, the process of expectation formation, and the transaction execution mechanism. Based on the combination of conceptual framework and empirical evidence, this article builds an analytical system, with a focus on the changes in investor behavior driven by algorithmic systems, including issues such as information convergence, data distortion, and algorithmic collusion. In response to the reviewers’ comments, this article positions it as a comprehensive study that integrates conceptual analysis and is supplemented by empirical tests, rather than a systematic literature review. Based on the daily data of GameStop (GME), AMC, SPY, and VIX from April 2021 to April 2026, this article examines the volume-volatility transmission, trading volume peaks, market structure before and after GenAI, and cross-market contagion during stress periods. The results show that the trading volume of GME has a Granger predictive effect on absolute returns (lagging by 1 period: F = 34.13, p < 0.001); The absolute return on the peak trading day is significantly higher. The cross-asset correlation rises during the VIX stress period. The above results only indicate statistical evidence consistent with the AI amplification mechanism and cannot constitute proof of the direct cause of AI. Finally, this article puts forward governance suggestions based on result-oriented supervision, algorithm auditing, platform responsibility, and cross-border collaboration.

Keywords: Sustainable Systems, AI and Applications, Decision Support Systems
Abstract: Sustainable AI is advocated as one of the great challenges of our times, however it is not a straight-forward task. Intertwined goals should be accomplished leveraging the efficient use of resources with the reduction of operational software cost as well as environmental emissions and waste. It is clear that in order for such efforts to be successful, besides environmental sustainability, economical sustainability of software systems should also be considered. In this work, we explore an approach to enable the computation of both environmental and economic sustainability of alternative deployments of inference engines. For this purpose, a emph{Four-stage approach to create a Sustainability measurement}, combining both factors in a single function, is introduced. Furthermore, we present the emph{Energy Consumption Measurement Testbed}, a standalone service that measures inference energy consumption in real time via CodeCarbon for alternative models and deployment setups, as a tool to explore environmental sustainability of inference engines. As a proof of concept, the proposed approach and tools were used to explore the sustainability metrics for alternative inference models, such as CNN and MLP, in alternative configurations (different precision types, 3 different providers, different IaaS instance types). Conclusions drawn on sustainable deployments of inference models is also discussed, as a first step towards employing viable solution for sustainable deployment of AI models.

Keywords: Representation Learning, Trust in Autonomous Systems, Neural Networks and their Applications
Abstract: Traceability is a fundamental requirement for explanation, diagnosis, and trust in intelligent systems, particularly in safety-critical and human-in-the-loop settings. However, conventional neural representations entangle input contributions into opaque activations, making it difficult to recover the input contributions underlying a decision. This paper introduces the Prime-Weighted Exponential Neuron (PWEN), a neural unit that encodes inputs multiplicatively using exponentiation with distinct prime-based weights. Under discrete input assumptions, this representation forms a uniquely factorized code, enabling exact and data-independent recovery of input contributions via prime decomposition. As a result, PWEN provides intrinsic traceability and neuron-level invertibility, where explanations arise directly from the computational structure rather than through post hoc analysis. We show that PWEN can represent fundamental logical operations, including AND, OR, and XOR, establishing it as a symbolically compatible computational primitive. We further demonstrate that PWEN neurons can be incorporated into a simplified sequential decision-making task through a grid-based foraging experiment. In this setting, PWEN-based controllers exhibit (i) behavioral equivalence to rule-based and logic-based systems, (ii) competent goal-directed behavior, and (iii) evolvability through genetic optimization. By tightly coupling encoding, computation, and decoding within a single representation, PWEN offers a mathematically grounded approach to traceable neural computation, supporting the design of transparent and auditable intelligent systems.

Keywords: Application of Artificial Intelligence, Machine Learning, Intelligent Internet Systems
Abstract: Decentralized finance (DeFi) protocols are vulnerable to high-frequency exploits leveraging flash loans, primarily due to their reliance on static rule-based defense mechanisms that lack adaptability to evolving adversarial strategies. This paper proposes the Hybrid Detection and Trust Throttling (HDTT) framework. HDTT combines post-execution off-chain ML scoring with on-chain reputation state management, enabling asynchronous trust-based access control. Methodologically, HDTT uses out-of-fold stacking with Random Forest and XGBoost base learners aggregated via L1-regularized Logistic Regression. The L1 penalty naturally provides feature selection, reducing weights on signals that degrade under adversarial perturbation. To align with real-world deployment constraints, we implement a precision-capped thresholding strategy prioritizing low false-positive rates. We evaluate HDTT on 674 real-world exploits and Ethereum transactions (2021–2025) under rigorous adversarial evasion settings, including gradient-based (e.g., PGD/FGSM) and black-box feature manipulations bounded by an adversarial budget. Rather than relying solely on gas metrics, HDTT extracts 54-dimensional semantic features. Relying on this multi-layer feature architecture, HDTT achieves 0.924 precision, 2.4% FPR, and 0.957 F1-score on 10,000 real-world transactions. Crucially, it exhibits graceful degradation against severe adversarial attacks, consistently outperforming standalone ML models and static/adaptive baselines. With a post-execution inference latency of 1.2 ms, HDTT demonstrates the feasibility of lightweight, robust asynchronous trust scoring. The dataset and code are available at textcolor{black}{url{https://github.com/x0x0z/HDTT}}.

Keywords: Autonomous Vehicle, AI and Applications
Abstract: This paper proposes an automated defect detection and classification system for analyzing the data from UAVs during bridge inspections. Our proposed system employs 5 state-of-the-art deep convolution neural network (CNN) architectures as baseline models, a CNN-based meta-classifier or pooling function, and a stacking-ensemble learning technique to increase the accuracy of defect detection and classification. Our experimental results reveal the effectiveness of the proposed approach, with the stacking ensemble learning model achieving the highest performance in terms of accuracy in detecting and classifying common concrete bridge surface defects, including spalling, cracking, and efflorescence compared to the individual baseline models. The success of our approach highlights the robustness and performance-boosting capabilities of deep ensemble learning for automated defect detection and classification in UAV-assisted bridge inspections. Overall, this approach can significantly improve the efficiency and accuracy of bridge inspections with enhanced and reproducible deliverables.

Keywords: Autonomous Vehicle, Deep Learning, Intelligent Transportation Systems
Abstract: End-to-end autonomous driving provides a unified perception-to-control framework, but signalized intersections remain challenging because traffic lights are few-pixel cues that directly determine stop/go behavior. This challenge is amplified by class imbalance and appearance shifts caused by illumination and weather, allowing small perception errors to propagate into unsafe closed-loop actions. We propose TLCAM (Traffic-Light Classification Auxiliary Module), a capacity-controlled auxiliary branch that enforces traffic-light cues in the shared encoder and provides the controller with a compact semantic embedding, while incurring only 0.45% more parameters and 0.02 ms of latency. TLCAM is trained jointly with the original control and trajectory losses plus a traffic-light classification loss, encouraging signal-aware representations without materially changing the base architecture. In MORAI urban driving, TLCAM achieves 96.79% traffic-light accuracy on observed frames and improves red-light compliance: whole-vehicle stop success rises from 90.22% to 95.38%, strict no-crossing success from 73.64% to 79.53%, and stopping-distance variance is reduced by 53%. In an out-of-distribution evaluation, where models are trained only on “Sunny + Noon” and tested across night and adverse weather conditions, TLCAM consistently improves stop–restart performance. These results show that targeted supervision of sparse, safety-critical cues can improve rule compliance and robustness in end-to-end driving with negligible computational overhead.

Keywords: Intelligent Transportation Systems, Autonomous Vehicle, Machine Learning
Abstract: Deceptive Global Navigation Satellite System introduces structured measurement biases that evade immediate detection, creating a vulnerable interval between the true spoofing onset and the eventual alarm. We address this by formulating onset localization as a change-point problem within the residual sequences of nonlinear state estimators. Accounting for estimator-induced temporal correlations, we derive a detectability bound based on a residual energy metric and explain why ramp-style attacks are inherently harder to localize than step biases. From this formulation, we derive a correlated generalized likelihood ratio (C-GLR) statistic for spoofing onset localization and develop a data-driven neural localizer. Evaluation on KITTI-derived trajectories shows that the C-GLR reduces localization error over standard cumulative sum baselines, while the neural approach offers the highest robustness across varying spoofing profiles.

Keywords: Autonomous Vehicle, Intelligent Transportation Systems, Deep Learning
Abstract: Predicting the future motion of vehicles is essential for autonomous driving systems. However, long-term prediction remains challenging because intention-related motion tendencies and interaction-driven short-term variations are entangled in the observed motion. Existing intention-aware models typically encode motion history into a single latent representation, which can entangle these factors and weaken maneuver-consistent long-term prediction. To address this issue, we propose DIAT, a Decomposition-based Intention-Aware Transformer for maneuver-conditional vehicle trajectory prediction. DIAT introduces a Decomposed Transformer Motion Encoder (DTME) that applies a causal exponential moving average (EMA)-based decomposition to separate each observed motion sequence into trend and residual components before temporal encoding. The trend component is intended to capture intention-related motion tendencies, while the residual component retains short-term variations influenced by social interaction. We additionally incorporate target-relative motion embeddings into spatial attention to improve interaction modeling between the target vehicle and its neighbors. Experiments on the NGSIM highway dataset demonstrate that DIAT reduces trajectory prediction RMSE across all prediction horizons and all evaluated maneuver classes compared with prior intention-aware baselines, while achieving higher balanced accuracy on less dominant maneuver classes. Ablation studies further confirm that both decomposition-based motion encoding and relative-motion-aware interaction modeling contribute to these improvements.

Keywords: Intelligent Transportation Systems, Smart Buildings, Smart Cities and Infrastructures, Machine Learning
Abstract: As urban areas expand, automatic monitoring of parking lots becomes essential for efficient and sustainable cities. This work proposes a self-supervised approach for parking spot occupancy recognition that requires no labeled samples from the target parking lot. Building upon a self-supervised transfer learning fine-tuning protocol, the proposed training strategy consists of two self-supervised stages: first on unlabeled generic data and then on unlabeled target-specific data, followed by supervised fine-tuning using only generic parking lot labels. We adopt SimCLR with a ResNet-50 encoder and evaluate the method under a leave-one-out cross-environment protocol on three public datasets: PKLot, CNRPark-EXT, and PLds. We also introduce a two-stage deployment strategy in which a Strong General Model is initially deployed, followed by a Specialized Model that incorporates unlabeled images collected during the first N days of deployment in a self-supervised manner. Experimental results show that the Strong General Model alone outperforms supervised and self-supervised baselines, achieving an average accuracy of 97.2%, which further improves to 97.8% with the proposed two-stage strategy. These results demonstrate that self-supervised learning enables a scalable and label-efficient solution for real-world parking occupancy monitoring. Our trained models and source code are publicly available at https://github.com/DSBD-Research/Parking-Spot-Occupancy-Recognition.

Keywords: Connected and Automated Vehicles, Autonomous Vehicle, Intelligent Transportation Systems
Abstract: Commercial 3D sensors and vehicle-to-everything (V2X) modules in automated driving typically provide sparse, post-processed object lists with explicit state information, rather than dense intermediate features, due to standardized output formats and data privacy constraints. This discrepancy introduces spatial, temporal, and semantic misalignment when performing late fusion with mainstream vision-centric 3D multi-object tracking (MOT) pipelines. To address this issue, we propose CLF-StreamTrack, an end-to-end cross-level-fusion-based (CLF) 3D MOT framework that fuses structured object-list priors with dense multi-view visual features, and further enhances track representation learning via a stream-based memory-driven temporal aggregation mechanism, enabling effective spatiotemporally consistent tracking. Extensive experiments on the nuScenes benchmark, using both real and synthetic object lists, demonstrate that the proposed approach consistently improves tracking performance over state-of-the-art vision-centric trackers and the existing CLF-tracking baseline, validating the effectiveness of integrating object-centric priors with temporal modeling.

Keywords: Computational Intelligence, Decision Support Systems, Transfer Learning
Abstract: Skin lesion classification is a critical task in medical image analysis, particularly for early detection of melanoma. Convolutional Neural Networks (CNNs) are effective in capturing local features, while transformer-based architectures provide strong global contextual modelling. However, standalone models often fail to jointly exploit both characteristics. In this paper, a hybrid CNN--Transformer architecture enhanced with multiple attention mechanisms is proposed. Specifically, Squeeze-and-Excitation (SE), Efficient Channel Attention (ECA), Convolutional Block Attention Module (CBAM), Coordinate Attention (CoordAtt) and self-attention are integrated individually into the hybrid framework to analyse their impact on discriminative feature learning. The proposed models are evaluated on both HAM10000 and ISIC 2017 datasets. The empirical results demonstrate that attention-enhanced hybrid models consistently outperform baseline CNN and transformer architectures. The proposed model integrated with SE and CBAM achieves the highest performance on HAM10000, while self-attention-based counterparts show stronger generalisation on ISIC 2017. These findings highlight the importance of combining channel, spatial, and global attention mechanisms for robust skin lesion classification.

Keywords: Optimization Control, Decision Support Systems, Supervisory Control
Abstract: This paper introduces a novel approach to the General Assembly Line Balancing Problem (GALBP) by utilizing a receding horizon optimal control framework. The proposed discrete model for the assembly line offers a flexible representation, avoiding assumptions about specific line configurations. The control actions aim to optimize the industrial assembly line by minimizing the completion time while adhering to constraints such as task precedence, workstation capacity, and resource requirements. Control actions are represented through task assignment and resource allocation matrices, assigning tasks to specific workstations and assigning resources to workstations, respectively. The optimization problem is formulated as a MixedInteger Nonlinear Programming (MINLP) problem. The inherent robustness of the receding horizon approach ensures optimal solutions for the assembly line, effectively adapting to sudden changes. Numerical experiments demonstrate the robustness and effectiveness of the proposed control synthesis in efficiently distributing tasks and resources, minimizing the overall completion time.

Keywords: Human-Machine Cooperation and Systems, Decision Support Systems, Intelligence Interaction
Abstract: Human-AI teaming (HAT) increasingly involves AI systems that provide real-time, context-aware guidance in complex tasks. While such systems can improve performance, their effectiveness depends on how they shape human cognition and behavior. In particular, AI assistance can introduce cognitive demands and influence attention, planning, and interaction with the task environment, with effects that can vary across levels of expertise. This work investigates these mechanisms in a simulated search and rescue (SAR) environment. We compare human performance under two LLM (Large Language Model)-guided conditions and a no-LLM baseline, and analyze interaction at multiple levels, including task performance, eye-tracking measures, and planning behavior. Eye tracking provides fine-grained insight into attention allocation and interaction with AI guidance, while behavioral measures capture how users structure and adapt their decisions over time. Results indicate that LLM guidance enhanced task efficiency (higher rewards and victims-per-step) but did not increase total victims saved. Eye-tracking data revealed an attention-guidance trade-off, with visual resources shifting to the chat interface alongside increased pupil size variability. Expertise moderated this effect: novices exhibited passive AI reliance, whereas experts maintained a "verification loop" through persistent environmental scanning. These findings suggest that LLM-mediated teaming efficacy depends on the operator’s ability to cross-reference AI guidance with ground truth to maintain situational awareness.

Keywords: Computational Intelligence, Metaheuristic Algorithms, Decision Support Systems
Abstract: While many metaheuristics perform well in complex search spaces, their reliance on opaque stochastic perturbations often makes the search process difficult to interpret and reproduce. To bridge this gap, this study proposes the holistic weighted jaguar algorithm (HW-JA), the first extension of the jaguar algorithm (JA) framework to real-world constrained portfolio optimization. Rather than relying on stochastic processes, HW-JA updates candidate solutions through deterministic movement rules informed by fitness feedback and local search-space structure. This design yields a systematic yet adaptive search process, in which a dynamic hunting operator and cyclic pairwise updates explicitly account for inter-asset dependencies. In addition, a rank-based neighborhood search mechanism provides a clear vector-based rationale for each search trajectory. Experimental results on financial benchmarks demonstrate that HW-JA achieves competitive performance relative to state-of-the-art metaheuristics. More importantly, the structured search decision rationales produced by HW-JA support reproducible and traceable search behavior, providing an auditable basis for explainable evolutionary optimization, thereby addressing a critical requirement in high-stakes financial decision-making.

Keywords: Deep Learning, Machine Learning, Generative AI
Abstract: Deepfake image detectors achieve high accuracy on standard benchmarks, yet their reliability collapses under adversarial perturbations that are visually imperceptible but optimized to shift a detector’s decision boundary. Recent defenses incorporate explainable AI (XAI) attribution maps as auxiliary inputs for attack detection; however, when explanations are generated from the same compromised detector, an adaptive perturbation can corrupt both the prediction and its explanation. We propose Cross-Model XAI Divergence (CMXD), a detector-agnostic framework that treats disagreement between explanation maps from two architecturally distinct deepfake detectors as an explicit signal of adversarial manipulation. Given a face image, CMXD computes attribution maps from XceptionNet and EfficientNet-B4ST, derives complementary cross-model features via the absolute difference and element-wise product of the maps, and fuses these with the RGB image into a compact five-channel representation. A lightweight ResNet-50 classifier trained on this representation distinguishes clean inputs from adversarially perturbed deepfake images. Experiments on FaceForensics++ (Face2Face) with PGD training and evaluation across unseen white-box and black-box attacks demonstrate that cross-model explanation signals substantially improve attack detection compared to single-detector baselines, while preserving practical inference cost through offline attribution precomputation.

Keywords: Machine Learning, Deep Learning, Quality and Reliability Engineering
Abstract: Abstract— Existing methods for testing deep neural networks (DNNs) primarily prioritize test inputs likely to reveal model faults under a fixed labeling budget. In practice, choosing that budget is difficult: too little testing misses failures, while too much incurs unnecessary labeling costs. This work studies the stopping problem in DNN testing. We formulate testing as a cost–benefit decision process in which labeling an input incurs cost c and discovering a fault yields value v. Based on this formulation, we introduce AdaStop, a framework that estimates the marginal fault discovery rate during testing and stops labeling when the estimated rate falls below the threshold τ = c/v. Experiments across multiple datasets, architectures, and selection strategies show that 65–84% of faults can be discovered using only 9–31% of the labeling budget.

Keywords: Deep Learning, Cognitive Computing, Transfer Learning
Abstract: Autism Spectrum Disorder (ASD) is associated with atypical gaze behavior during social perception. To our knowledge, we present the first multimodal framework that combines scanpath images with statistical captions (i.e., textual summaries of gaze features) for ASD classification during point-light walker (PLW) biological motion viewing. Each image--caption pair is embedded using a vision--language transformer and classified via a lightweight MLP. To ensure rigor, we benchmark against a non-textualized baseline in which standard classifiers receive the same underlying gaze-derived feature family, including all quantities used across the caption variants, as raw numeric vectors rather than captions; we also detail data quality-control procedures for calibration error, off-screen fixations, and blink handling. Among five caption variants tested, those encoding both spatial and temporal descriptors achieved the best performance, reaching 95.6% fold-averaged accuracy and a macro F1 score of 0.957 under leave-one-subject-out cross-validation. Unimodal ablations show that scanpaths are more informative for ASD classification, whereas captions better capture neurotypical patterns, underscoring their complementary strengths. To interpret model behavior, we apply structured perturbations: occluding core body regions (head, torso, legs) substantially reduces image-only performance, while caption perturbations that disrupt temporal or semantic structure likewise degrade classification. This work provides new insight into how multimodal gaze representations reflect attentional differences in ASD, and, to our knowledge, is the first to unify gaze-derived language and vision within a single model for biomotion-based analysis.

Keywords: Machine Learning, Deep Learning, Computational Intelligence
Abstract: Siamese verification networks are widely used to compare items such as faces, cars, or signatures. In these scenarios, the network is trained to learn an embedding space in which similar objects are mapped closer together, while dissimilar objects are mapped further apart. Two objects are considered to belong to the same class (e.g., the same person in two different images) when the distance between their embeddings falls below a predefined threshold. Defining this threshold, however, is a non-trivial task and typically requires labeled data. In this work, we assume that the distribution of distances produced by a siamese verification network can be approximated by a bimodal function. Based on this assumption, we propose an unsupervised method to determine the verification threshold by identifying the minimum point between the two modes. The proposed approach does not require annotated samples, enabling the verification threshold to be updated directly in the deployment environment without the cost of manual labeling. We evaluate our method on four datasets: MNIST, CIFAR-10, LFW, and PKLot. The results indicate that the proposed approach achieves an average verification accuracy of 94%, comparable to the EER method, while eliminating the need for labeled data.

Keywords: Deep Learning, Generative AI, Machine Vision
Abstract: Deep learning-based methods for medical image analysis are often limited by the availability of large, well-annotated datasets, particularly in specialized domains such as rheumatoid arthritis (RA), where expert labeling is costly and time-consuming. In this work, we present a data-efficient pipeline for the augmentation of hand radiograph datasets that combines active learning-based landmark annotation with conditional diffusion models. We first introduce an iterative active learning framework to annotate postero-anterior (PA) hand radiographs with a structured representation of anatomy using anatomical landmarks. We use these annotations as structured conditioning signals in a ControlNet-based diffusion framework to generate anatomically realistic hand radiographs. The model is trained in a two-stage process: first, learning anatomical priors from a large dataset of healthy hands, followed by fine-tuning on a small RA dataset with expert-provided disease staging. By conditioning on landmark configurations that encode joint spacing and alignment, the proposed method enables controlled synthesis of radiographs exhibiting varying degrees of pathology.

Our results demonstrate that the integration of anatomical priors with diffusion-based generation produces realistic and clinically consistent images, enabling dataset augmentation and improving the accuracy of downstream tasks. This approach provides a way to incorporate geometric structure into generative models and highlights the potential of combining active learning and conditional diffusion for medical imaging applications.

Keywords: Deep Learning, Neural Networks and their Applications, Big Data Computing
Abstract: Transformer-based language models have become powerful tools in natural language processing tasks with strong performance. While training language models from scratch is time-consuming, progressive training is a popular approach to speed up pre-training of transformer models. This method trains a shallow transformer first and expands the model by reusing pre-trained weights. Existing progressive training approaches take advantage of higher convergence speed, but do not consider that different sized models exhibit different learning capacities, which limits the potential of progressive training methods. To realise this potential, we propose an efficient progressive pre-training framework which aligns model growth with dataset complexity. Our approach begins by training a shallow transformer on a small subset of low-complexity samples. In later stages, the depth of the model and the number of the training samples are increased concurrently. An effective curriculum for the training samples is constructed by a training subset selection approach based on the cluster density of the samples. Compared to the state-of-the-art progressive training approaches, our approach saves 74.1% of the pre-training time to achieve a validation loss of 2.0, when training a 16-layer RoBERTa model on the BookCorpus dataset. Experimental results on the GLUE benchmark also demonstrates a better model generalization using our method.

Keywords: Neural Networks and their Applications, Transfer Learning, Deep Learning
Abstract: Zero-resource Named Entity Recognition (NER) has recently garnered increasing attention due to its independence from labeled data. However, existing methods often oversimplify the task by assuming either an identical set of entity types between source and target domains or the availability of type names and descriptions in the target domain. These assumptions rarely hold in real-world scenarios, where entity definitions are dynamic and continually evolving, thereby limiting the applicability of such methods. To address this challenge, we propose CDZO, a Consistency-constrained Decomposed framework for Zero-resource Open-world named entity recognition. Specifically, CDZO decomposes the NER task into two subtasks: entity boundary identification and entity type classification, thereby reducing the complexity of cross-domain knowledge transfer. To mitigate the boundary deviations across domains, CDZO introduces a boundary smoothing strategy. In addition, CDZO designs a non-entity-aware prototype generation algorithm that leverages non-entity tokens in the source domain to extract latent, type-specific semantic representations. Finally, CDZO proposes a reweighting mechanism and a semantic consistency loss to enforce output consistency across subtasks. Experiments on ten cross-domain transfer settings demonstrate that CDZO consistently outperforms state-of-the-art baselines, achieving significant improvements in performance.

Keywords: Deep Learning, Machine Learning, Generative AI
Abstract: Machine unlearning for text-to-image diffusion models aims to selectively remove undesirable concepts from pre-trained models without costly retraining. Current unlearning methods share a common weakness: erased concepts return when the model is fine-tuned on downstream data, even when that data is entirely unrelated. We adapt Projected Gradient Unlearning (PGU) from classification to the diffusion domain as a post-hoc hardening step. By constructing a Core Gradient Space (CGS) from the retain concept activations and projecting unlearning updates into its orthogonal complement, PGU ensures that subsequent fine-tuning cannot undo the achieved erasure. Applied on top of existing methods (ESD, UCE, Receler), the approach eliminates revival for style concepts and substantially delays it for object concepts, running in roughly 6 minutes versus the ~2 hours required by Meta-Unlearning. PGU and Meta-Unlearning turn out to be complementary: which performs better depends on how the concept is encoded, and retain concept selection should follow visual feature similarity rather than semantic grouping.

Keywords: Transfer Learning, Deep Learning, Machine Learning
Abstract: Transfer learning (TL) has become a cornerstone of deep learning, enabling the adaptation of powerful pretrained models to a wide range of downstream tasks. However, this paradigm introduces a critical security vulnerability that backdoor defects in pretrained models can be inherited by downstream models, that exhibits a severe threat. In this paper, we propose a novel backdoor defense framework, Distribution-Aware Pruning and Alignment (DAPA), for safe transfer learning (STL). DAPA synergistically weakens backdoor-related redundant connections via distribution-aware pruning, while introducing Gaussian Mixture Model (GMM) guided feature distribution alignment during the fine-tuning. This process forces the model's internal features to align with a benign distribution built from clean data. It eliminates the anomalous feature regions on which the backdoor relies. Furthermore, extensive experiments on typical backdoor attacks and benchmark TL datasets are conducted. DAPA significantly shows the low defect inheritance rate, and almost entirely preserves the accuracy on the main task. Its performance surpasses existing cutting-edge methods. In summary, this work provides an efficient and general model purification solution for STL.

Keywords: Biometric Systems and Bioinformatics, Deep Learning, Computational Life Science
Abstract: Protein-protein interactions (PPIs) reveal important biological processes and are involved in a wide range of diseases, including cancer, neurodegenerative disorders, and infectious diseases. Binding affinity, an important measurement of PPIs quantified as the changes in Gibbs free energy (ΔG), is therefore essential. Current computational methods rely on protein structural data, which is often unavailable due to experimental limitations. It constrains large-scale affinity prediction and slows the discovery of potential therapeutics. We present AffinityBERT, a deep learning framework that predicts ΔG directly from protein sequences. Our approach evaluates six model variants that combine custom and ProteinBERT embeddings with convolutional neural network (CNN), transformer, and graph attention network (GAT) architectures. On a curated sequence-ΔG dataset, the transformer-based AffinityBERT-PT model achieves a mean absolute error (MAE) of 1.57 kcal/mol and a Pearson correlation of 0.5985. Compared to state-of-the-art (SOTA) methods on a benchmark of 82 protein samples, which is widely utilized in related studies, it delivers an MAE of 1.44 kcal/mol and a Pearson correlation of 0.58. To ensure the feasibility and validate its utility in predicting disease-related effects, we performed a real-world application to SOX17 mutation, a cardiovascular disease-related protein. The main contributions are: (1) we curated a high-quality sequence-ΔG dataset for systematic benchmarking for future broader research; (2) compared custom and ProteinBERT embeddings for ΔG prediction to show whether pre-trained model improves the results; (3) implemented and evaluated CNN, transformer, and GAT architectures to show which model achieves the best results; (4) demonstrated structure-independent prediction of ΔG to allow scalable and efficient PPI affinity estimation without reliance on structural data; and (5) achieved SOTA performance compared to current sequence-based methods. AffinityBERT bypasses the need for protein structures and allows scalable PPI studies and supports a wide range of protein-related biological research, such as drug discovery and design.

Keywords: Digital Twin, Control of Uncertain Systems, System Modeling and Control
Abstract: Reliable digital-twin operation requires update mechanisms that remain effective under sens- ing noise, model mismatch, and decision-to-execution delay. This paper proposes a delay-aware update- trigger synthesis framework combining Conditional Value-at-Risk (CVaR) with a grid-supported Wasser- stein robustness score. A residual-based risk signal is augmented with delay compensation, and candi- date policies are evaluated by post-drift tail risk and executed-update cost over a normalized deployment grid. The framework is evaluated through 21,760 PyBullet runs over 64 labeled cells and 34 pol- icy configurations; the four ρ levels are trajectory- equivalent. When configurations and runs are pooled within each policy family, triggered policies reduce aggregate Lp95 from 0.500 to 0.481, but increase mean executed updates from 74.5 to 502.8. At the candidate level, the non-dominated triggered design reduces the maximum observed post-drift loss from 0.521 to 0.519 relative to period-10 updating while requiring approx- imately 394 additional updates. Sensing noise and trigger selection dominate intervention variability, whereas execution delay has only a modest aggregate effect. No candidate satisfies the service bound over all labeled cells. Because the ambiguity radius enters through a common certification surcharge, it raises the robust score without changing candidate ordering or simulated trajectories. Overall, the marginal tail- risk improvement and substantial intervention cost motivate explicitly cost-aware and more selective up- date mechanisms.

Keywords: Cyber-physical systems, Intelligent Transportation Systems, Human-Centered Transportation
Abstract: This study proposes a cyber-physical data-driven framework for proactive detection of highway traffic hotspots and hot regions. The proposed framework bridges users’ online search records in cyberspace as early indicator. To handle large-scale and irregular search records, we propose an Origin–Destination–Time (ODT) tensor model to represent the spatio-temporal structure of route search data and accelerate computation. Using destination-wise inflow sequences derived from these records, we develop a systematic method to automatically identify anomalous surges that indicate emerging traffic hotspots and further the regions. To validate the framework, we conduct experiments using a one-year real-world dataset covering 2,728 interchange (IC) nodes within a highway network. Furthermore, we integrate and compare search data with actual traffic volumes for evaluation. The results reveal a strong correlation between search intensity and traffic flow, demonstrating that online search behavior serves as a reliable proxy for anticipating traffic dynamics. These findings suggest that route search records in cyberspace can be effectively utilized for proactive traffic monitoring and highlight the potential for early prediction of congestion patterns.

Keywords: Connected and Automated Vehicles, Robotic Systems, Control of Uncertain Systems
Abstract: When autonomous vehicles navigate unsignalized intersections, they face significant decision-making challenges due to multi-directional interactions among road users and inherent uncertainties. Although reinforcement learning (RL) has gained widespread acceptance for addressing such challenges, safety and partial observability remain major concerns. To address these issues, this paper proposes a Mamba-based Selective State-Space Model (Mamba-SSM) integrated with a distributional reinforcement learning network (Mamba-DRLN). The Mamba network provides context-aware representations by selectively updating latent states over input sequences, enabling effective handling of partial observability. Meanwhile, the distributional RL implemented, in particular with a Quantile Regression Deep Q-Network handles uncertainty by approximating the distribution of returns. The proposed framework further incorporates a risk-averse policy based on Conditional Value-at-Risk (CVaR) to select actions that minimize risk and enhance safety. This novel integration provides a robust, uncertainty-aware, and risk-sensitive decision-making policy for autonomous driving at intersections. The proposed method is validated using the high-fidelity CARLA simulator. Experimental results demonstrate that Mamba-DRLN outperforms baseline methods in producing safer and more reliable decisions.

Keywords: Neural Networks and their Applications
Abstract: Non-stationary wind energy systems exhibit complex multi-scale dynamics where critical information is often distributed across varying spectral bands. This paper proposes a Multi-Scale State Decomposition (MSSD) framework utilizing Adaptive Frequency-Domain Conceptors (AFDC) within the Echo State Network (ESN) paradigm. Unlike traditional forecasting methods that treat the reservoir as a monolithic state space, the proposed framework partitions the internal dynamics into frequency-aligned subspaces. By introducing a cybernetic feedback mechanism, the AFDCs dynamically regulate the system's focus, allowing it to autonomously transition between low-frequency trends during steady-state operations and high-frequency transients during volatile environmental regimes (e.g., storm fronts). We provide a rigorous theoretical analysis of the state-space contraction properties under frequency-domain decomposition. Experimental validation on real-world wind energy datasets demonstrates that the adaptive framework not only reduces Mean Squared Error by an order of magnitude compared to standard ESNs but also provides a principled approach for state-space management in non-stationary cybernetic systems.

Keywords: Application of Artificial Intelligence, Adaptive Systems, Human Performance Modeling
Abstract: Deep learning-based knowledge tracing (DLKT) has shown strong promise for adaptive learning, yet existing models are almost exclusively designed for binary responses. This constraint is unrealistic for practical applications given the prevalence of partial credit and graded assessments in real educational settings. In this paper, we extend two established DLKT architectures, the LSTM-based Deep Knowledge Tracing (DKT) model and a Self-Attentive Knowledge Tracing (SAKT) model, to support continuous-valued student responses, replacing the binary output head with a regression formulation trained under MSE loss. Since there are no large-scale public datasets with continuous scoring or partial credit responses, we generate synthetic data using two polytomous Item Response Theory (IRT) models, the Graded Response Model (GRM) and the Generalized Partial Credit Model (GPCM), which differ structurally in how they assign probability mass across response categories. Our experimental results demonstrate that both extended DLKT models successfully learn from continuous response data, substantially outperforming trivial baselines. Furthermore, we find that the choice of generative IRT model meaningfully affects training difficulty with both architectures performing considerably better on GPCM data than GRM data. This work contributes towards establishing a foundation for continuous and partial credit DLKT and highlights the need for real-world datasets with more comprehensive assessment scoring.

Keywords: Generative AI, Cyber Systems and Engineering, Logistic and Manufacturing Automation
Abstract: Manufacturing systems consist of interconnected machines, resources, and processes that harmonically transform raw materials into finished products. The performances of a manufacturing system are often pertinent to the rate of throughputs, the cost of production, the quality of finished products, the timeliness of delivery, the resilience to disruptions, and/or the lifecycle sustainability. However, it is not uncommon that manufacturing systems face unexpected disruptions such as machine failures, material shortages, human labor challenges, privacy breaches, demand variations, and tooling problems. These disruptions not only affect immediate operations for work-in-process (WIP) but also compromise the system’s short- or long-term capability in meeting key performance indicators (KPIs). Consequently, manufacturing resilience refers to the ability to adapt, reconfigure, and then recover system functionality and production performance in response to unexpected disruptions. Traditional methods for addressing these disruptions tend to be time-consuming and resource-intensive, which cannot effectively maintain the resilience of manufacturing systems. Despite recent advances in digital twins (DTs) and artificial intelligence (AI), very little has been done to mitigate high computational demands and generate resilient designs of manufacturing system under uncertainty. Therefore, this talk will present a new Generative AI approach for the on-the-fly design of manufacturing systems in response to production disruptions. First, we propose a novel Generative Adversarial Network (D-GAN) to generate diverse, adaptive system designs that align production performance with target key performance indicators (KPIs). Second, DT models are coupled with statistical metamodeling to sequentially generate large amounts of training data samples under different scenarios. Experimental results show the high potential of the proposed D-GAN approaches to generate cost-effective system designs and enhance manufacturing resilience. We hope this talk can help catalyze more in-depth studies on the novel design of GenAIs for general applicability to a board range of manufacturing scenarios and factory configurations.

Keywords: Cybernetics for Informatics, Cyber Systems and Engineering, Adaptive Systems
Abstract: Large language model systems are a useful substrate for studying cybernetic efficiency: they transform costly inputs into scored outputs, are organized into staged workflows, and are routinely modified by prompts, retrieval, verifiers, and pipelines. Standard cost-benefit comparison is necessary, but it does not say whether a system is operating far below a declared frontier, near its ceiling, or in an intermediate band where constraints may become visible. We present Compression Physics as a regime diagnostic for cybernetic efficiency systems and demonstrate it on LLM inference workflows. The diagnostic maps output-per-cost into a declared frontier-normalized coordi- nate, constructs cost-weighted pairwise merge observations, and compares a reciprocal response model with constant and linear nulls. Applied across model scale and declared workflow stages, the method combines surface efficiency with external response readouts. System-prompt experiments show changes in numbered structure, reference-evidence coverage, and self-check marker incidence at different declared stages. These results support an auditable stagewise diagnostic.

Keywords: AI and Applications, Affective Computing, Artificial Immune Systems
Abstract: Enterprise production operations are increasingly adopting artificial intelligence and agentic automation to accelerate incident triage, root-cause analysis, and recovery. In mission-critical environments, however, automation cannot be evaluated only by speed or diagnostic accuracy. Incident response is a socio-technical control problem in which automated decisions may affect system stability, customer experience, compliance posture, and human accountability. This presentation introduces a trust-aware, human-in-the-loop Agentic AIOps framework that combines telemetry-driven reasoning, portfolio-level risk context, explicit guardrails, and bounded execution to support trustworthy incident response at enterprise scale.

Keywords: Medical Informatics, Computational Life Science
Abstract: Traditional pediatric respiratory management relies heavily on physical auscultation to diagnose common pulmonary conditions. However, human interpretation of adventitious breath sounds suffers from low interrater reliability, highlighting a need for objective, automated screening tools. While digital stethoscope systems offer a promising solution, their development is structurally bottlenecked by data-collection bias, environmental acoustic artifacts, and label uncertainty inherent in existing repositories. This paper details how stationary speech-based features fail to capture transient upper-airway obstructions, misclassifying 10 out of 30 actual croup samples as normal breath sounds and yielding a poor baseline sensitivity of 46.7%. To address these data-layer vulnerabilities prior to algorithm training, we outline a high-level conceptual data-ingestion framework. By exploring real-time hardware-level quality filtering, high-level multidisciplinary data auditing, and edge-computing localization constraints, this research establishes a mathematically efficient and privacy-compliant pathway optimized for low-power intelligent stethoscope implementation at the clinical bedside.

Keywords: AI and Applications, Robotic Systems, Medical Informatics
Abstract: We propose a transformer-based foundation model to analyze robotic surgical videos for comprehensive, multi-task surgical scene understanding. The model simultaneously performs: (1)~surgical task classification (knot tying, needle passing, suturing), (2)~objective skill-level assessment on a standardized 1--5 proficiency scale, and (3)~natural language generation of fine-grained surgical gesture descriptions, all within a unified multi-task learning framework with adaptive uncertainty weighting. Evaluated on the JIGSAWS benchmark using the da~Vinci surgical robot, the proposed framework achieves 100% task classification accuracy, 91.7% skill assessment accuracy (Spearman~rho = 0.904), and BLEU-4 of 0.847, with an overall score of 92.1%. The model's ability to generate interpretable natural language descriptions of surgical gestures opens new possibilities for automated surgical education, real-time feedback, and scalable objective assessment in robotic surgery training programs.

Keywords: Robotic Systems, Modeling of Autonomous Systems, Mechatronics
Abstract: Robotic autonomy in unstructured outdoor environments remains challenging due to uncertainty, sensor degradation, partial observability, and dynamic environmental variability. We present a closed-loop embodied AI framework for robust field robotics integrating multimodal perception, uncertaintyaware online adaptation, reinforcement learning (RL), and model predictive control (MPC). The proposed architecture consists of four tightly coupled modules: (i) an RGB-D perception encoder producing latent environmental representations, (ii) an uncertainty estimation module based on predictive entropy, (iii) an online self-supervised adaptation mechanism for representation refinement under distributional shifts, and (iv) a hybrid RL+MPC control stack in which RL generates nominal actions while MPC enforces safety and dynamic feasibility constraints. Experiments are conducted in NVIDIA Isaac Sim across randomized forest-navigation scenarios involving varying terrain slope, vegetation density, sensor noise, and partial occlusions. Compared with non-adaptive RL, MPC-only, and vision-only baselines, the proposed approach improves task success rate by 21% and reduces collision frequency by 35%. Ablation studies demonstrate that uncertainty-triggered adaptation and MPC safety filtering are both critical for robust autonomy in dynamic outdoor environments.

Keywords: Fault Monitoring and Diagnosis, Intelligent Transportation Systems, Machine Learning
Abstract: Automotive systems generate massive sensor data requiring sophisticated anomaly detection approaches. We explore parameter-efficient fine-tuning strategies with four LLMs: Orca2:7b, Marco-O1:7b, Cogito:8b, and Llama3-Groq:8b. Our unified framework processes four hardware datasets (CIA Engine, Battery, SaferCar) and two log benchmarks (BGL, Thunderbird) using Sentence Transformers, PCA dimensionality reduction, HDBSCAN clustering, and Logistic Regression classification. For hardware fault detection, our PEFT methods achieve 94.4–94.6% accuracy with safety scores of 0.940–0.945 while reducing computational requirements by 50–99%, enabling real-time edge deployment. For log-based anomaly detection, transformer-driven pipelines achieve macro F1 scores of 0.989 (BGL) and 0.990 (Thunderbird) with ROC-AUC scores exceeding 0.999 and false positive rates below 1%. The unified framework enables production-ready automotive fault intelligence across diverse hardware signals and operational logs.

Keywords: Digital Twin, Intelligent Transportation Systems, Cyber-physical systems
Abstract: Urban bus-corridor digital twins often use door-to-door observations that combine traffic motion with holding, layover, and stop service. We introduce an operator-aware Robust Calibration Method for Digital Twins (RCMDT) that separates bus and stop calibration from background-traffic updates. Both levels are frozen before cross-day evaluation. Rule C flags link-hour medians when travel time exceeds 325 s, effective speed is below 5 km/h, and distance does not exceed 1500 m. We evaluate the rule through development-set sensitivity analysis and comparisons with median absolute deviation (MAD) and Isolation Forest. Intelligent Road Network (IRN) data provide an external diagnostic without influencing threshold fitting. Level 1 (L1) compares Gaussian-process Bayesian optimization (BO), using expected improvement, with continued Latin Hypercube Sampling (LHS) under equal 40-evaluation budgets. Level 2 (L2) applies three iterative ensemble smoothing (IES) updates with 10 ensemble members and moving-only observations. Rule C retains 56.6% of eligible link keys on the December 19 development set and 47.6% on the December 30 cross-day set. Its full-window speed Kolmogorov–Smirnov (K-S) distance is lower than those of MAD and Isolation Forest on both dates, although retention is lower. Across four common ablation seeds, fully audited RCMDT (Full-RCMDT) yields a mean cross-day K-S distance of 0.348 ± 0.035. RCMDT using unaudited raw observations (Raw-RCMDT) yields 0.351 ± 0.027, while the Bayesian-optimization-only configuration (BO-only) gives the lowest value at 0.339 ± 0.035. BO also achieves a lower mean final L1 score than LHS in four of five search seeds. Performance varies by data split and evaluation criterion.

Keywords: Application of Artificial Intelligence, Machine Learning, Transfer Learning
Abstract: In incomplete data classification, missing values are usually imputed, but it will create distribution difference between imputed and complete data. To address this issue, we propose Covariance Alignment guided Incomplete Data classification (CAID) for the tasks with incomplete training and complete test data. In CAID, we first impute the missing attribute values, then reduces Global Covariance Difference (GCD) between imputed training and test data to obtain a feature transformation matrix for learning a classifier with the transformed imputed training data. Afterwards, we estimate Local Covariance Difference (LCD) using the predictions of test data, and update the matrix by minimizing both GCD and LCD. This operation iteratively refines the predictions of test data, and reduces the distribution difference across datasets until convergence is achieved. CAID was tested on several UCI datasets, and the experimental results demonstrate that CAID can significantly improve the accuracy compared to many related incomplete data classification methods.

Keywords: Cyber Systems and Engineering, Complex Network, Distributed Intelligent Systems
Abstract: State sharding improves blockchain scalability by distributing nodes, transactions, and account states into multiple shards for parallel execution. However, naive account distribution strategies often ignore real transaction interaction patterns, resulting in excessive cross-shard transactions (CSTXs) and severe workload imbalance, which jointly degrade throughput and confirmation latency.

In this paper, we focus on alleviating the cross-shard bottleneck from the perspective of account distribution. We propose CA-LB, a community-aware load-balanced account partitioning algorithm for state-sharded blockchains. CA-LB introduces a hybrid scoring model that integrates node propagation importance, structural similarity, internal edge ratio, and explicit load-penalty factors, enabling the clustering of highly interactive accounts while preventing shard overload.

Experiments on real Ethereum transaction traces using the BlockEmulator demonstrate that CA-LB significantly reduces cross-shard transaction rates while maintaining balanced shard workloads, achieving up to 61.2% higher throughput and up to 51.2% lower confirmation latency compared with representative graph partitioning and community detection baselines.

Keywords: Generative AI, Image Processing and Pattern Recognition, Information Assurance and Intelligence
Abstract: Diffusion models can generate high-quality images, but it is hard to track ownership and provenance of the outputs. We present a watermarking framework for diffusion models that supports both noise-domain and frequency-domain embedding, and we study how using error-correcting codes (ECC) changes bit recovery and prompt alignment. In the noise domain, we embed an ECC codeword into the initial sampling noise. Compared with repetition, ECC causes a small drop in bit accuracy under common perturbations, while keeping prompt alignment more stable (CLIP changes are near zero and slightly positive on average). In the frequency domain, we use a dual-ring layout to separate detection and payload, which improves reliable verification and message recovery. Experiments on Stable Diffusion v2.1 (512×512) with nine common perturbations, and comparisons with prior baselines (DWT-DCT, DWT-DCT-SVD, RivaGAN, Tree-Ring, and Gaussian Shading), show that our method offers a practical trade-off among invisibility, robustness, and capacity for copyright protection and provenance in diffusion-generated images.

Keywords: System Modeling and Control, Cyber-physical systems, Trust in Autonomous Systems
Abstract: This paper introduces the ACAI Navigator, a cybernetic regulatory overlay for Large Language Model (LLM) agents that augments inference-time generation with persistent feedback state, bounded homeostatic regulation, and policy-level control. Contemporary LLMs operate primarily as reactive sequence predictors and do not maintain explicit internal variables for long-horizon behavioral stability. ACAI treats the LLM as a stochastic plant and the Navigator as an external controller that senses dialogue disturbances, updates an internal state vector consisting of Emotion, Mind, Context, and Burnout, and maps this state to mode selection, prompt framing, and temperature constraints. The revised formulation provides explicit state-update equations and an implemented control law for mode gating and temperature modulation under API-level constraints. We further describe a multi-scenario evaluation protocol comparing ACAI-controlled generation against a vanilla LLM baseline under stress, urgency, adversarial tone, and instruction-drift scenarios. The framework does not claim global stability of the underlying LLM plant; rather, it establishes boundedness of the controller state and admissible control actions while empirically measuring task compliance, affect deviation, mode switching, temperature modulation, and degenerate behavior under disturbance.

Keywords: Medical Informatics, Visual Analytics, Communication, Machine Learning
Abstract: Accurate medical image segmentation supports lesion localization and quantitative assessment in computer-aided clinical workflows. Self-supervised vision foundation models, such as DINOv3, provide strong hierarchical representations, but adapting their patch tokens for lightweight dense prediction remains challenging. Features from different depths differ in semantics and spectral composition, so direct fusion can weaken spatial coherence and boundary detail. We propose FlexDINO, a compact segmentation framework that uses a frozen DINOv3-S encoder with two lightweight modules. The SpecBridge Adapter aligns cross-layer tokens into spatially coherent multi-scale features, while the FlexFusion Head uses frequency-aware gates to balance low-frequency semantic context and high-frequency structural detail. Under a controlled local reimplementation protocol, FlexDINO achieves the lowest HD95 on TN3K, Kvasir-SEG, and ISIC. It also obtains the best Dice and IoU on Kvasir-SEG and ISIC while reducing trainable parameters by 4.8% relative to the baseline decoder. These results show that frequency-guided token adaptation and adaptive fusion can improve dense prediction with frozen foundation-model encoders.

Keywords: Big Data Computing, Visual Analytics, Communication, Cyber Systems and Engineering
Abstract: Ride cancellations in urban mobility applications remain frequent and, from the passengers' perspective, are often poorly explained events that undermine service reliability. This paper investigates cancellation patterns in Rio de Janeiro through a data-fusion and visualization-oriented artifact grounded in Design Science Research. The dashboard integrates anonymized ride data with public and georeferenced sources, including territorial layers, urban incident records, socioeconomic indicators, and user complaints. The contribution is not a new visualization technique in isolation, but a reproducible analytical pipeline that connects operational, spatial, temporal, and textual evidence to support hypothesis generation about service instability, perceived safety, and signals consistent with socio-spatial disparities. To avoid causal overclaiming, the study treats redlining-related findings as exploratory indicators rather than proof of discriminatory intent. A pilot evaluation with 14 participants reports positive perceptions of visualization clarity, navigation, and analytical usefulness, while also identifying the need for broader task-oriented validation and stronger statistical modeling in future work.

Keywords: Application of Artificial Intelligence, Environmental Sensing, Smart Buildings, Smart Cities and Infrastructures
Abstract: Semi-supervised change detection is increasingly important for high-resolution remote sensing applications, where dense pixel-level annotations are costly and difficult to obtain. However, existing semi-supervised methods often suffer from limited inter-temporal feature interaction, weak sensitivity to small-scale changes, and pseudo-change noise caused by illumination variations, seasonal differences, and imperfect image registration. These issues are further amplified in teacher–student frameworks, where inaccurate pseudo-labels may propagate errors during consistency training. To address these challenges, this paper proposes a Feature Cross-Fusion Semi-Supervised Change Detection Network (FCFSSCDNet) that enhances temporal interaction, multi-scale contextual modeling, and hierarchical feature alignment within a unified framework. Specifically, a Cross-Attention Enhancement Module models cross-temporal dependencies to suppress pseudo-change responses, a Global Context Module captures scale-variant contextual information to improve sensitivity to small changes, and a Feature Cross Fusion Module progressively aligns and integrates multi-level features during decoding. Extensive experiments on the WHU-CD and LEVIR-CD benchmarks demonstrate that the proposed method consistently outperforms state-of-the-art semi-supervised approaches under varying labeled-data ratios. For example, under the 10% labeled setting on WHU-CD, FCFSSCDNet achieves an F1-score of 89.04% and an IoU of 80.24%, while on LEVIR-CD it reaches an F1-score of 90.44% under the 30% labeled setting. These results indicate that FCFSSCDNet provides a robust and effective solution for semi-supervised remote sensing change detection in low-annotation scenarios.

Keywords: Human-centered Learning, Generative AI, Human-Machine Cooperation and Systems
Abstract: Tutor response generation should identify and localize student mistakes, provide concrete guidance, and remain actionable without inducing over-reliance. SRL-oriented scaffolding can support such behavior through planning, monitoring, and reflection prompts, yet in modern LLM tutors these supports are often implicit and difficult to inspect or compare. We address this problem for tutor response generation (TRG) with a structured multi-role workflow that exposes intermediate diagnosis and strategy artifacts and enforces a rubric-first response contract with adaptive SRL micro-scaffolds phrased as brief learner check-ins. We evaluate on BEA 2025 MRBench v3 with a calibrated LLM judge anchored to human-labeled dev annotations for four pedagogical dimensions: Mistake Identification, Mistake Location, Providing Guidance, and Actionability. Under this calibrated rubric protocol, our routed workflow reaches MI 0.937, ML 0.864, PG 0.840, and AC 0.987 on test and scores above compared baselines, including a rubric-in-prompt configuration, while a paired router control shows near-parity quality (BEA avg 0.907 vs. 0.914) with about 88% fewer generation-side tokens. These results position the routed system as a practical operating point for training-free agentic tutoring while keeping conclusions bounded by judge calibration quality.

Keywords: AI and Applications, Computational Intelligence, Systems Safety and Security
Abstract: Vulnerability detection is crucial for software security. However, mainstream methods face dual challenges: syntax-based approaches fail to capture multi-hop semantic dependencies across statements, while large language models (LLMs), despite strong code comprehension, produce nontransparent decisions due to their black-box nature. These issues undermine the reliability and practical utility of vulnerability detection. To address these issues, we propose VulPESR (Vulnerability Detection via Program Execution Semantic Graph-based Reasoning). It recovers multi-hop semantic dependencies by constructing a Program Execution Semantic Graph (PESG). Furthermore, it ensures decision transparency and traceability through a structured, three-stage chain-of-thought mechanism. Moreover, to alleviate class imbalance, LLMs are leveraged to generate diverse vulnerable samples. Experiments on BigVul and Devign show maximum F1-score improvements of 21.44% and 20.19% over baselines, and traceable evidence chains unify high detection performance with decision transparency.

Keywords: Cybernetics for Informatics, Systems Safety and Security, AI and Applications
Abstract: Software-Defined Networking (SDN), particularly in emerging 6G architectures, is increasingly exposed to sophisticated multi-stage cyberattacks that current AI-based anomaly detection systems struggle to detect due to limited and non-representative training datasets. Most existing datasets lack fine-grained correlation between network traffic and host-level process execution, creating a context gap that degrades detection accuracy under encrypted and high-rate traffic conditions. This paper proposes a methodology for generating high-fidelity, context-rich security datasets by correlating packet-level network flows with kernel-level host telemetry using extended Berkeley Packet Filter (eBPF). An optimization-based correlation framework is formulated and implemented in a heterogeneous SDN testbed executing realistic multi-stage attack scenarios. The resulting dataset enriches each flow with causal host-process context, enabling intent-aware labeling. Experimental evaluation based on learning curves and ROC analysis demonstrates faster convergence and high detection performance, achieving F1-scores of 99.70% with LSTM and 99.54% with Random Forest models, significantly outperforming packet-only benchmarks. These results confirm that causal host--network context is a critical missing component for effective dataset generation in next-generation SDN and 6G networks.

Keywords: Image Processing and Pattern Recognition, Multimedia Systems, Machine Learning
Abstract: 3D Gaussian Splatting (3DGS) has replaced the Neural Radiance Fields (NeRFs) as the most advanced scheme for scene rendering. Benefiting from the advantages of 3DGS in fast rendering and explicit representation, some works migrate 3DGS to scene perception tasks with the help of feature embeddings provided by 2D foundation models. However, due to the limitations of 2D foundation models, it is hard to provide accurate semantic annotations for all pixels of each view. The feature embeddings of the same category in different views may be discrepant to some extent, which may compromise the accuracy of feature distillation to Gaussians. To this end, we propose cross-view semantic consistency learning (CSC-learning), which leverages the similarity of feature embeddings across different views to guide Gaussian features toward consistency, while simultaneously further refining Gaussian geometry and appearance attributes. In addition, we introduce the state-of-the-art general foundation model, DINOv3, which can generate high-quality feature embeddings to enhance the accuracy of feature distillation and the effect of CSC-learning. Based on these, we construct a cross-view semantic consistency-guided feature distillation framework for 3DGS (CSC-GS). Extensive experimental results demonstrate that our method achieves state-of-the-art performance in both scene rendering and perception tasks.

Keywords: Generative AI, Multimedia Computation, Affective Computing
Abstract: ，政治新闻的多模态情感分析是 这对民意监测至关重要，但已有方法 脸 处理语义等问题时的局限性 不匹配 文本与图像之间，特征冗余和跨模态 噪音。这种缺陷在 资源有限、高度敏感的南方环境 朝鲜语 政治，有效的解决方案依然难以捉摸。就是这样 纸张 提出一种轻量级多模态情感分析方法 为韩国政治场景量身定制。通过定制 指导Qwen3-VL-plus大型语言的提示工程模型，政治新闻图像被明确转换为 高保真韩语文本描述。这变换 复杂体 跨模态比对问题转化为高效、语义上 增强的单模语义理解任务。建筑 在此基础上，我们构建了一个分析框架 基于KoELECTRA，采用多头差速器 捕捉细致区分的注意力机制 立场与情感之间的抉择。此外，我们还设计了 多任Ó

Keywords: Systems Safety and Security, Heuristic Algorithms, Distributed Intelligent Systems
Abstract: The Delegated Proof of Stake (DPoS) protocol is widely utilized as a consensus mechanism. However, it faces challenges such as the selection of witness nodes that do not reflect the preferences of the majority of stakeholders and the vulnerability of the voting process to manipulation. To address this problem, this paper formulates witness election as a permutation-based preference aggregation problem. Specifically, Kendall's rank correlation is adopted as the fitness function to measure the consistency between a candidate witness ranking and stakeholders' voting preferences. Based on this formulation, we adapt the Equilibrium Optimizer (EO) to search for a witness ranking that maximizes aggregate stakeholder preference consistency. To reduce the computational overhead caused by repeated fitness evaluation, a Spark-based parallel implementation is further designed. Experimental results reveal that the witness nodes selected through EO are consistent with the preferences of the majority of stakeholders, and EO performs better than the five new heuristic algorithms. Furthermore, our proposed DPoS protocol has been proven to possess high scalability and resilience against vote manipulation.

Keywords: Deep Learning, System Modeling and Control
Abstract: Anomaly detection plays a critical role in ensuring the reliability and efficiency of industrial systems. Traditional anomaly detection models typically rely on the availability of both normal and anomalous data. However, in many industrial scenarios, anomalous data is scarce or even entirely absent, posing significant challenges for robust model training. To mitigate this issue, synthetic anomalies are often generated, but such anomalies may not accurately represent real-world irregularities, leading to overfitting and reduced generalization. This work introduces the Spectral-Regularized Denoising Autoencoder (SR-DAE), a novel approach for anomaly detection in high-dimensional spectral data, designed to address the challenge of scarce or unavailable anomalous data. SR-DAE incorporates a Physically-Informed Spectral Regularization mechanism, allowing the model to dynamically focus on the most informative frequency bands for anomaly detection. This mechanism enhances the model's ability to detect anomalies with greater accuracy by prioritizing relevant features and improving its sensitivity to subtle variations in the data. We evaluate SR-DAE on two vibration dataset from turbochargers, comparing its performance with several baseline models. Experimental results demonstrate that SR-DAE significantly outperforms the baseline models across key evaluation metrics, offering an optimal balance between minimizing false positives and maximizing anomaly detection.

Keywords: Image Processing and Pattern Recognition, Virtual and Augmented Reality Systems, Generative AI
Abstract: 本文展示了高效的动态人类 称为动态人类高斯化身的重建方法 结合空间——时间特征和点——水平 功能（DHAG）。该方法能够重建 人体从单眼视频数据中获得的，针对 现有方法在处理非刚性问题上的不足 人体的变形和自我阻塞。DHAG 使用SMPL模型初始化三维高斯点，且 构建了基于以下内容的人类姿态优化方法 时空特征（HPO-ST），其 捕获由 动态人体中的非刚性变形。与此同时， 基于 点级特征（DRO-P）缓解了细节损失 由人类通过导电投射引起的自我遮蔽 对三维高斯点进行可见性分析，对齐 具有图像特征的可见空间投影点，以及 提取点级特征。此外，由 计算高斯分布之间的瓦瑟斯坦距离 点，即分裂、克隆和合并的过程 高斯点被优化，从而减少冗余 在重建过&#

Keywords: Manufacturing Automation and Systems, Reinforcement Learning, System Modeling and Control
Abstract: To address the challenges of scheduling policy transferability across diverse robotic cellular manufacturing systems (RCMSs), this paper proposes PNc-GenRL, a generalizable scheduling framework. To formally capture the operational logic of such systems, place-timed Petri nets (PTPNs) are employed as the modeling tool, providing an environment for reinforcement learning. On this basis, PNc-GenRL transforms PTPN states into high-dimensional feature representations through multi-layer perceptrons. Subsequently, PNc-GenRL generates optimal scheduling decisions by mapping these informed state features to expected rewards using Q-learning. We develop an end-to-end training mode where a stochastic on-the-fly simulator generates varying RCMSs in each episode, enabling PNc-GenRL to learn adequately across heterogeneous systems. Extensive experiments on 12 configurations derived from three real-world RCMS benchmarks demonstrate that PNc-GenRL achieves stable convergence and exhibits superior zero-shot generalization. The model consistently produces competitive makespans on unseen systems without the need for retraining or complex parameter tuning, thereby significantly enhancing production efficiency in dynamic manufacturing environments. For reproducibility, the source code is available at: url{https://github.com/MartiniiiK/PNc-GenRL}.

Keywords: Adaptive Systems, Social simulations
Abstract: In today's highly competitive global market, multinational companies heavily rely on efficient expert allocation to maximize their market share across diverse regions. However, existing resource allocation methods often overlook the non-linear diminishing marginal utility of different markets and fail to address nuanced strategic risks, such as resource overflow in saturated regions and market collapses in vulnerable ones. To address this complex Maximizing Market Share via Expert Allocation (MMSEA) problem, this paper formalizes the issue by extending the Group Role Assignment (GRA) model, a core sub-model of the Environments-Classes, Agents, Roles, Groups, and Objects (E-CARGO) framework. Specifically, we introduce a novel allocation mechanism that incorporates agent categorization, mathematically models non-linear diminishing marginal returns, and utilizes a penalty-based soft-constraint mechanism to balance resource utilization against strategic business priorities. Extensive large-scale simulation experiments are conducted to evaluate the model across various business strategies. The results demonstrate that the proposed conservative strategy significantly improves global market stability, drastically reducing collapsed markets from 0.24 to 0.01 and resource overflow from 1.235 to 0.270, with only a minimal trade-off in absolute total performance. Ultimately, it bridges the gap between macroscopic business strategy and mathematical optimization, providing a robust, quantitative decision-making tool for executives to align global expert deployment with specific risk tolerances and corporate visions.

Keywords: Adaptive Systems, Social simulations
Abstract: 在当代企业招聘中，一个关键挑战是从庞大的候选人库中挑选既具备卓越能力又具备长期部门契合度的管理培训生。越来越多的企业将关注点从短期实习表现转向员工的发展潜力。完全依赖基于绩效的评估往往忽视了长期因素，进而导致工作效率下降等问题。公司在为特定部门选拔合适实习生时，难以在短期成就与长期考量之间取得平衡。本文通过扩展群体角色分配（GRA）模型形式化了该问题，并进行了大规模社会模拟实验。通过引入三维加权机制，研究企业如何在短期绩效与长期潜力之间取得平衡，全面评估管理培训生并实现最佳招聘策略。这种方法使各部门能够理性地选择所需的学员，有效地实现了学员、部门和公司三振的成果

Keywords: Deep Learning, Machine Learning, Machine Vision
Abstract: Deep learning methods for detecting forged remote sensing (RS) images have made great progress in terms of performance. However, most of them are based on a single RGB color space and may have some limitations in feature extraction. Thus, we propose a new method called Differential evolution-optimized Evidence fusion across Multi-color Spaces (DEMS) for RS image forgery detection. Firstly, the RS images are converted to other related color spaces (e.g., HSV and LAB). Then, parallel convolutional neural networks with the same structure are trained to extract complementary information. A differential evolution-based weight optimization algorithm is designed to dynamically determine the weights of these color spaces. Finally, a weighted evidence fusion strategy is applied to integrate the complementary information for forgery detection. Experimental results demonstrate that DEMS exhibits superior performance against two categories of baseline methods across the Fake-Vaihingen and Fake-LoveDA datasets: it outperforms RGB-only methods by 14.10% and 11.63%, and exceeds the best single non-RGB color space method by 8.76% and 8.31%, respectively. These results comprehensively verify DEMS's effectiveness for detecting forged RS images.

Keywords: Transfer Learning, Deep Learning, Image Processing and Pattern Recognition
Abstract: Transfer learning enables models to utilize knowledge from multiple source domains to improve performance in a target domain. However, many existing methods treat all sources uniformly, overlooking the inter-source heterogeneity. To address this issue, we propose a Distributed Evidence Fusion (DEF) framework that preserves domain-specific features and handles uncertainty through evidence theory. In DEF, a deep sub-domain adaptation network is trained independently on each source domain to capture fine-grained domain-specific features. Subjective logic is then used to quantify the prediction uncertainties of each domain, which are combined via evidence theory to reach a unified and reliable decision. Furthermore, to better align the target distribution, we introduce a multi-round pseudo-labeling strategy that gradually refines the model’s classification performance. Experimental results on multiple benchmark datasets demonstrate that our DEF method consistently outperforms state-of-the-art approaches, particularly in scenarios with challenging domain shifts.

Keywords: Machine Learning, Distributed Intelligent Systems, Deep Learning
Abstract: Federated Learning (FL) suffers from significant performance degradation under data heterogeneity, where client data distributions are non-IID. Beyond accuracy loss, heterogeneous training often produces miscalibrated models that make overconfident and unreliable predictions. Existing calibration-aware FL methods primarily rely on post-hoc calibration, which adjusts prediction probabilities after training but does not influence the optimization dynamics of FL. Under heterogeneous data, this separation between calibration and training leads to biased client updates and unstable aggregation. To address this limitation, we propose FedCal+, a calibration-aware and communication-efficient FL framework that incorporates calibration signals into the training process. FedCal+ estimates client reliability through local calibration and uses this information to guide aggregation, while adaptive gradient quantization reduces communication overhead. We further improve robustness under distribution shifts through federated batch normalization and optional client fine-tuning. Theoretically, we provide convergence guarantees for non-convex objectives, calibration-dependent generalization bounds, and communication complexity analysis. Experiments on SVHN, CIFAR-10, and CIFAR-100 demonstrate consistent improvements in accuracy, calibration quality, and communication efficiency under both IID and non-IID settings.

Keywords: Human-Machine Cooperation and Systems, Manufacturing Automation and Systems, Generative AI
Abstract: Driven by the human-centric vision of Industry 5.0, shop-floor scheduling is expected not only to optimize production efficiency but also to account for worker capability, human–machine matching, and fairness in task allocation. To this end, this paper investigates a dual-resource flexible job shop scheduling problem with human-factor constraints and proposes a scheduling method that integrates a humancentric dispatching model with large language model (LLM)-based rule evolution. In the proposed method, worker capability constraints are treated as hard feasibility conditions for candidate assignments, while a human–machine preference matrix and mismatch cost are introduced to characterize the effect of human–machine mismatch on scheduling execution. In addition, a worker-level fairness budget is imposed to prevent the excessive concentration of mismatch tasks on a small number of workers. On this basis, an LLM-based rule generation mechanism is developed to automatically generate interpretable priority rules within a restricted rule-expression space. A dual-layer memory mechanism, consisting of shortterm elite-rule feedback and cross-generation stable heuristic knowledge, is further incorporated to drive continuous rule evolution. Experimental results show that the proposed method achieves promising performance in balancing makespan and worker-level mismatch burden.

Keywords: Human-Machine Cooperation and Systems, Machine Learning, Decision Support Systems
Abstract: Accurate preoperative identification of difficult airways remains challenging due to the subjectivity inherent in traditional assessment methods. We propose DW-MEF, a Dynamic Weighted Multimodal Evidence Fusion framework that integrates multiple facial images, video sequences, clinical variables, and textual descriptions generated by multimodal large language models (MLLMs). DW-MEF employs a head-specific gating-based knowledge-augmented dynamic weighting mechanism to model the diagnostic relevance of visual data, and adopts a two-stage fusion strategy combining multi-head attention with Dempster-Shafer evidence theory to quantify uncertainty and resolve multimodal conflicts. Experimental results on a real-world clinical dataset demonstrate that DW-MEF outperforms existing representative methods. Our framework provides a transparent and trustworthy solution for airway assessment assisted by artificial intelligence.

Keywords: Fault Monitoring and Diagnosis, Generative AI, Trust in Autonomous Systems
Abstract: Outcome-level metrics are often too coarse for diagnosing why evidence-grounded research agents fail when trajectories mix search, evidence collection, synthesis, and stop decisions. We formulate structured failure diagnosis as a pilot-scale problem over real agent trajectories: assigning a primary failure type, an earliest critical root-cause step, and optional propagated symptom steps from an evidence-state graph and a synchronized proxy-state trace. On a reviewed pool of 14 real failures drawn from a 10-task, 20-run source benchmark, the full pipeline attains 14/14 failure-type accuracy and 12/14 critical-step exact accuracy, compared with 4/14 and 3/14 for a minimal single-call flat-log judge under the same protocol. This contrast is explicitly local to the reviewed N=14 pool, and root-cause versus symptom separation is the only aggregate change observed there. Graph and proxy-state ablations remain tied with the full system, so the value of those components remains unresolved on the reviewed real-failure pool. Because the reviewed real-failure pool is small and the graph/proxy-state ablations remain tied with the full system, we present the result as pilot-scale evidence for a structured diagnosis formulation rather than as validation of graph-specific superiority or mature benchmark-level performance.

Keywords: Biometric Systems and Bioinformatics, Biometrics and Applications
Abstract: The fundamental challenge in facial paralysis assessment lies in identifying facial motor biomarkers with diagnostic significance. However, the diagnostic contribution of individual biomarkers varies considerably, exhibiting substantial inter-patient heterogeneity. Existing automated methods typically operate on fixed network architectures, which are ill-suited for modeling hierarchical relationships among biomarkers and accommodating individual variability. To address these limitations, we propose a Hierarchical Dynamic Attention Patient-Adaptive Network (HiDAPA) for fine-grained facial biomarker importance analysis. HiDAPA organizes 313 clinical indicators from 11 facial actions into a three-level hierarchy: type level, action-type interaction, and fine-grained indicators. A dynamic graph learning module integrates data-driven correlations with medical priors to construct personalized adjacency matrices. A patient-adaptive graph attention module generates transformation parameters and attention weights conditioned on patient characteristics. Experiments on 63 clinical samples show HiDAPA outperforms several state-of-the-art baselines (Pearson 0.979, Acc@10 87.3%), with learned importance distributions aligned with clinical cognition.

Keywords: Machine Learning, AI and Applications, Representation Learning
Abstract: In implicit feedback-based recommender systems, noisy interactions—such as misclicks and non-preference behaviors—can severely distort the optimization process. Existing methods typically address this issue by down-weighting or entirely discarding noisy feedback. However, these approaches are prone to misidentifying noisy interactions or exacerbating data sparsity by mistakenly filtering out informative hard samples. In this paper, we propose a novel denoising framework named HiT-D, based on hierarchical temporal evidence accumulation. First, we introduce an energy-based probabilistic resampling mechanism to reduce the optimization frequency of high-energy noise. Subsequently, we characterize interactive behaviors using global selection frequency and statistical stability, leveraging the empirical lower bound of the loss to structurally distinguish genuine hard samples from stubborn noise. Finally, for persistently abnormal interactions, we accumulate historical evidence across training epochs, incorporating time decay and an exponential penalty to achieve progressive down-weighting rather than aggressive one-off truncation. Extensive experiments on three real-world datasets with varying noise ratios and two representative backbone models validate the effectiveness and robustness of the proposed framework in fostering trustworthy, human-centric recommender systems.

Keywords: Decision Support Systems, Computational Intelligence in Information
Abstract: Hyperspectral satellite constellations are increasingly required to perform time-critical Earth observation missions while operating under stringent resource constraints, including intermittent visibility, limited downlink capacity, and tightly coupled energy dynamics. These factors make observation planning and data transmission inherently interdependent, turning mission design into a large-scale joint optimization problem. To address these challenges, we develop a unified observation-downlink scheduling model that integrates satellite-target assignments, observation intensities, compression retention factors, and per-slot downlink allocations within a single constraint-aware formulation. Building on this problem, we propose the Adaptive Teaching-Learning Evolutionary Framework (ATLEF), a population-based optimizer that couples swarm intelligence with a lightweight controller for adaptive operator selection. A hybrid encoding scheme enables continuous optimization of discrete scheduling variables. Extensive experiments on synthetic and realistic multi-hyperspectral scenarios demonstrate that ATLEF achieves higher overall utility, faster convergence, and improved constraint satisfaction compared to representative heuristic and learning-based baselines. It outperforms the best-performing baselines by an average of 8.92% in total utility across various constellation scales. These results highlight ATLEF as an effective and extensible optimization paradigm for next-generation space-air-ground integrated remote-sensing systems.

Keywords: Representation Learning, Machine Vision, Computational Intelligence
Abstract: Current CNN-based video models often struggle to connect low-level frame evidence with high-level behavioral semantics, leading to the loss of critical local details or global temporal context. To address this issue, we propose S-TRGR, a Spatio-Temporal Relation Graph Reasoning framework that explicitly targets mid-level semantic modeling. Our architecture combines a 2D CNN backbone for static feature extraction, a dual-Transformer design to capture spatial and temporal interactions, and a Graph Convolutional Network (GCN) for higher-order relational reasoning. By organizing video content into structured interaction graphs, S-TRGR encodes contextual cues that are essential for anomaly perception. Experiments on UCF-Crime and XD-Violence show that S-TRGR outperforms state-of-the-art methods, delivering a 5.0% absolute gain in Top-1 accuracy on UCF-Crime.

Keywords: AI and Applications, Cloud, IoT, and Robotics Integration, Agent-Based Modeling
Abstract: The integration of Artificial Intelligence into blockchain governance is frequently presented as a natural evolution toward scalable, automated oversight. This paper challenges that narrative by examining the fundamental paradoxes, adversarial vulnerabilities, and systemic failure modes that emerge when algorithmic authority is embedded within decentralized blockchain ecosystems. We argue that AI governance in trustless systems creates a recursive contradiction: the introduction of centralized inferential logic into architectures designed to eliminate centralized trust. Through analysis of adversarial dynamics, scaled human-oversight limitations, and cascading failure scenarios, we demonstrate that conventional AI governance frameworks underestimate the sophistication of adaptive adversaries and overestimate the reliability of human-in-the-loop safeguards. Rather than proposing yet another governance architecture, we advocate for anti-fragile governance design principles that anticipate, absorb, and learn from their own failures. The paper concludes that sustainable governance in decentralized ownership systems requires not better algorithms but better epistemological humility about algorithmic limitations.

Keywords: Machine Learning, Computational Intelligence, Distributed Intelligent Systems
Abstract: In response to growing privacy concerns, Federated Learning (FL) enables collaborative model training without sharing raw data. While most FL methods focus on supervised tasks, unsupervised clustering---especially with uncertain data---remains underexplored. To address this issue, we propose Federated Evidential C-Means (F-ECM), which incorporates evidential clustering into the FL framework. F-ECM follows a client--server structure: each client runs local ECM iterations on private data to generate partition matrices and belief assignments, updates its local cluster centers, and sends the centers to the server. The server updates global prototypes by aggregating client updates and broadcasts them for the next round. Experiments on synthetic and real datasets show that F-ECM effectively models uncertainty and significantly outperforms related clustering baselines, particularly under Non-IID and noisy settings.

Keywords: AI and Applications, Machine Learning, Deep Learning
Abstract: Vision Transformers (ViTs) achieve state-of-the-art performance in vision tasks but face computational challenges in real-time multimedia and signal processing applications. To address this, we propose a parametric activation-based pruning framework that uses trainable scaling factors to automatically identify and prune redundant components during training. Unlike many existing methods, our approach requires no auxiliary modules, attention-based scoring, or architectural changes, making it lightweight, easy to implement, and broadly applicable. Experiments show that our method reduces FLOPs by up to 60% and compresses parameters by over 65%, with less than 2% accuracy loss. It also generalizes well to object detection and semantic segmentation with minimal performance degradation. These results demonstrate that our activation-driven pruning offers an efficient and deployment-friendly solution for Vision Transformers, particularly suited for resource-constrained multimedia and edge scenarios.

Keywords: AI and Applications, Human-Machine Interaction
Abstract: As organizations increasingly adopt artificial intelligence (AI) technologies in human resource management, AI-powered chatbot mentors are emerging as a novel tool to support employee onboarding. However, limited research has examined how the interaction characteristics of AI systems influence newcomers’ socialization outcomes and job attitudes. Drawing on organizational socialization theory, this study conceptualizes AI chatbot mentors as socio-technical socialization agents and investigates how two interaction characteristics “personalization” and “spontaneity” affect newcomers’ role clarity, self-efficacy, and expected acceptance by others. The socialization outcomes are examined in relation to job satisfaction and intention to exit. Survey data were collected from 845 Japanese Generation Z employees, and the proposed model was tested using partial least squares structural equation modeling. The results show that both personalization and spontaneity significantly enhance newcomers’ role clarity, self-efficacy, and expected acceptance by others. In turn, role clarity and self-efficacy significantly increase job satisfaction and reduce intention to exit. However, expected acceptance by others does not significantly influence job satisfaction, suggesting that AI-mediated onboarding may be more effective in supporting cognitive and task-related adjustment than interpersonal social integration. These findings contribute to the organizational socialization literature by identifying AI chatbot mentors as an emerging digital socialization mechanism. The study also highlights the importance of cultural context, particularly in Japan where workplace relationships and mentorship traditions emphasize interpersonal interaction. Practical implications suggest that organizations should integrate AI-enabled onboarding tools with human mentorship to support both task adaptation and social integration.

Keywords: Neural Networks and their Applications, Autonomous Vehicle, Image Processing and Pattern Recognition
Abstract: Reliable perception in autonomous vehicles requires the conversion of raw sensor data into structured semantic representations for downstream decision-making. However, vision-based sensors are vulnerable to noise, adverse weather, motion blur, and low-light conditions, which can significantly degrade perception performance. Modern autonomous driving systems rely on deep learning models to interpret complex urban scenes from multimodal sensor input, including cameras, LiDAR, and radar. In particular, object detection and segmentation architectures such as YOLO, Faster R-CNN, DeepLab, and Mask R-CNN enable scene understanding at both the pixel and instance levels, providing the semantic information necessary for safe navigation. In this work, the robustness of deep learning feature representations in object detection and segmentation models under sensor noise is investigated, with particular emphasis on how noise-invariant and sparsity-regularized encodings preserve structural and semantic information in simulated driving environments. Sparse autoencoders achieve lower reconstruction error and improved structural fidelity compared to baseline and denoising approaches; however, consistently retaining rich semantic information remains challenging. Although noise-invariant encodings may appear robust under moderate noise levels, they can inadvertently suppress subtle structural details that are essential for precise semantic understanding. Consequently, robustness to noise alone is insufficient, and the preservation of the informative signal structure is critical to reliable semantic representations in the perception of autonomous vehicles.

Keywords: AI and Applications, Image Processing and Pattern Recognition, Systems Safety and Security
Abstract: 探测显示深渊存在严重漏洞 神经网络（DNN）。然而，高效搜索 高维离散空间，同时满足严格 区域预算 依然是一个重要的瓶颈。现有白盒 方法 由于连续而遭受离散化误差 放松， 而黑箱启发式方法则受限于简单 几何原语，未能利用复数 拓扑 建筑。确定理论上的脆弱性 上游 被绑定，我们提出拓扑-像素解耦层级 优化（TPDHO）。该框架正交 将二进制掩模生成分解为两个协作程序 等级：A 宏观决策层面，其中强化学习（RL） 智能体动态调度拓扑算子（例如， 分裂， 用于全球探索的瞬移，以及微型执行 Level， 其中自适应混合求解器采用启发式搜索 用于低维调整和梯度引导 贝叶斯 高维像素注入的优化（BO）。 就是这样 混合范式绕过了二元化错误和 探索 无&

Keywords: Image Processing and Pattern Recognition, Intelligent Transportation Systems, Machine Vision
Abstract: Video anomaly detection is a vital task in video analysis. Previous methods adopt simplistic fusion strategies when attempting to utilize dynamic optical flow features, failing to fully exploit cross-modal complementary information. To solve this problem, we propose a Memory-based Multi-modal Fusion anomaly detection Network (MMFNet), which embeds an adaptive Memory-based Multi-modal Fusion (MMF) module. This updatable MMF module records the modal weight allocation patterns across multiple scenarios, thereby enabling dynamic and flexible fusion of RGB and optical flow modalities. Extensive experiments on the multi-scenario TSD dataset show that our MMFNet achieves an AUC improvement of over 9.8% against state-of-the-art models. Moreover, MMFNet delivers competitive performance on three benchmark datasets (UCSD Ped2, Avenue, and ShanghaiTech), which fully verifies its superior effectiveness and strong generalization capability in video anomaly detection.

Keywords: Application of Artificial Intelligence, Consumer and Industrial Applications, Deep Learning
Abstract: Remote sensing image change detection is essential for applications such as urban monitoring and disaster assessment. However, in semi-supervised settings with limited annotations, complex multi-class scenes often suffer from semantic ambiguity and error accumulation caused by noisy pseudo-labels. To address these challenges, we propose a Semantic Feature Guided Semi-Supervised Change Detection Network (SFGSSCDNet), which integrates external semantic knowledge into a teacher-student framework through a consistency-driven learning paradigm. Specifically, a Multi-attention Feature Fusion Module (MAFM) is designed to enhance multi-scale feature representation, while a Semantic Feature Guidance Module (SFGM) introduces semantic priors from a vision-language model and performs adaptive pseudo-label calibration by modeling the consistency between semantic priors and model predictions. This design enables more reliable supervision and improves discrimination in semantically complex scenarios. Extensive experiments on the SYSU-CD dataset under different labeled-data ratios demonstrate that the proposed method consistently outperforms several state-of-the-art approaches. In particular, SFGSSCDNet achieves an F1-score of 80.64% and an IoU of 67.56% at a 30% label ratio. These results indicate that incorporating semantic priors with consistency-driven guidance provides an effective solution for robust change detection under low-label conditions.

Keywords: Systems Safety and Security, Deep Learning, Application of Artificial Intelligence
Abstract: With the widespread application of federated learning in privacy-sensitive scenarios, its open collaborative training mechanism prevents the central server from verifying local data, thereby exposing the system to severe and stealthy security threats such as backdoor attacks. Particularly in the ubiquitous non-independent and identically distributed (Non-IID) data scenarios, the inherent heterogeneity of client data significantly enhances the stealthiness of malicious updates. This makes it challenging for conventional defense mechanisms to effectively distinguish between benign parameter divergence and malicious feature injection. To address these issues, this paper proposes FedMDR, a multi-dimensional detection and repair backdoor defense method for federated learning. This method precisely quantifies anomalous features from multiple dimensions by extracting the Normalized Update Energy (NEUP) of the model's output layer, the parameter update cosine distance, and the Dynamic Prediction Difference (DDIF). Subsequently, it integrates a Gaussian Mixture Model (GMM) with HDBSCAN to achieve mean fusion and adaptive clustering of these features, effectively overcoming the baseline drift caused by heterogeneous data and enabling precise interception. Furthermore, the proposed scheme leverages a Generative Adversarial Network (GAN) to reverse-reconstruct pseudo-triggers and actively repair the global model via adversarial fine-tuning. Experimental results demonstrate that in Non-IID scenarios, FedMDR can maintain high Main Accuracy (MA) while strictly restricting the Backdoor Accuracy (BA) to an acceptable level, thereby significantly enhancing overall system security.

Keywords: Robotic Systems, Human-Machine Interaction, Autonomous Vehicle
Abstract: Advances in large language models (LLMs) and vision-language models (VLMs) facilitate multi-robot zero-shot object-goal navigation in unfamiliar environments, yet key limitations persist: LLM-based methods rely on language descriptions, losing critical visual geometric details; VLM-based methods, by contrast, use VLMs to reason over global maps but overlook egocentric local information, which leads to suboptimal decisions in complex scenes. To address these issues, we propose Co-EBENav, a human-mimetic multi-robot navigation method that emulates a collaborative ``eye-brain-eye" cycle, thereby fully leveraging both local and global information. Specifically, each robot’s egocentric views from ``eye" are parsed by VLM to mimic right-brain spatial intuition processing, and collectively analyzed by LLMs to emulate left-brain logical reasoning; these two processes jointly update a collaborative global memory system. Actions are determined by projecting the collaborative ``brain’s" memory back to each robot’s ``eye" inputs, unifying distributed local perception with the team’s collaborative global memory. This integration enables Co-EBENav to outperform existing multi-robot zero-shot methods on HM3D and MP3D datasets, achieving superior team success rates and team exploration efficiency.

Keywords: Autonomous Vehicle, Deep Learning, Reinforcement Learning
Abstract: Autonomous driving motion planning requires both stable policy learning from large-scale datasets and adaptive decision-making in dynamic, uncertain environments. However, existing imitation learning (IL) and reinforcement learning (RL) approaches each exhibit critical limitations: IL models often suffer from distribution shift and causal confusion, while pure RL methods struggle with inefficient exploration, reward design complexity, and poor sim-to-real transferability. To address these challenges, we propose DriverTS, a two-stage framework that stabilizes the transition from IL to RL via a progressive actor–critic schedule. In the offline stage, imitation learning pretraining enables the planner to efficiently learn feasible and rule-compliant behaviors from large-scale expert trajectories. In the online stage, reinforcement learning fine-tunes this pretrained policy via interactive environment feedback, improving adaptability and robustness in out-of-distribution (OOD) scenarios. The transition between IL and RL is carefully managed: DriverTS first trains only the critic network before progressively introducing joint actor–critic optimization, ensuring stable policy evolution and preventing catastrophic collapse during fine-tuning. Quantitative experiments in the Waymax simulator demonstrate that DriverTS outperforms all RL-based baselines and remains competitive with the best IL-based methods. More importantly, qualitative analyses reveal that DriverTS effectively mitigates distribution shift and generates novel yet rule-compliant driving behaviors that extend beyond the training data distribution. These results highlight that DriverTS not only bridges the gap between IL and RL but also advances towards more generalizable and interpretable autonomous driving policies.

Keywords: Transfer Learning, Representation Learning, Machine Vision
Abstract: Due to domain shift and annotation noise, existing crowd counting models often suffer significant performance degradation when deployed in unseen domains. To address this challenge, we propose ZxCount, a single-source domain generalization framework that enhances the robustness of density estimation without requiring target-domain data. ZxCount improves domain generalization by leveraging stable, noise-robust patch-wise density representations as domain-invariant structural priors, and adopts a memory-mask-memory architecture with two density priors, namely Global Density Memory and Pure Density Memory, to regularize feature learning. In addition, a Domain-Specific Filtering Unit is introduced to generate reliability masks that explicitly disentangle domain-invariant and domain-specific representations. To further stabilize prototype learning, we employ a Similarity Alignment Loss, while a Density-Aware Constraint is introduced to preserve semantic consistency. Notably, ZxCount eliminates memory-bank querying during inference, enabling faster inference than prior memory-based methods. Extensive experiments across multiple cross-domain benchmarks demonstrate that ZxCount achieves strong performance among existing approaches.

Keywords: Agent-Based Modeling, AI and Applications, Robotic Systems
Abstract: Object Goal Navigation (ObjectNav) refers to an agent navigating to an object in an unseen environment, which is an ability often required in the accomplishment of complex tasks. While existing methods demonstrate proficiency in isolated single object navigation, their limitations emerge in the restricted applicability of lifelong memory representations, which ultimately hinders effective navigation toward continual targets over extended periods. To address this problem, we propose OVAL, a novel lifelong open-vocabulary memory framework, which enables efficient and precise execution of long-term navigation in semantically open tasks. Within this framework, we introduce memory descriptors to facilitate structured management of the memory model. Additionally, we propose a novel probability-based exploration strategy, utilizing a multi-value frontier scoring to enhance lifelong exploration efficiency. Extensive experiments demonstrate the efficiency and robustness of the proposed system. The code of OVAL included in our supplementary materials will be open-sourced.

Keywords: Application of Artificial Intelligence, Multimedia Systems, Interactive and Digital Media
Abstract: LLMs have rapidly advanced in recent years, particularly in multimodal tasks. Among them, Video Montage is a challenging video-related multimodal task. It requires automatically synthesizing a cohesive multimodal narrative by retrieving and sequencing semantically relevant clips from diverse sources, aligning them with structured knowledge and exemplar references, and generating stylistically consistent narration or captions. This task holds significant value in the film and advertising industries, yet it has received limited attention in prior research. To fill this gap, we propose VMAgent, a few-shot LLM-driven agent framework that mimics the human montage workflow through five modular components: preprocessing, analysis, copywriting generation, clip matching, and rendering. We decouple semantic reasoning from temporal control by introducing precise timestamp modeling only at the rendering stage, allowing LLMs to focus on high-level semantic tasks. To facilitate evaluation, we further construct a large-scale dataset called Ad-VM, consisting of 25,000+ videos. Experiments show that VMAgent outperforms end-to-end baselines and ablated variants in copywriting quality, clip alignment, and output validity. We also validate VMAgent in real-world deployment scenarios, and the observed user engagement further demonstrates its practical effectiveness.

Keywords: Affective Computing, Reinforcement Learning, Deep Learning
Abstract: Multimodal sentiment analysis integrates language, visual, and acoustic cues to predict sentiment, but modality quality can vary substantially across utterances due to occlusion, background noise, motion blur, or imperfect transcripts. Such variations make conventional fusion strategies prone to over-relying on unreliable modalities. We propose MRUF, a reliability-aware multimodal fusion method for multimodal sentiment prediction through multi-granularity routing and uncertainty-aware fusion. MRUF first applies temporal gating to obtain sentiment-relevant utterance summaries. It then performs routing at both the subspace level and the modality level, where modality routing is trained with task-aware supervision derived from leave-one-out loss increases to learn interpretable utterance-level modality importance. In addition, MRUF predicts modality-wise uncertainty scores and refines routing weights through inverse-variance reweighting, thereby reducing the influence of unreliable modalities. We further introduce modality-invariant contrastive alignment to stabilize shared representations under heterogeneous inputs. Experiments on CMU-MOSI and CMU-MOSEI under both aligned and unaligned settings show consistent improvements over strong baselines. Further analysis shows that modalities with higher predicted uncertainty tend to receive lower fusion weights, supporting the intended reliability-aware behavior.

Keywords: Image Processing and Pattern Recognition, Systems Safety and Security, Deep Learning
Abstract: Automatic image colorization enables large-scale and low-cost reuse of grayscale media (e.g., manga panels and archival photographs), facilitating unauthorized reuse and redistribution. Once released online, grayscale content can be readily turned into unauthorized colorized derivatives using off-the-shelf models, creating a practical need for proactive, content-side protection at publication time. Building on Uncolorable Examples (UE), which add imperceptible perturbations to released grayscale images to degrade unauthorized colorization, we propose Semantic Color Naturalness Breaker (SCNB)---a semantic-level UE framework that drives colorization outputs toward content-inconsistent colors while preserving the visual fidelity of the released grayscale media. We further introduce Content-aware Color Distributional Distance (CaCDD), a ground-truth-free, content-aware measure of color plausibility derived from semantic color priors, used both as the optimization objective of SCNB and as an evaluation metric. Experiments on ImageNet show that our method remains effective under small perturbation budgets and common post-processing, supporting practical deployment in real-world content-sharing pipelines.

Keywords: Machine Vision, Human-Machine Cooperation and Systems, Intelligence Interaction
Abstract: Understanding 3D affordances, i.e., identifying functionally relevant regions on objects, is critical for robotic manipulation and human–robot interaction. However, sparse point clouds and the lack of texture make it challenging to leverage pretrained 2D vision foundation models. Existing methods encode multi-view depth renderings, but the domain gap between natural images and sparse geometric observations limits part-aware feature learning and feature discriminability. We propose a geometry-aware adaptation framework for vision foundation models, instantiated with DINOv2 in this work. The framework integrates a trainable geometry branch and LoRA modules, using depth and normal information to adapt DINOv2 to sparse 3D observations. A two-stage design facilitates cross-domain knowledge transfer while enforcing multi-view functional consistency. Experiments demonstrate improved affordance localization and robustness under pointcloud corruptions, supporting more reliable perception for downstream robotic interaction.

Keywords: Deep Learning, Machine Learning, Application of Artificial Intelligence
Abstract: With the rapid advancement of deep learning-based speech enhancement (SE), research into lightweight SE models has gained significant momentum. A multitude of high-performance speech enhancement models have recently emerged. However, due to their substantial parameter counts and computational complexity, they are often impractical for speech tasks requiring lightweight deployment and low latency, such as deployment on resource-constrained edge devices. In this paper, we propose HanSENet, a lightweight model designed for low-latency speech enhancement. We introduce a Grouped Prime-Kernel Convolution (GPKC) to capture time-frequency features from both global and local perspectives. Furthermore,a Dual-Path enhancement module based on Mamba (DPM) is proposed to bolster the network’s modeling capability while maintaining low latency. To further improve speech quality, we incorporate an unwrapped loss function based on group delay and instantaneous frequency to refine phase modeling. Experimental results demonstrate that the proposed HanSENet achieves a PESQ score of 3.43 on the VoiceBank+Demand dataset. With only 77.14k parameters and 41.16M Multiply-Accumulate Counts per second (MACs), the model exhibits promising denoising performance.

Keywords: Deep Learning, Machine Vision, Machine Learning
Abstract: Adversarial examples expose critical security vulnerabilities in deep learning, yet existing pixel-space methods often suffer from visible high-frequency noise and structural artifacts. Current approaches struggle to balance high attack success rates with perceptual invisibility, frequently compromising global feature integrity. To address this, we propose a novel framework leveraging multi-scale frequency disentanglement in variational latent spaces. Using differentiable wavelet decomposition, we decompose images into multiple frequency components and encode the high-frequency features into a compact latent representation, which is then optimized via a well-trained hierarchical Vector Quantized VAE (VQ‑VAE). We also enhance latent‑space expressiveness via adaptive codebook optimization and a dead‑code reset mechanism. Experimental results on ImageNet show that the proposed method achieves an attack success rate (ASR) of over 98% across both CNN and Transformer architectures, while preserving superior perceptual fidelity relative to competing methods, with a PSNR of 41.8 dB and an SSIM of 0.97. It enables more effective vulnerability assessment of security-sensitive vision systems through the generation of high-fidelity adversarial examples.

Keywords: Image Processing and Pattern Recognition, Machine Vision, Application of Artificial Intelligence
Abstract: 表结构识别（TSR）旨在恢复 从文档图像逻辑组织表格， 包括行-列关系和合并单元格布局。 尽管近期取得了进展，强劲的TSR仍然面临挑战 具有生成单元格的复杂表、多层头部， 边界薄弱，布局模糊，结构性中 规律性通常仅在隐式中学习，且 预测可靠性很少被明确建模。前往 针对这些限制，我们提出了SPB-TSR，一种网格感知型 以及用于表格结构的信心感知框架 认可。SPB-TSR首先构造候选网格 之前通过OCR文本块间距模式来约束 结构假设空间，然后对此先验进行积分 带有文本语义和几何布局提示 结构感知解码。进一步提升鲁棒性 在歧义性下，我们引入贝叶斯后验解码 为了聚合多个随机预测，估计 结构级置信度，并拒绝不可靠的输出。 PubTabNet 和 TableBank 上的实

Keywords: Generative AI, Deep Learning, Transfer Learning
Abstract: Large Language Models (LLMs) excel at linguistic and reasoning tasks but remain underexplored in time series forecasting due to modality mismatch and high adaptation costs. We propose MemDec-TS, a lightweight memory decoder that enables LLMs to perform time series forecasting without up dating the LLM backbone. MemDec-TS combines a pretrained temporal memory bank for pattern matching across domains, a distributional interpolation adapter for aligning temporal representations with LLM token distributions, and a reversible plug-and-play interface for numerical–token conversion. Evaluated on five datasets spanning energy, traffic, finance, and meteorology, MemDec-TS demonstrates competitive forecast ing performance, promising cross-domain generalization, and limited forgetting. Compared with prior LLM-based methods that rely on fine-tuning or more complex adaptation strategies, MemDec-TS leverages distributional alignment to support efficient adaptation. This design enables frozen LLMs to be applied to time series forecasting in a plug-and-play manner, offering a practical direction for energy, traffic, and finance applications.

Keywords: Big Data Computing, Machine Learning, Representation Learning
Abstract: Heterogeneous Graph Neural Networks (HGNNs) have exhibited remarkable efficacy in modeling complex systems with multiple types of nodes and relations, yet their training on large-scale heterogeneous graphs remains computationally prohibitive. Although graph condensation methods can effectively improve learning efficiency on large-scale graphs, existing condensation processes are mainly designed for homogeneous graphs and typically rely on computationally expensive gradient matching or bilevel optimization paradigms, rendering them impractical for heterogeneous settings. To address these limitations, we propose HGC-RC, a simple yet effective role-aware heterogeneous graph condensation framework. Specifically, HGC-RC first extracts semantically enhanced node embeddings via lightweight propagation. It then introduces a role-aware hybrid clustering strategy consisting of class-partitioned clustering for labeled target nodes to preserve class distributions and unsupervised type-wise clustering for non-target nodes to retain critical cross-type connectivity. Finally, a compact heterogeneous graph is efficiently reconstructed based on the resulting cluster assignments. Extensive experiments demonstrate that HGC-RC outperforms state-of-the-art baselines, offering a practical pathway to accelerate HGNN training on large-scale heterogeneous graphs without sacrificing task performance.

Keywords: Autonomous Vehicle, AI and Applications, Generative AI
Abstract: With the rapid advancement of large language models (LLMs) technologies, their application in the domain of autonomous driving has become increasingly widespread. However, existing methods suffer from unstructured reasoning, poor generalization, and misalignment with human driving intent. While Chain-of-Thought (CoT) reasoning enhances decision transparency, conventional supervised fine-tuning (SFT) fails to fully exploit its potential, and reinforcement learning (RL) approaches face instability and suboptimal reasoning depth. We propose ThinkDrive, a CoT guided progressive RL fine-tuning framework for autonomous driving that synergizes explicit reasoning with difficulty-aware adaptive policy optimization. Our method employs a two-stage training strategy. First, we perform SFT using CoT explanations. Then, we apply progressive RL with a difficulty-aware adaptive policy optimizer that dynamically adjusts learning intensity based on sample complexity. We evaluate our approach on a public dataset. The results show that ThinkDrive outperforms strong RL baselines by 1.45%, 1.95%, and 1.01% on exam, easy-exam, and accuracy, respectively. Moreover, a 2B-parameter model trained with our method surpasses the much larger GPT-4o by 3.28% on the exam metric.

Keywords: Deep Learning, Knowledge Acquisition, Human-centered Learning
Abstract: Knowledge Tracing (KT) aims to model students' evolving knowledge states and predict their performance on future exercises based on historical learning interactions. Despite recent progress, existing KT methods still face two major challenges: pronounced heterogeneity across schools and strict privacy constraints on educational data sharing. Most approaches follow a one-size-fits-all paradigm and therefore generalize poorly across heterogeneous schools. Although federated learning (FL) offers a promising privacy-preserving solution, standard parameter aggregation is insufficient to capture the structural and behavioral diversity of KT data. To address these challenges, we propose FSBKT, a federated KT framework built on school-level structural and behavioral prototype learning. Specifically, each school extracts structural prototypes from knowledge relations and behavioral prototypes from student interaction patterns, thereby alleviating cross-school heterogeneity and enabling effective knowledge transfer. On the server side, we introduce the Mixture-of-Prototypes (MoP) aggregation with an Exponential Moving Average (EMA) to produce stable global prototypes, and further devise a Similarity-based Personalized Aggregation (SPA) strategy to compute client-specific model parameters for subsequent local updates. Experiments on three real-world KT benchmarks demonstrate that FSBKT delivers superior overall predictive performance under highly heterogeneous settings, while maintaining favorable privacy--utility trade-offs as confirmed by membership inference and re-identification attack evaluations.

Keywords: Application of Artificial Intelligence, Complex Network, AI and Applications
Abstract: ，社交推荐系统 1033;用用户的社交ࡃ ; ; ; ; 1;能 关系以提升个性 1270;推荐 表演。然而，现 6377;方法存在一些໾ ; ; ; ; 2;题 例如一阶社会信 4687;不足， 隐式高阶传播中 0340;不可控噪声， 以及用于用户嵌 0837;的刚性融合策஽ ; ; ; ; 3;。这篇论文 提出了DGEMSR，一种á 22;于 显式多阶社会图
9982;双门自适应图 融合。它明确建 7169;了一阶和二阶 社会视角分别使 9992;可学习的 alpha-门控机制，用&# 20110;动态整合它们 在二阶截断传播 0026; 减少噪音和过度 4179;滑。此外，还੍ ; ; ; ; 7; beta-门控机制，基 0110;用户等级， 自适应地融合用 5143;-项目交互嵌入 社会嵌入——增强对& #31038;会信息的依赖 适合低活跃用户 0248;先考虑个人事ࡃ ; ; ; ; 3; 高活跃用户偏好
2290;实验 LastFM和Ciao数据集显示D GEMSR表现优于此 关键推荐指标的 2522;线。

Keywords: Deep Learning, Image Processing and Pattern Recognition, Intelligent Transportation Systems
Abstract: Event cameras provide high-temporal-resolution and high-dynamic-range visual sensing, making them well suited for semantic segmentation in autonomous driving scenarios involving fast motion and challenging illumination conditions. However, event-frame representations are inherently sparse and noisy, which can limit the performance of conventional CNN-based models due to their relatively limited receptive fields and insufficient capability for modeling long-range dependencies. In this paper, we propose Evformer, an EMA-based Transformer framework for event-frame semantic segmentation. Evformer incorporates an Efficient Multi-scale Attention (EMA) module to effectively capture both global context and local structural information through low-rank proxy attention and cross-dimension spatial aggregation. To improve the representation of small and distant objects, a Feature Enhancement Aggregation (FEA) module is introduced to adaptively emphasize informative regions using multi-scale receptive fields. Furthermore, a Multi-scale Perception Decoder (MSD) facilitates hierarchical feature fusion and enhances spatial detail recovery through attention-guided skip connections. Extensive experiments on the DSEC-Semantic and DDD17-Semantic datasets demonstrate that Evformer achieves competitive performance compared with existing methods, attaining mIoU scores of 52.42% and 56.25%, respectively. Qualitative results further indicate that the proposed framework is effective in preserving fine structural details and boundary integrity in complex urban scenes.

Keywords: Generative AI, Deep Learning
Abstract: Text-guided local attribute editing aims to modify target attributes such as color and material while preserving object structure and unedited regions. Existing diffusion-based editing methods have achieved strong semantic alignment, but they still suffer from two common failure modes in local attribute editing: spatial leakage of the editing signal into neighboring regions, and structural degradation caused by globally rewriting the foreground appearance. We formulate local attribute editing as a region-constrained reverse bridge under diffusion inversion, and propose a region-structure-driven editing method that transforms globally semantic-driven updates into structure-preserving regional updates. The method decomposes each intermediate update into a background-preserving term and a foreground editing term. The former stabilizes non-edited regions through background-prioritized reconstruction, while the latter injects target-specific changes through localized residual projection within reliable foreground regions. To realize this formulation, we introduce a soft regional mask jointly constructed from cross-attention seeds and source-image appearance priors, together with a late-stage latent anchoring scheme that suppresses boundary corruption and texture collapse in latent space rather than through naive image-space compositing.

Keywords: Adaptive Systems, Control of Uncertain Systems, Autonomous Vehicle
Abstract: This paper investigates the problem of maintaining high-fidelity vibration isolation and robust tracking in vehicle active suspension systems subjected to unpredictable payload mutations and continuous nonlinear disturbances. First, to effectively handle the parameter variations and system nonlinearities, the quarter-car suspension dynamics are established under a linear parameter-varying (LPV) framework. To tackle unknown severe payload mutations and continuous nonlinear disturbances, a novel constructive adaptive robust control scheme is proposed. This scheme integrates a homogeneous polynomial parameter-dependent (HPPD) gain scheduling mechanism with a tanh-based smooth robust compensator. Furthermore, by introducing a constructive Lyapunov-based approach and designing energy-driven dual adaptive laws, the proposed framework actively estimates the mass mutations and the unknown upper bounds of matched uncertainties. Rigorous mathematical analysis proves that the closed-loop system is strictly uniformly ultimately bounded (UUB). Finally, comprehensive numerical simulations, particularly under extreme over-compensation payload traps and random road excitations, validate the superior robustness, adaptability, and passenger comfort of the proposed strategy compared to traditional fixedgain methods.

Keywords: Machine Learning, Deep Learning, Representation Learning
Abstract: Legal Judgment Prediction (LJP) aims to predict the corresponding charge, relevant law articles, and term of penalty from case fact descriptions. To address the limitations of existing methods in exploiting legal structural information, modeling subtask dependencies, and distinguishing similar cases, we propose SoftRoute-LJP. This unified framework integrates graph-enhanced soft evidence routing with prototype memory. Specifically, the model constructs a law article relation graph and performs law-aware re-encoding of case facts through community-level legal semantic modeling, uses charge prediction as an intermediate signal to build soft routing for law article prediction and penalty prediction, and further introduces historical case prototypes organized by charge to provide supplementary discriminative evidence through prototype retrieval and gated fusion. Experimental results on CAIL2018 show that the proposed method achieves competitive performance on all three subtasks and obtains the best results on 11 of 12 evaluation metrics. Ablation studies further verify the effectiveness of graph enhancement, soft routing, and prototype memory. These results demonstrate the effectiveness of jointly modeling legal structure, inter-task dependencies, and historical case evidence for legal judgment prediction.

Keywords: Generative AI, Deep Learning, Machine Vision
Abstract: 人体运动预测是应用中的关键任务，涵盖从人机交互到自动驾驶。尽管近期的潜扩散模型已证明 在生成多样且真实感方面取得了显著成功 运动时，他们通常使用固定的结构处理骨骼数据， 预定义的运动学树。这种静态拓扑 假设严重限制了网络的能力 捕捉非相邻者之间复杂、动态的交互 关节在多样动作中。为了克服这一限制， 我们提出了自适应图扩散模型，这是一本新颖的作品 生成框架，学习数据驱动的骨架 连接性。具体来说，我们嵌入了一个自适应图 学习模块，位于扩散去噪网络中。作者 共同优化可学习的邻接矩阵 固有的物理骨架，我们动态地采取了方法 推断长程空间依赖关系而不依赖 仅限于手动图设计。广泛的实验 跨越三个

Keywords: Application of Artificial Intelligence, Medical Informatics, Deep Learning
Abstract: Efficient suicide risk assessment in psychological support hotlines remains challenging due to the unstructured nature of long-form clinical dialogues and the lack of validation against real-world clinical outcomes. Current AI methods primarily focus on short-window social media or non-clinical data, often failing to provide the clinical accountability required for high-stakes psychiatric triage. This paper proposes a Human-AI Collaborative Cybernetic System for continuous suicide risk assessment using a real-world dataset of 1,438 subjects with average hour-long sessions. We systematically compare two diagnostic dimensions: Cognitive Tracking (Text Stream), which utilizes an LLM-driven sequential memory stream to trace conscious intent, and Physiological Stress (Audio Stream), which employs deep acoustic decoders to capture subconscious arousal. To ensure clinical safety, we introduce an expert-anchored fusion mechanism that integrates data-driven LLM inference with established psychiatric manual scale ratings. Validated against 12-month clinical follow-up ground truth, our system achieves an F1-score of 78.34% on a 200-case test set. This represents an improvement of 8.25% over manual scales and 8.60% over standalone deep acoustic methods. The results demonstrate that incorporating expert knowledge with long-context LLM reasoning effectively mitigates AI hallucinations and enhances diagnostic robustness in clinical settings.

Keywords: Computational Intelligence, Neural Networks and their Applications, Machine Vision
Abstract: Tensor completion is a crucial task because real-world visual data is often incomplete. Although recent tube transform-based methods have advanced the field, these methods primarily exploit intra-tube correlations while overlooking intra-slice local similarities, resulting in limited performance when handling tube-missing data. To overcome this limitation, we propose a novel dual-module alternating completion framework for low-rank tensor recovery. This framework achieves comprehensive correlation modeling by alternately executing an intra-slice local similarity module and an intra-tube global correlation module. Notably, we introduce a learnable convolutional kernel tensor with adaptive weights in the intra-slice local similarity module to effectively capture the intrinsic local structures of the data. Experiments show that our method outperforms current state-of-the-art approaches in both quantitative and qualitative evaluations, particularly excelling in handling tube-missing data.

Keywords: Machine Vision, Machine Learning, Deep Learning
Abstract: Out-of-Distribution (OOD) image detection aims to identify samples that fall outside the training distribution of a classifier, which is essential for reliable deployment in open-world scenarios. Existing post-hoc methods usually rely on a single reliability cue, such as output confidence or feature distance, and may fail when OOD samples exhibit both unreliable activations and misleading feature structures. In this paper, we propose PFR-OD, a fully post-hoc OOD detection framework that performs progressive reliability assessment within a single feature representation. PFR-OD first introduces Gradient-Guided Adaptive Rectification (GGAR) to regulate unreliable activation magnitudes and improve confidence estimation, and then evaluates structural consistency using a Top-k Prototype Deviation Score (TPD-Score). By progressively combining activation reliability and feature structural consistency, PFR-OD achieves improved ID/OOD separability without retraining or modifying the pre-trained classifier. Extensive experiments on large-scale image benchmarks demonstrate that PFR-OD achieves state-of-the-art OOD detection performance while requiring only an additional gradient computation during inference.

Keywords: Affective Computing, Machine Learning
Abstract: Emotion Recognition in Conversations (ERC) aims to accurately identify speakers' emotions in multi-turn dialogue scenarios. Existing ERC datasets commonly suffer from severe class imbalance, causing minority emotions to exhibit sparse and unstable representations in the feature space. Meanwhile, data augmentation methods generate samples that remain highly coupled with original utterances, thereby limiting the diversity of generated samples. In addition, contrastive learning approaches commonly adopted in ERC tasks typically rely on equal-weight averaging or Top-K prototype construction, which is prone to incorporating noise and leads to suboptimal information utilization. To address these challenges, we propose a Weak-Feature Diffusion and Power-Perturbed Contrastive Learning framework (WDCL) for ERC, which introduces Emotion-Aware Diffusive Weak Feature Learning (EDWF) that injects emotion conditions into an Emotion-Aware Conditional Diffusion (EADF) model to reconstruct and enhance weak features in the continuous embedding space, thereby improving the representation quality of sparse samples and the model's discriminative capability. Power-Perturbed Contrastive Learning (PPCL) integrates Gaussian smoothing and power transformation to redefine prototype selection and similarity measurement, thereby mitigating potential semantic drift introduced by EDWF-generated samples during diffusion. This design reduces the influence of semantically shifted samples and enables global optimization of contrastive learning by pulling positive samples closer while pushing negative samples apart. Extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed framework, validating the effectiveness of both EDWF and PPCL in improving representation robustness and discrimination for ERC.

Keywords: Machine Learning, Deep Learning, Distributed Intelligent Systems
Abstract: Personalized federated learning (pFL) mitigates instability and accuracy degradation of standard federated learning (FL) on non-independent and identically distributed (non-IID) data, a common challenge in edge-intelligence deployments where inference is performed locally and model updates are coordinated periodically. We empirically observe that, when decomposing a model into a feature extractor and a classifier, the optimal classifier varies significantly across clients. In contrast, the feature extractor remains relatively stable, implying these components should be treated differently. However, existing pFL methods often overlook this distinction, hindering classifier collaboration and underexploiting classifier knowledge. We propose a novel algorithm, Personalized Federated Learning with Classifier-Centric Mutual Learning (FedCCML), which decouples the optimization of the feature extractor and classifier. A hyperhead synthesizes classifiers from compact edge prototypes, and classifier-latent mutual learning further exploits classifier knowledge to enhance both the hyperhead and feature extractors. Experiments on five real-world datasets show FedCCML consistently improves personalized accuracy under label-skew and feature-skew settings and outperforms state-of-the-art baselines.

Keywords: Image Processing and Pattern Recognition, Deep Learning, Machine Vision
Abstract: 基于草图的3D形状检索（SBSR）旨在通过草图作为查询来检索相关的3D形状。然而，由于草图与三维形状之间存在较大的模态差异，SBSR依然具有挑战性。这一问题因草图绘制过程中产生的噪声而加剧，单个绘画风格和抽象层次的差异常导致视觉表现不一致且失真。为解决这些问题，我们提出了一种噪声感知能量的多级对比学习（NEMCL）框架，用于稳健的跨模态检索。在NEMCL中，引入了基于能量距离的度量，以显式表征草图噪声并将其纳入概率特征表示中，使模型能够适应绘图引起的变异性。此外，设计了一种多层次的关节对比学习策略，通过双向样本中心和中心对中心匹配，逐步对齐草图和三维形状，同时保持层级间的判别语义结构。广泛使用

Keywords: Fault Monitoring and Diagnosis, Intelligent Industrial Systems, Neural Networks and their Applications
Abstract: Reliable fault diagnosis of rotating machinery is essential for intelligent industrial maintenance. However, it faces formidable challenges from unknown fault types and domain shifts induced by varying operating conditions, which is formally formulated as the open-set domain generalization (OSDG) problem. Existing methods are mainly data-driven, thereby overlooking the cascaded propagation of uncertainty across feature extraction, topological learning, and decision-making stages. To address this critical issue, this paper proposes PGU-OD, a novel Physics-Informed Graph Learning framework with Uncertainty Awareness for Open-set Domain generalization. First, it designs a physics-informed spectral attention module to extract condition-robust fault features, thereby suppressing perceptual uncertainty caused by frequency shifts. Further, it constructs an uncertainty-aware adaptive graph learning mechanism to dynamically adjust the edge weights of the sample graph guided by class-scale Gaussian distribution parameters, which mitigates the structural propagation of uncertainty. Finally, a Gaussian-distribution-based adaptive boundary loss function and a dual-criteria open-set inference strategy are developed to optimize decision boundaries and reliably reject unknown faults. Extensive experimental evaluations on two public and widely used rotating machinery fault datasets demonstrate that the proposed PGU-OD outperforms state-of-the-art baselines in both known fault classification and unknown fault rejection under domain shifts.

Keywords: Deep Learning, Representation Learning, Image Processing and Pattern Recognition
Abstract: Few-Shot Class-Incremental Learning (FSCIL) aims to continuously acquire novel knowledge from scarce samples while preserving previously learned concepts. Although recent methods leverage the transferable representations of vision-language models, they still struggle with unreliable prototype estimation and the progressive degradation of decision boundaries as classes accumulate. To address these issues, we propose prototype-guided cross-modal contextualization (PGC), a method that dynamically integrates multi-modal context to construct adaptive representations. We calibrate prototypes using language priors and base-class statistics to rectify statistical bias. Building on calibrated prototypes, we perform contextualized fusion via cross-modal attention to refine instance representations. To prevent decision boundary degradation, we employ virtual compositional boundary supervision, which enforces structural consistency through semantically mixed examples. During inference, we introduce the dual-view complementary inference strategy that jointly exploits visual discriminability and textual generalization to enhance prediction robustness. Extensive experiments conducted on miniImageNet, CIFAR100, and CUB200 demonstrate that our method outperforms existing methods.

Keywords: Machine Learning, Deep Learning, AI and Applications
Abstract: Analytical orbit propagation models are widely used for large-scale space object prediction due to their high computational efficiency. However, their long-horizon accuracy and stability are fundamentally constrained by parameter bias and structural approximations in analytical propagation. These approximations do not generate a single homogeneous error, but rather coupled multi-timescale error components: a slow-varying secular drift associated with accumulated parameter and model biases, and a fast-varying periodic oscillation arising from phase-dependent structural approximation residuals. To address this challenge, we propose a physics-aware dual-end cooperative correction framework (PAC2) for analytical orbit prediction, which jointly models multi-timescale error dynamics while preserving the physical structure of analytical propagation. Experiments on real-world Starlink LEO constellation data show that PAC2 reduces the long-horizon prediction RMSE from 42.4 km (SGP4) to 11.5 km over the 6-72 h prediction horizon, significantly improving prediction accuracy while maintaining the computational efficiency of analytical propagation.

Keywords: Image Processing and Pattern Recognition, Human-Computer Interaction, Generative AI
Abstract: Existing visual cropping methods for finegrained Visual Question Answering (VQA) often struggle to balance accuracy and efficiency. We propose a training-free visual cropping method supporting both fully automatic and user-guided modes within a unified framework. In user-guided mode, coarse annotations constrain the search space, while value features extracted from the Multimodal Large Language Model (MLLM) Key-Value (KV) cache generate precise semantic heatmaps. Incorporating a Gaussian center prior further enables Top-1 direct localization. In fully automatic mode, Value heatmaps are computed over the entire image and Top-K candidate verification is employed for robust localization. Both modes utilize a second-stage refinement mechanism, where selected regions of interest (ROIs) and precisely cropped and augmented red circle visual prompts to enhance final inference. Evaluations on V*Bench demonstrate that our user-guided mode achieves 86.3% accuracy with only 2 forward passes, while fully automatic mode achieves 74.5% accuracy with an average of 6 forward passes (FPs), yielding a 6.0-7.8× speedup compared over prior methods. Ablation studies validate the localization advantage of Value features (99.7% recall vs. 38.2% for Attention weights). Our method is natively compatible with FlashAttention and exhibits strong generalization across diverse base models.

Keywords: Systems Safety and Security, Generative AI, Application of Artificial Intelligence
Abstract: Instruction-driven image editing technology improves content creation efficiency but brings severe copyright challenges. Existing watermarking methods struggle to simultaneously resist semantic-level editing attacks while meeting the dual requirements of original copyright identification and editing traceability. To address this problem, the dual-watermark collaborative framework for instruction-driven editing is designed. To the best of our knowledge, it is the first framework for instruction-driven editing that integrates external watermarking and in-generation watermarking. The external watermarking is designed based on the deep robust watermark framework to protect the original copyright of images. The in-generation watermarking embeds the editor's identity information implicitly during editing through a lightweight modification of the VAE decoder, which is used for edit traceability. During the watermark training process, we introduce image editing as an adversarial attack to enhance the watermark's robustness against semantic-level distortions. Furthermore, we design a joint training strategy, pre-training independently and then optimizing jointly, ensuring that the two types of watermarks do not interfere with each other when working together. Experimental results show that under strong editing attacks such as InstructPix2Pix, the bit error rate (BER) of both the external watermarking and the in-generation watermarking are below 9%, showing stronger robustness compared to state-of-the-art (SOTA) methods.

Keywords: Information Assurance and Intelligence, Systems Safety and Security, Big Data Computing
Abstract: To address the challenges of leakage traceability and accountability for high-value materials science data during offline circulation, this paper proposes a robust watermarking scheme for tabular materials data. Specifically designed for the vulnerability of traditional perimeter-based defenses once data are exported, the proposed method leverages content-based SimHash fingerprinting to enable dynamic mapping and segmented watermark embedding, thereby fully eliminating dependence on physical primary keys and row indices. By further integrating a zero-width character mechanism, the scheme ensures high physical fidelity while completely preserving the original numerical precision and visible semantics of the data. Experiments on real-world materials datasets show that the proposed method not only effectively resists deep semantic tampering, but also demonstrates remarkable robustness against extreme row deletion and arbitrary row reordering attacks. These results provide a secure and high-quality solution for precise traceability and accountability attribution of materials data in offline scenarios.

Keywords: Human-Collaborative Robotics, Assistive Technology, Environmental Sensing
Abstract: Guide dogs offer independence to Blind and Low-Vision (BLV) individuals, yet their limited availability leaves the vast majority of BLV users without access. Quadruped robotic guide dogs present a promising alternative, but existing systems rely solely on the robot's ground-level sensors for navigation, overlooking a critical class of hazards: obstacles that are transparent to the robot yet dangerous at human body height, such as bent branches. We term this the viewpoint asymmetry problem and present the first system to explicitly address it. Our Co-Ego system adopts a dual-branch obstacle avoidance framework that integrates the robot-centric ground sensing with the user's elevated egocentric perspective to ensure comprehensive navigation safety. Deployed on a quadruped robot, the system is evaluated in a controlled user study with sighted participants under blindfold across three conditions: unassisted, single-view, and cross-view fusion. Results demonstrate that cross-view fusion significantly reduces collision times and cognitive load, validating the necessity of viewpoint complementarity for safe robotic guide dog navigation.

Keywords: Machine Learning, Deep Learning, Generative AI
Abstract: Personalized federated learning (PFL) offers a potential solution to the challenge of privacy sensitivity and high heterogeneity of medical data for Large Language Models (LLMs). However, we conduct a systematic preliminary study on the sample alignment issue in federated medical instruction tuning, revealing that data heterogeneity acts as a persistent barrier throughout the tuning process; the majority of local samples fail to benefit from model aggregation during the entire training phase, and this phenomenon is particularly pronounced under extreme non-IID scenarios. Existing PFL methods often rely on architectural modifications, which introduce additional computational and communication overhead while overlooking the intrinsic value of medical data. To address this issue, we propose Personalized Medical Data Identification via Dynamic Enrollment for Federated LLMs (FedMIDE), a data-driven federated instruction tuning method that achieves efficient personalization solely through data-level control. FedMIDE incorporates a dynamic data inclusion strategy to enhance personalization performance. Extensive experiments demonstrate that FedMIDE achieves the highest accuracy among the compared methods, while maintaining excellent robustness and stability.

Keywords: Distributed Intelligent Systems, Generative AI, Optimization and Self-Organization Approaches
Abstract: With the rapid integration of large language models (LLMs) and edge intelligence, real-time decision-making through edge deployment has emerged as a critical research frontier. Edge devices are frequently constrained by limited memory capacity, which hinders the deployment of large-scale models on individual nodes. Although conventional pipeline parallelism alleviates memory pressure, it suffers from severe load imbalance in heterogeneous environments and fails to account for the divergent resource demands of the prefill and decode stages. To address these challenges, this paper proposes EdgePD, a hardware-aware prefill-decode disaggregated inference system tailored for heterogeneous edge clusters. EdgePD leverages an automated hardware profiling module to characterize node capabilities and incorporates a grouping mechanism along with a stage-aware heuristic for layer allocation to ensure balanced workloads across heterogeneous resources. Furthermore, the system introduces a multi-threaded dynamic batching mechanism to effectively hide the network communication overhead through fine-grained concurrency. Experimental results on the ShareGPT dataset using OPT models demonstrate that, compared to conventional unified pipelines, EdgePD improves throughput by up to 3.4× in representative settings, while reducing end-to-end latency.

Keywords: Autonomous Vehicle, Application of Artificial Intelligence, Deep Learning
Abstract: In recent years, multi-view 3D object detection has advanced significantly in autonomous driving scenarios. However, image degradation caused by low light, adverse weather, or motion blur often leads to the loss of texture and structural information, severely impairing the detection of distant or occluded objects. To address this, we propose a Degradation Probability-guided Feature Compensation method. We introduce a Degradation Probability-aware Feature Enhancement (DPFE) module that explicitly estimates degradation and performs targeted compensation through spatiotemporal interactions. Additionally, we construct adaptive 3D queries focused on high-degradation areas to enhance target localization. Finally, a dual-branch foreground feature fusion mechanism dynamically integrates original and enhanced features to balance information preservation and recovery. Experimental results on nuScenes and nuScenes-C show that the proposed method consistently improves performance under various degradation conditions. In particular, under low-illumination scenarios, our method achieves a 3.2% improvement in Resilience Rate over the baseline.

Keywords: Deep Learning, Image Processing and Pattern Recognition, Systems Safety and Security
Abstract: Deep neural networks are vulnerable to adversarial examples, which pose significant security threats due to their transferability across different models in black-box settings. Although input transformation strategies are widely adopted to enhance the transferability of adversarial attacks, existing methods often fail to achieve an optimal balance between geometric diversity and semantic structure. To address this limitation, we propose the Geometric Invariance Attack (GIA), a novel adversarial framework driven by a block-wise perspective and flip mechanism that draws inspiration from physical perspective distortion. Unlike shuffling-based methods that disrupt the global topology of images, this mechanism introduces non-rigid micro-deformations within local blocks while strictly preserving macro-semantic coherence. This strategy leverages the intrinsic geometric invariance of deep neural networks to generate adversarial gradients that exhibit both geometric diversity and semantic consistency. Extensive experiments demonstrate that the proposed GIA significantly outperforms existing baseline methods. Notably, it achieves state-of-the-art success rates against eight advanced defense mechanisms, including adversarial training and purification-based defenses. These results empirically confirm that modeling intrinsic geometric deformations is essential for evading deep purification defenses.

Keywords: Human-Collaborative Robotics, Human-Machine Interaction, Assistive Technology
Abstract: Physical Human-Robot Interaction (pHRI) naturally emerges in collaborative tasks involving force exchange, such as collaborative object transportation, in which robots assist human operators in handling heavy or bulky objects. In such scenarios, force-based information is commonly used to estimate the human's intention of motion. However, relying solely on force-based estimation may lead to ambiguous, if not unpredictable, behaviors, leading to suboptimal coordination and reduced operator comfort. Building upon a baseline force-based interaction framework, the purpose of this paper is thus to experimentally assess to what extent the introduction of an explicit speech-based intention declaration module offers measurable improvements in collaborative transportation. The resulting multimodal human-robot collaboration architecture combines physical interaction forces with spoken commands through an agreement-aware mechanism that safely coordinates verbal and force-based cues. We experimentally evaluate the impact of the additional cognitive channel on a collaborative transportation task involving an industrial robot and multiple human subjects, comparing the baseline force-based information only system against the speech-augmented version. The evaluation combines both objective performance metrics and subjective user feedback collected via questionnaires. The results demonstrate that the introduction of explicit intention declaration through speech reduces user effort, while preserving overall comfort during task execution. These findings prove that adding a cognitive communication channel can effectively complement physical interaction and improve performance and interaction quality in collaborative transportation, leading to a more intuitive and comfortable human-robot collaboration.

Keywords: Image Processing and Pattern Recognition
Abstract: Ultra-high-resolution image segmentation is often associated with significant memory consumption challenges and reduced inference speed. Many networks have been developed to reduce memory consumption until the emergence of ISDNet and its variant networks. However, ISDNet’s adoption of dual independent input introduces inherent structural redundancy, resulting in parameter inefficiency and suboptimal inference speeds. Furthermore, the inclusion of a super-resolution module significantly increases memory demands and degrades speed performance during the training phase. Therefore, we propose SBUPNet, a novel ultra-high resolution segmentation network that effectively reduces structural redundancy by feeding the same input to different branches and sharing partial convolutional layers. In this network, a Detail and Context Enhancement Fusion module (DCEF) is designed to compensate for the information loss caused by convolution operations. To address the issue of excessive training resource consumption, Uncertain Pixel Prediction Loss is introduced to replace the super-resolution methods used in previous studies. Uncertain Pixel Prediction Loss calculates uncertain pixels based on an error-prediction network to enhance detailed information. It does not slow down the inference speed, but greatly accelerates the training speed and reduces the memory overhead. Comprehensive evaluations on the Inria Aerial, DeepGlobe-Land Cover, and DeepGlobe-Road benchmarks demonstrate that SBUPNet provides a highly efficient solution for ultra-high-resolution tasks. In particular, our approach achieves a competitive 76.5 mIoU on the Inria Aerial dataset with an inference speed of 9.49 FPS. Compared to the leading open-source framework ISDNet, SBUPNet not only enhances segmentation accuracy but also delivers a 20% increase in processing throughput. Furthermore, from a system-maintenance perspective, SBUPNet significantly optimizes the training pipeline, reducing both total training duration and peak memory consumption by approximately 35%. This substantial reduction in computational overhead underscores the model’s suitability for rapid deployment and iterative development in large-scale remote sensing systems.

Keywords: Representation Learning, Machine Learning
Abstract: Existing contrastive multiview methods focus on feature learning but neglect inter-sample contexts, limiting feature discriminability. Furthermore, overlooking multi-layered structures leads to distribution conflicts across granularities, from individual instances to global clusters. These issues are exacerbated by static alignment strategies during training that ignore varying view importance, allowing noise from low-quality modalities to corrupt the shared semantic space. To address these issues, we propose a novel framework, Context-enhanced Dynamic Guidance Multiview Clustering with Multi-level Alignment (MCDMC). Initially, the context clustering enhancement (CCE) module models inter-sample contexts to enhance feature discriminability. Subsequently, the dynamic view guidance (DVG) module learns view-specific weights to guide the model to focus on high-quality views during training. Furthermore, the multi-level distribution alignment (MDA) strategy ensures cross-view consistency via hierarchical contrastive learning and distribution constraints. Extensive experiments demonstrate the effectiveness of the proposed framework.

Keywords: Generative AI, System Modeling and Control, AI and Applications
Abstract: Scenario-based requirement analysis is an important technique in agile practice, used to elaborate user stories by systematically analyzing the interactions between the system and its participants.However, traditional scenario analysis relies heavily on manual effort and is constrained by the analyst’s experience, making it difficult for analysts to rapidly construct scenarios that cover diverse interaction branches and to fully uncover potential requirements within a limited time, while such analytical experience is also hard to systematically transfer. The rapid advancement of large language models provides new opportunities for automating this process. This paper proposes TSGRM, a three-stage framework for scenario generation and requirement mining based on large language models. Given user stories as input, the framework simulates the human analysis process to generate structured scenarios and subsequently identifies implicit requirements, helping analysts obtain richer scenario candidates within limited time, thereby supporting rapid iterative communication and requirement validation. The results show that, across 45 user stories from three projects, the proposed method outperforms the baseline in both scenario expansion capability and implicit requirement mining.

Keywords: Application of Artificial Intelligence, Machine Learning, Machine Vision
Abstract: Single image dehazing in maritime environments presents a unique challenge due to the complex coupling of non-uniform atmospheric scattering and high-frequency water surface reflections. Existing methods, primarily designed for terrestrial scenes with homogeneous haze assumptions, often fail to address this domain-specific degradation, resulting in severe artifacts and detail loss. To address this, we propose a novel Physics-Aware Dual-Stream Network, PiD-Net that leverages explicit frequency-based decomposition to restore maritime imagery. Our framework consists of two core contributions: (1) A physics-guided Maritime Haze Synthesis Pipeline (MHSP) that models the distinct characteristics of sea breeze-induced haze and reflection noise; (2) A dual-stream architecture that decomposes the input signal into low-frequency transmission and high-frequency detail components, processing them via specialized streams to effectively suppress reflection interference while recovering global structure. Extensive experiments demonstrate that our approach outperforms state-of-the-art baselines on both synthetic and real-world datasets, achieving superior perceptual quality (0.1652 LPIPS). Furthermore, we validate the practical utility of our method for downstream maritime surveillance, showing that PiD-Net preprocessing significantly enhances object detection performance, improving the Recall rate by up to 11% compared to raw inputs.

Keywords: Application of Artificial Intelligence, Intelligent Industrial Systems, Machine Learning
Abstract: Ballot integrity relies on accurately detecting unauthorized marks, yet traditional template matching fails un der complex textual–graphical backgrounds, and Transformer based detectors suffer from high latency and weak robust ness to dense small anomalies. We propose MAGC-YOLO, a YOLOv11-based real-time detector integrating a Multi-scale Aggregation with Cross Attention (MACA) module, a Text Suppression Guidance Module (TGSM), and a Geometric Constraint IoU (GC-IoU) loss augmented by PolyLoss to handle class imbalance. We also release PASBD, the first public ballot scribble dataset (4,800 images). Experiments on PASBD and a private 1,500-image dataset show MAGC-YOLO outperforms YOLOv11-s by +14.6% mAP50 / +13.5% mAP50:95 (PASBD) and +23.4% / +25.7% (private), while maintaining real-time speed (> 87 FPS on RTX 3090). Results confirm its superiority for industrial ballot inspection.

Keywords: Human-Machine Interaction, Deep Learning, Machine Learning
Abstract: Speaker diarization aims to determine "who spoke when" in multi-speaker recordings. In practical multi-party human-robot interaction (HRI) scenarios, diarization systems must operate in a streaming manner under adverse far-field conditions with strong reverberation and background noise, which remain challenging for existing online methods. This work proposes RS-EEND, a robust streaming diarization system that builds upon an end-to-end neural diarization backbone and incorporates a Mel-spectrogram enhancement front-end. The enhancement module directly predicts enhanced Mel-spectrograms for downstream diarization and is adapted for low-latency online processing. We investigate diarization-oriented enhancement targets, including direct-path speech, reverberant speech, and reverberation time shortening (RTS), and propose a speaker-aware enhancement training strategy. Experiments on simulated data and the real-world AMI corpus demonstrate that: 1) RTS provides a better trade-off by suppressing late reverberation and preserving useful early reflections for speaker discrimination; 2) the enhanced Mel-spectrograms can be directly used to improve far-field online diarization, and the proposed speaker-aware training strategy further improves performance. Furthermore, the proposed RS-EEND approach outperforms existing online diarization methods on both simulated and AMI datasets.

Keywords: Information Assurance and Intelligence, Systems Safety and Security, System Architecture
Abstract: Self-sovereign digital identity (SSDI) returns control of personal data from centralized authorities to individual users, addressing growing concerns over large-scale data breaches and identity theft. However, current SSDI protocols predominantly rely on smart contracts deployed on public blockchains, inheriting their well-known limitations in privacy and throughput scalability. In this paper, we present BYOW (Bring Your Own Witness), a privacy-preserving, general-purpose transactional protocol for N self-sovereign participants. BYOW introduces a decentralized witness network in which witnesses attest to transaction occurrence without learning any transaction details, requiring zero durable distributed storage. Transactions and their associated state transitions are negotiated directly among participants, decoupling throughput from global consensus overhead. We show that BYOW provides Byzantine fault tolerance under the condition that the intersection of all participants' trusted witness sets is non-empty. We implement the protocol in Rust and evaluate it on a cluster of 42 Raspberry Pi 4 nodes across 125 experimental configurations spanning 2-32 witnesses, 2-32 concurrent clients, and 64-1024 request batches. Our experiments demonstrate that per-node throughput increases with request volume, that latency grows linearly under exponentially increasing loads, and that per-node communication overhead remains constant as the witness network scales. While BYOW assumes a known set of N participants rather than an open-membership model, it achieves strong scalability and privacy guarantees without relying on any distributed ledger.

Keywords: Machine Vision, Visual Analytics, Communication, Human-centered Learning
Abstract: This study investigates the impact of encoder architectures and optimization algorithms on the performance of U-Net-based models for human segmentation. A comprehensive evaluation was conducted using EfficientNet, ResNet, and MobileNet encoders combined with ten optimization algorithms. Model performance was assessed using training and validation loss, along with Intersection over Union (IoU). Descriptive statistics and visual analyses were employed to compare encoder-optimizer combinations. Furthermore, a one-way Analysis of Variance (ANOVA) was performed to examine the statistical significance of performance differences. Among all encoder–optimizer combinations, EfficientNet-B6 combined with the RAdam optimizer achieved the highest validation IoU (0.871), whereas EfficientNet-B0, EfficientNet-B1, and EfficientNet-Lite0 consistently provided competitive performance with substantially lower computational complexity, offering a better balance between segmentation accuracy and efficiency. Statistical analysis confirms that encoder selection has a significant impact on segmentation performance (p < 0.001), whereas differences among optimizers are not statistically significant (p > 0.05). Overall, the findings demonstrate that encoder architecture plays a more critical role than optimizer choice in determining segmentation accuracy, highlighting the effectiveness of efficient encoder backbones for human segmentation tasks.

Keywords: Machine Vision, Deep Learning, Machine Learning
Abstract: Multi-Head Self-Attention (MHSA) is the cornerstone of Vision Transformers, allowing models to capture diverse feature representations by projecting tokens into independent subspaces. However, the standard MHSA strictly isolates these heads, preventing any information exchange during the attention computation itself. In this paper, we propose Multi-Overlapped-Head Self-Attention (MOHSA), a novel mechanism that replaces the hard division of attention heads with a soft, overlapping division. By allowing queries, keys, and values to partially overlap with those of adjacent heads, MOHSA fosters rich inter-head communication directly within the attention mechanism. We extensively investigate various overlap dimension scheduling strategies to identify optimal configurations. Comprehensive experiments in several Transformer architectures demonstrate that MOHSA outperforms standard MHSA on the CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1k datasets, offering a significant performance boost with negligible computational overhead. Code link: url{https://github.com/ZTX-100/MOHSA}.

Keywords: Systems Safety and Security, Smart Sensor Networks, Deep Learning
Abstract: Multivariate time-series (MTS) anomaly detection remains challenging. This task requires learned representations to capture both temporal dynamics and spatial dependencies among sensors. In existing methods, these two views are often modeled in an insufficiently coordinated manner, which limits representation quality and degrades detection performance. To address this issue, we propose SDFT-AD (Self-supervised learning with Differentiated Fine-Tuning for multivariate time-series Anomaly Detection), a two-stage framework for MTS anomaly detection that combines cross-dataset self-supervised pre-training with differentiated fine-tuning. SDFT-AD consists of two key components: 1) a cross-dataset self-supervised pre-training scheme on a shared dual-branch backbone, which separately extracts temporal and spatial features to learn transferable representations; and 2) a differentiated task-weight schedule during fine-tuning, which reconstructs data from both views and adaptively adjusts task weights to better capture temporal-spatial interactions. Extensive experiments on four public MTS benchmarks show that SDFT-AD consistently outperforms strong baselines, demonstrating the effectiveness of the proposed framework for robust MTS anomaly detection.

Keywords: Systems Safety and Security, Machine Learning, Multimedia Systems
Abstract: ，联邦学习容易受到后门影响 攻击。然而，传统的辩护通常依赖于 未能识别的粗粒度全球指标 恶意更新，因为隐蔽攻击者可以明确地 限制他们的更新以规避侦测。此外， 这些防御的表现受到了显著影响 非独立且完全相同的降级 分布式场景。我们观察到恶意更新 后门表现出判别性相对异常 参数尺寸。基于这一见解，我们提出以下建议 AlphaGuard，一种利用后门检测方法 细粒度的维度相对放大。具体来说，AlphaGuard 通过 将客户端更新与构造引用进行比较，以及 根据 由此产生的放大模式。广泛的实验 证明AlphaGuard的表现远超大多数 现有的防范多种后门攻击，包括像F3BA和A3FL这样的自适应攻击，甚至在 数据异质性高的情景。

Keywords: Reinforcement Learning, Robotic Systems, Multimedia Computation
Abstract: Audio-Visual Navigation (AVN) requires an agent to localize and navigate toward a continuously vocalizing target relying solely on visual observations and acoustic cues. Currently, systems lack the ability to adaptively correct and replan when faced with incomplete or misleading visual perception. Furthermore, relying on physical collisions to compensate for missing visual information results in inefficient and unsafe navigation, whereas existing methods are overly dependent on passive visual perception. To address these issues, we propose the Transformer-based Token Fusion and Dynamic Graph Planning (TDGP) model, which incorporates high-level perception layers and leverages the Transformer model to fuse multimodal cues for precise local planning. Next, a low-level planning layer was designed that uses physical collision penalties to remove edges that collide with the map in real time and apply corresponding penalties, forcing the agent to automatically re-plan to compensate for the lack of visual information. Experiments show that our TDGP model outperforms baseline models on the Replica and Matterport3D(MP3D) datasets, and that the model’s sound enhancement strategy significantly improves generalization in unheard acoustic scenarios.

Keywords: Human-Machine Interaction, Human-Collaborative Robotics, Application of Artificial Intelligence
Abstract: ，人类习惯性地描述问题并下达命令 在沟通中使用歧义引用。现存 人机交互方法难以解释 自然语言指令中的模糊性。这篇论文 提出一个智能人机交互引擎 对这些命令的精准理解。我们引入 多轮对话互动方式与 人类习惯。它利用了大型企业的能力 用于分析自然语言指令的语言模型。作者 整合机器人视场的视觉信息 通过这些观点，模糊的元素得以澄清 对话。该过程实现了意图识别和 生成清晰、格式化的命令序列，包含显式 需求和目标对象。实验结果显示 我们的引擎能够准确解析人类的意图 命令。采用多轮对话，它获得了93%的评分 复杂指令的识别率。模糊之后 机器人的适应性和执行效率得到了澄清 在无结构环境中，效果显著增强。对É

Keywords: Systems Safety and Security, Generative AI
Abstract: System code vulnerability detection facilitates the timely identification and remediation of security flaws, thereby mitigating potential attacks, safeguarding data and system integrity, and preventing significant losses. Recently, Retrieval-Augmented Generation (RAG) frameworks have been adopted to perform security checks on software source code. However, conventional retrieval methods within RAG often yield suboptimal results because standard embeddings treat code merely as natural language. To address this, this paper proposes a method based on Code Property Graphs (CPGs) for structural representation. By employing graph embedding techniques, code syntax, control flow, and data dependencies are transformed into structure-aware vectors. Compared to variants without graph embeddings, our method increases the F1 score by 7.89 percentage points. Additionally, a hierarchical prompt strategy is designed to guide Large Language Models (LLMs) in accurately localizing and reasoning about vulnerabilities. This effectively enhances the accuracy of LLMs in code security detection, with our method achieving an optimal F1 score of 76.18%.

Keywords: Environmental Sensing, Smart Sensor Networks, Robotic Systems
Abstract: This paper presents Range-SLAM, a real-time, lightweight SLAM system designed to address the challenges of simultaneous localization and mapping (SLAM) in environments with smoke and other harsh conditions using Ultra-Wideband (UWB) signals. Since optical sensors like LiDAR and cameras struggle in low-visibility environments, UWB signals provide a robust alternative for real-time positioning. The proposed system uses general UWB devices to achieve accurate mapping and localization without relying on expensive LiDAR or other dedicated hardware. By utilizing only the distance and Received Signal Strength Indicator (RSSI) provided by UWB sensors in relation to anchors, we combine the motion of the tag-carrying agent with raycasting algorithm to construct a 2D occupancy grid map in real time. To enhance localization in challenging conditions, a Weighted Least Squares (WLS) method is employed. Extensive experiments, including smoke-filled environments and simulated scenarios, demonstrate the efficiency and robustness of the proposed system.

Keywords: Systems Safety and Security, Deep Learning, Machine Learning
Abstract: Detecting anomalies in network traffic is critical for identifying and mitigating cyberattacks. However, most existing approaches rely on large amounts of labeled data for supervised training, which is commonly impractical. Moreover, as attack measures continuously evolve, unknown attacks pose severe threats to cyberspace. To address these challenges, we propose a novel one-class classification framework for network intrusion detection. Our hybrid framework integrates a Variational Autoencoder (VAE) to effectively compress network traffic with the Deep Support Vector Data Descriptor (Deep SVDD) algorithm for anomaly detection, where Deep SVDD inherits the encoder of VAE. By leveraging the VAE’s advanced ability to learn data distributions, our approach alleviates the hypersphere collapse problem in Deep SVDD. The proposed method is trained solely on normal traffic in an unsupervised manner. Consequently, it reduces the reliance on labeled data and is capable of detecting previously unknown attack types. Experimental results on the three commonly used public datasets, including NSL-KDD, UNSW-NB15, and CICDDoS2019, demonstrate that our framework outperforms existing methods across multiple evaluation metrics.

Keywords: Deep Learning, Neural Networks and their Applications, Intelligent Industrial Systems
Abstract: Multivariate Time Series (MTS) forecasting is crucial for the condition monitoring and decision support of complex systems, playing a vital role in domains such as modern power grid scheduling, industrial asset maintenance, and financial risk early warning. However, due to the complex dynamic interactions within real-world systems, the collected data commonly exhibit multiple periodic structures and com- plex non-stationary fluctuations, posing significant challenges to accurate MTS forecasting. To address these challenges, this paper proposes a novel dual-branch forecasting framework, DBMFNet. The framework comprises two core components: (i) the Multi-scale Period Modeling (MPM) branch, which reorganizes 1D sequences into 2D tensors to explicitly decouple intra-period variations and inter-period trends, and synergizes 2D convolutions with a Transformer to capture multiple pe- riodic dependencies, effectively overcoming the limitations of traditional models in identifying multiple periodicities; and (ii) the Multi-scale Frequency Modeling (MFM) branch, which adaptively adjusts the spectral decomposition resolution based on energy density, and integrates independent normalization with channel attention mechanisms to achieve precise charac- terization of complex non-stationary features. Extensive exper- iments on eight real-world long-term time series benchmark datasets demonstrate that DBMFNet achieves strong predictive performance across various forecasting scenarios.

Keywords: Neural Networks and their Applications, Intelligent Industrial Systems, Computational Intelligence
Abstract: Single-source anomaly detection frameworks struggle to capture the complex, early precursors of cattle diseases triggered by the coupling of environmental stress and individual physio-behavioral deviations. To address this, we propose a weakly supervised early warning model based on a Multi-source Spatiotemporal Attention Fusion (MSTAF) mechanism. We construct an Environmental Stress Index (ESI) and a Health Deviation Score (HDS) using high-frequency rumen sensor data, environmental readings, and individual metadata to quantify multimodal interactions. We design a multi-channel spatiotemporal attention network to decouple disease-relevant historical windows and modal contribution weights through temporal and channel attention. Validated on continuous monitoring data from 120 Holstein cows over 12 months in a commercial farm, MSTAF achieves a sensitivity of 93.2% and a specificity of 90.5% for early warnings 48 hours prior to clinical onset, operating without explicit pathological labels. Its discriminative capacity and robustness significantly outperform traditional fusion baselines, establishing an interpretable multi-source monitoring framework for precision livestock farming.

Keywords: Machine Learning, Robotic Systems, Adaptive Systems
Abstract: Test-time adaptation (TTA) is essential for improving robot tactile perception in real-world settings. However, current TTA methods are predominantly designed for visual datasets and face two major challenges when applied to tactile tasks: (1) the deep structured coupling between tactile textures and environmental noise; and (2) the severe class imbalance inherent in data streams generated during robotic interactions. To address these issues, we propose T^4, a trustworthy test-time adaptation framework for robust tactile perception. We first introduce Debias Batch Normalization to ensure fairness by balancing class-wise statistics under imbalanced streams. Simultaneously, we design the Geometric Noise Decoupling module, which leverages singular value decomposition to isolate common noise directions and achieves feature purification through orthogonal projection. Furthermore, we propose a Noise-Calibrated Trustworthy Sampling strategy to manage the memory bank based on noise intensity and semantic confidence, ensuring stable model updates. We evaluate the reliability of our approach on several classic tactile datasets with algorithmic corruptions and construct RTM15-C, the first tactile material dataset featuring real-world interferences. Compared to existing state-of-the-art methods, T^4 achieves an accuracy improvement of at least 5.0% across all benchmarks, thereby providing a systematic and robust solution for tactile material recognition in complex field scenarios. Our code and datasets will be made publicly available upon the acceptance of this paper.

Keywords: Machine Learning, Image Processing and Pattern Recognition, Representation Learning
Abstract: Time series classification has been widely studied, yet many existing approaches struggle to generalize across diverse datasets and representation forms. In this work, we conduct a comprehensive empirical study on transforming time series into image representations and leveraging state-of-the-art pretrained visual models for classification. Four visual encoding techniques---Line, Gramian Angular Field, Markov Transition Field, and Recurrence Plot---are applied to convert univariate sequences into two-dimensional images. Six pretrained models, including ConvNeXt, MLP-Mixer, DINOv3, Swin Transformer, ResNet, and ViT, are used to extract deep visual embeddings. These features are subsequently classified using multiple algorithms, including SVM, Random Forest, and ensemble learning strategies. Experiments on 128 UCR datasets demonstrate that visual-based representations yield competitive performance compared to existing time-series classifiers such as ROCKET, Catch22, and ShapeletTransform. The proposed framework achieves a mean stacking accuracy of 0.7753, confirming the effectiveness of visual representations for time series classification and providing a solid benchmark for future research.

Keywords: Soft Computing, Socio-Economic Cybernetics, Human-centered Learning, Game Theory and Multi-Agent Systems
Abstract: In the short-video ecosystem, platforms, creators and users are interdependent, forming a people-centered complex interactive system. As a core regulatory strategy, the design of innovation incentive mechanisms directly determines the sustainable development and operational resilience of the digital content ecosystem. Therefore, this paper constructs a three-party Stackelberg game model to obtain optimal strategies at equilibrium, with conclusions verified by numerical simulation. The research findings are as follows: (1) Platform innovation incentives have significant strategic value. Reasonable subsidies can enhance the creators' optimal innovation degree, driving the coordinated growth of user utility and platform revenue; (2) The advertising sharing ratio and incentive intensity have a dynamic trade-off relationship. The platform needs to anchor the sharing threshold to flexibly adjust the incentive strategy; (3) The user innovation sensitivity coefficient is a key regulating variable for the incentive effect. The incentive intensity is not necessarily the stronger the better; blindly increasing it will lead to a decrease in the creators' revenue; (4) Innovation incentives can effectively resist the excessive intervention of advertisers in the short-video ecosystem, maintaining user experience and the long-term interests of creators. This research provides theoretical basis and management implications for the design of incentive mechanisms in short-video platforms.

Keywords: Computational Intelligence in Information, Application of Artificial Intelligence, Deep Learning
Abstract: Cytochrome P450 (CYP450) inhibition is a major cause of adverse drug-drug interactions and a key consideration in early-stage drug discovery. However, accurate CYP450 inhibitor identification remains challenging because individual molecular representations often capture only partial chemical information, limiting predictive performance. Here, we develop FDG-CYP, a multimodal prediction framework that integrates complementary molecular representations derived from fingerprints, descriptors, and graph structures for CYP450 inhibitor prediction. Using a large-scale dataset comprising more than 13,000 compounds from the PubChem BioAssay database, FDG-CYP achieves competitive performance relative to representative baseline methods across balanced accuracy, Matthews correlation coefficient, and F1-score. Ablation studies further reveal the complementary contributions of different molecular modalities to predictive performance. FDG-CYP therefore provides an effective approach for distinguishing CYP450 inhibitors from non-inhibitors and may facilitate the early identification of compounds with potential metabolic safety liabilities.

Keywords: Soft Computing, Socio-Economic Cybernetics, Game Theory and Multi-Agent Systems, Consumer and Industrial Applications
Abstract: As product functions grow increasingly complex, consumer electronics and home appliance firms face a strategic dilemma in the voluntary disclosure of non-mandatory preventive measure information. In the context of human-centric operational management, this study incorporates consumer perception to explore firm disclosure strategies. This paper introduces the consumer perceived effect coefficient into a duopoly Hotelling model to analyze the optimal disclosure strategies of firms with different quality levels and examine the impacts of disclosure on pricing, profits and social welfare. The results show that preventive measure disclosure is a core competitive instrument for duopolies, whose effect depends mainly on the consumer perceived effect coefficient. When rivals disclose, high-quality firms can strengthen their premium advantages, while low-quality firms achieve catch-up in both quality and profit. Social welfare increases monotonically once the perceived effect coefficient exceeds the critical threshold. Firms can realize a win-win between competitive advantage and social welfare by adopting differentiated disclosure strategies based on their own quality. This study enriches research on non-mandatory information disclosure in operations management and provides theoretical support for firm decisions and industry order optimization.

Keywords: Human Factors, Human Performance Modeling, Machine Learning
Abstract: In autonomous driving scenarios, the occurrence of motion sickness symptoms becomes more frequent, which significantly affects passenger ride comfort. Existing studies on ride comfort evaluation in autonomous driving mainly adopt data-driven methods from two perspectives: vehicle motion states and passenger physiological responses. However, methods based on vehicle motion states cannot adequately reflect individual differences in passenger perception, while physiological-response-based methods are often influenced by multiple factors and show relatively weak associations with driving behavior. As a result, there is still room for improvement in ride comfort evaluation performance. To address this issue, this paper develops a ride comfort evaluation framework from the mechanism of motion sickness generation based on the Sensory Conflict Theory. First, a perceptual modeling approach is employed to compute sensory conflict from vehicle motion states, which serves as an internal perceptual representation of motion inconsistency. Then, electroencephalography (EEG) and electrocardiography (ECG) signals are incorporated to introduce passenger-specific physiological response characteristics. Finally, a multimodal fusion model, termed SCFusion, is constructed using sensory conflict, EEG, and ECG signals as inputs for comfort evaluation. To support this study, a multimodal ride comfort dataset, termed the MM-RC dataset, is constructed under a driving simulation environment. Experimental results show that, compared with raw IMU data, the sensory conflict representation more effectively reflects motion discomfort severity and provides greater information gain in the ride comfort evaluation task. Moreover, SCFusion achieves better prediction performance than single-source models, reaching an accuracy of 87.16% in the five-class classification task.

Keywords: Deep Learning, Machine Vision, Human-centered Learning
Abstract: Student behavior recognition in classrooms is pivotal for smart education, yet achieving real-time performance without compromising accuracy remains a significant challenge. Although Transformers possess superior global modeling, their computational overhead often hinders latency-sensitive deployment. To address this, we propose RERC-DETR, a lightweight multi-scale Transformer detector optimized for classroom analytics. Specifically, we introduce RAD-Net, an efficient backbone designed to minimize redundancy, and a Recalibrated Cross-Scale Feature Pyramid Network (RCFPN) that preserves high-resolution semantics and resolves feature aliasing. Furthermore, we introduce a Quality-Aware Logical Distillation strategy utilizing teacher confidence to selectively transfer structural knowledge, enhancing the student detector without inference overhead. Experiments on STBD-08 and SC7 datasets demonstrate that RERC-DETR achieves an optimal accuracy-efficiency trade-off. Compared with the baseline RT-DETR, our model reduces the parameter count by 44% to only 11.2M, while attaining 93.9% mAP50 at 156 FPS on STBD-08 and 89.7% mAP50 at 80 FPS on SC7, outperforming multiple SOTA detectors under a compact parameter budget.

Keywords: Machine Learning, Deep Learning, Machine Vision
Abstract: Audio-visual embodied navigation aims to enable robots to locate audio sources by fusing visual observations and auditory signals from the target. The core challenge lies in building effective semantic interactions between heterogeneous visual and auditory modalities. Existing methods typically use simple feature fusion strategies, which fail to capture the inherent geometric and semantic connections between the two modalities, leading to information loss in complex scene environments. To address these issues, we propose the Cycle-Temporal Attention Network (CTAN), which focuses on active semantic-enhanced fusion rather than basic multimodal combination. Specifically, our framework introduces the Audio-Visual Reconstruction Cross-Attention (AVRCA) module: it uses a bidirectional cycle-consistency constraint to mutually enhance the spatial semantic attributes of visual and auditory features, promoting more robust multimodal interaction. Additionally, we design a Temporal Cross-Modal Memory (TCMM) mechanism to dynamically integrate real-time enhanced multimodal features with historical context, reducing performance drops caused by auditory dead zones. Extensive experiments on the Replica and Matterport3D datasets demonstrate that our method outperforms existing audio-visual navigation approaches in success rate (SR), path efficiency (SPL), and scene adaptation (SNA).

Keywords: Reinforcement Learning, Application of Artificial Intelligence, Logistic and Manufacturing Automation
Abstract: The two dimensional irregular bin packing problem has broad applications across various industries. This paper proposes an end-to-end deep reinforcement learning framework to generate packing solution for two-dimensional irregular items. The packing process is decomposed into three sequential steps: item selection, rotation, and placement. To enable direct prediction of placement positions, we introduce a structure-aware placement strategy, which pre-generates a set of candidate placement positions and uses the model to select the most suitable one for the current item. Additionally, a state fusion module is designed to facilitate the model in comprehending environmental features and to enable more appropriate decision-making. Experimental results demonstrate that the proposed method achieves competitive performance on two dimensional irregular bin packing tasks.

Keywords: Deep Learning, Machine Learning, Application of Artificial Intelligence
Abstract: Fully Few-shot Class-incremental Audio Classification (FFCAC) requires audio models to learn each session from very few labeled examples, where lightweight closed-form updates often lose feature-space separability and heavier adaptation incurs high cost. We propose RAFTER-Net, an AST-based framework that combines parameter-free Lightweight Attention Fusion (LAF) with support-set Adaptive Feature Refinement (AFR). LAF fuses multi-layer AST features through dot-product weights, while AFR calibrates labeled support features by prototype attraction, nearest-negative repulsion, and uncertainty-aware smoothing before closed-form classifier updates. On NSynth-100 and LS-100, RAFTER-Net outperforms strong FFCAC baselines; on NSynth-100, it improves average accuracy over MAR by 16.0 percentage points and reduces performance dropping by 9.33 percentage points, with no learnable fusion parameters and low additional backbone-level MAC/parameter overhead.

Keywords: Complex Network, Social simulations, Cyber-physical systems
Abstract: The governance of online public opinion has become a critical challenge in modern socio-technical systems. While prior studies acknowledge the role of emotion in opinion formation, the impact of emotional intensity on large-scale opinion dynamics remains insufficiently understood. To address this gap, this paper proposes a computational framework that integrates Emotional Social Sharing and Emotional Contagion to model the role of emotional intensity in online social networks. The proposed model captures the co-evolution of individual emotions and opinions, where emotional intensity acts as a nonlinear amplifier of social influence in information diffusion processes. Simulation results on both synthetic and real-world networks show that emotional self-confidence and opinion thresholds jointly shape convergence patterns, while increased emotional intensity significantly delays consensus formation and amplifies system fluctuations. Moreover, a small proportion of agents with extreme emotional states can disproportionately destabilize the system, leading to heightened polarization risks. These findings identify emotional intensity as a key driving factor in opinion dynamics and provide a systematic framework for understanding and regulating opinion evolution in online environments.

Keywords: Machine Learning, Big Data Computing
Abstract: 基于会话的推荐旨在预测用户的下一个内容 基于匿名且简短的行为序列的动作。 现有方法常常面临数据稀疏性和 本地会话中的无意用户点击。 此外，大多数最新型号会压缩整个会话 转换为单一表示向量，或依赖重型 用于功能对齐的辅助任务，严格来说 限制了模型的表现力，并未能捕捉 用户兴趣多样且多方面。在这篇论文中， 我们提出了一种名为时空的新方法 多意图提取与协同去噪 （中至中）。具体来说，我们首先介绍一个协作组织 去噪模块，利用全局超图 注入高阶协作结构先验， 根据粗会话自适应地衰减噪声项目 意图。接下来是双路径时空建模模块 处理纯化后的表征，使用门控 图神经网络（GGNN）与并行变换器 捕捉细粒度的局部'

Keywords: Image Processing and Pattern Recognition, Machine Learning, Intelligent Industrial Systems
Abstract: Visual anomaly detection is crucial for industrial inspection. However, feature-based methods such as Reverse Distillation (RD) are limited by the domain gap between natural and industrial images, as well as the student’s excessive alignment with teacher features, which may reduce its sensitivity to anomalies. To address these issues, we propose Domain-Adaptive Reverse Distillation (DARD), a unified framework that jointly performs feature-domain adaptation and anomaly-aware student distillation. Specifically, we introduce a lightweight Convolutional Feature Adaptation (CFA) module together with a mask-guided contrastive adaptation objective to align normal industrial features while enhancing feature discrepancies in anomalous regions. Furthermore, we develop an anomaly-aware distillation mechanism that integrates a Fusion Bottleneck (FBN), Distribution-Specific Compensation (DiSCo), and denoising constraints to preserve anomaly-relevant information and prevent excessive teacher–student alignment. Experiments on four industrial anomaly detection benchmarks demonstrate that DARD achieves leading overall performance among the compared methods in both anomaly detection and localization.

Keywords: Neural Networks and their Applications, Deep Learning, Representation Learning
Abstract: Heterogeneous graph neural networks (HGNNs) effectively capture complex semantics in graphs with multiple node and edge types, but they often suffer from high computational costs and memory limitations when applied to large-scale graphs. Although precomputation-based paradigms have been introduced to improve scalability, existing methods still face challenges such as insufficient semantic extraction, overly simplified feature compression, and limited robustness. To address these issues, this paper proposes PEPHGNN, a Perturbation-Enhanced Heterogeneous Graph Neural Network based on pre-computation. In the precomputation stage, the model constructs multi-view semantic features through multi-path and multi-hop propagation and introduces a Multi-round Perturbation-Enhanced Projection (MPEP) module to achieve efficient dimensionality reduction while enhancing feature diversity and robustness. In the training stage, a lightweight architecture combining local semantic encoding and global semantic fusion is employed to model intra-path multi-hop information and inter-path complementary semantics. Experimental results on four small-scale and three large-scale benchmark datasets demonstrate that PEPHGNN consistently outperforms existing end-to-end and precomputation-based methods, achieving superior performance, better scalability, and faster convergence.

Keywords: Machine Learning, Deep Learning, Energy Systems
Abstract: ，准确的容量预测对健康至关重要 锂离子电池的管理和可靠运行。

现有方法通常模拟 耦合方式，使得缓慢区分变得困难 由于短期波动和 跨变量相互作用，限制了预测 不同预测视野的表现。 为解决这个问题，我们提出了DFTformer，其中 将输入序列分解为趋势序列和残差序列 组件，建模低频趋势 频域，并利用 频率增强的变压器。 公共磷酸铁锂电池的实验 数据集显示，所提方法能够实现强有力的效果 在不同预测视野内的性能 保持相对较低的计算成本。

Keywords: Evolutionary Computation, Swarm Intelligence, Optimization and Self-Organization Approaches
Abstract: Heuristic algorithms are popular for solving complex problems such as path planning and resource scheduling, but their design relies heavily on human expertise. Recent studies have shown that heuristic algorithm design can be automated by combining large language models (LLMs) with evolutionary computation, where LLMs generate candidate programs and evolutionary evaluation selects improved ones through multiple iterations.However, existing methods make limited use of the differences between generated programs, restricting the program diversity and search guidance. To address this issue, this paper proposes a semantic differential evolution (SDE) framework for automated heuristic algorithm design, which extends differential evolution into the code semantic space. SDE introduces a hybrid semantic distance to evaluate program diversity by combining the code structural features with deep semantic representations. A quality-diversity semantic elite archive is built to preserve diverse high-quality programs. Based on the archive, three semantic differential operators, i.e., directed, dialogue, and drift operators, are developed to support exploitation, heuristic recombination, and exploratory perturbation. Experimental results on three widely used benchmarks, i.e., circle packing, multi-objective traveling salesman problem, and online bin packing, show that SDE outperforms existing LLM-driven baselines in terms of solution quality.

Keywords: Big Data Computing, Machine Learning, Representation Learning
Abstract: Graph fraud detection plays a pivotal role in safeguarding the security and integrity of modern digital ecosystems. Graph Neural Networks (GNNs) are commonly adopted for graph fraud detection. However, the practical performance of existing GNN-based detectors is severely hindered by incomplete node attributes and extreme class imbalance within graphs. To mitigate these limitations, this paper proposes a novel framework for Graph Fraud Detection with Grouped attribute completion and Confidence-aware Contrastive learning, named GFD-GC. Specifically, it first imitates heterogeneous neighborhood structures to implement group-wise aggregation, which obtains informative complete node features by capturing fine-grained graph contextual patterns. Further, it introduces a confidence-aware supervised contrastive learning strategy to augment scarce labeled fraud nodes with high-confidence pseudo-fraud nodes, which enhances the compactness of fraud representations and their separability from non-fraud nodes. Extensive experiments demonstrate the superiority of the proposed GFD-GC over state-of-the-art baselines on the graph fraud detection task, thereby providing an effective solution for real-world fraud scenarios.

Keywords: Human-Machine Interaction, AI and Applications, Application of Artificial Intelligence
Abstract: Speech-based moderation pipelines increasingly rely on text-to-speech (TTS) and speech-to-text (STT) components, yet their combined impact on robustness, bias, and downstream classification remains underexplored. In this paper, we systematically evaluate TTS–STT pipelines for toxic speech detection, focusing on the effects of speech type (synthetic vs. human), model pairing, and gender. Our results show that synthetic speech is generally easier for STT models to transcribe than human audio, but introduces systematic biases that vary across TTS–STT combinations. We further find that toxic content is substantially more sensitive to transcription errors, leading to increased classification instability. In addition, human perceptual preferences for synthetic speech do not align with model-level performance, revealing a divergence between human-likeness and machine understanding. These findings highlight critical robustness, fairness, and reliability challenges in speech-based moderation systems and underscore the importance of jointly evaluating TTS and STT components.

Keywords: Application of Artificial Intelligence, Deep Learning, Environmental Sensing
Abstract: 3D human pose estimation in in-bed scenarios is an important task with applications in rehabilitation care and smart homes. Pressure-sensitive bedsheets offer an emerging solution, providing advantages over RGB-based methods, including enhanced privacy and occlusion robustness. Pose estimation using pressure-sensitive bedsheets is emerging as a promising alternative approach. However, current approaches face two key limitations: (1) labels are typically generated from RGB data without leveraging pressure information, leading to 3D spatial errors due to occlusion and depth ambiguity; and (2) existing models follow vision-centric pipelines without fully exploiting pressure-specific features. To address these issues, we propose: (1) a label optimization method aligning RGB-generated labels with pressure cues for realistic outputs, and (2) a pressure-aware feature extraction pipeline and loss function. On the TIP dataset, our method improves label-to-pressure coverage (bed plane) from 29.2% to 94.9% and reduces MPJPE from 79.2,mm (original) to,56.9 mm. Our work introduces pressure-aware optimization and tailored architectures, advancing the prospects for pressure application.