The exponential growth of multimedia data necessitates advanced cross-modal retrieval methods capable of bridging the semantic gap between heterogeneous modalities, such as images and text. Although deep cross-modal hashing techniques have demonstrated strong performance by jointly integrating feature extraction and hash code generation, existing approaches often struggle to generate genuinely unified hash codes. These limitations mainly arise from modality-specific feature discrepancies and insufficient mechanisms for robust semantic alignment. To address these challenges, we propose a Cross-Attention Aware Fusion-based Cross-Modal Hashing (CAFH) method. The CAFH framework introduces a cross-attention aware fusion module that effectively captures and integrates shared semantic information across modalities, thereby producing coherent representations of semantically related data points. Unified hash codes are subsequently generated from these integrated features, improving retrieval accuracy in cross-modal hashing tasks. Furthermore, the model incorporates semantic similarity learning to enhance the separability of dissimilar samples, thereby improving both robustness and retrieval precision. Extensive experiments on three benchmark datasets, namely MIRFLICKR25K, NUSWIDE-10K, and MSCOCO, demonstrate the superior performance of the proposed method in both image-to-text and text-to-image retrieval tasks. By achieving state-of-the-art results and addressing key limitations of existing methods, CAFH provides a robust and effective solution for large-scale multimedia retrieval.

Recent advances in foundation models and AI agents are shifting mobile AI systems from fixed-capability inference to agentic interaction, where systems exercise closed-loop intelligence through multi-step interaction. We refer to this emerging class as agentic mobile AI systems. Its defining property is runtime variability: e.g., task demands, interaction states, environments, user intent, and device resources change jointly during execution. This shift exposes the limitation of post-hoc optimization, which cannot make step-wise decisions over representation, reasoning, action, memory, and execution under runtime constraints. We argue that the core systems problem therefore changes to runtime-elastic realization, i.e., online selection and coordination of sufficiently capable and cost-aware configurations for the current interaction. This survey develops a systematic view of this transition. Rather than surveying mobile foundation models, general AI agents, or efficient LLM inference in isolation, we organize existing work around five coupled runtime decisions: what to represent, how much reasoning and capability to activate, how to realize actions under feedback, what state to retain, and how to execute under mobile constraints. We further synthesize representative instantiations, evaluation metrics, runtime substrates, and future directions, aiming to clarify this emerging systems problem and stimulate research on runtime-elastic agentic mobile AI systems.

The cooperative tracking of Multi-Autonomous Underwater vehicles (AUV) has shown great potential in fields such as ocean environmental monitoring, marine resource exploration, and underwater security. However, limited underwater acoustic communication range, sparse deployment of sensor nodes, and environmental uncertainties often lead to incomplete target trajectory information and partially observable states, resulting in decreased tracking accuracy and unstable performance in complex scenarios. To solve the above problems, this paper proposes a Hierarchical Deep Reinforcement Learning framework, termed ASF-HDRL (Attention-guided high-level Switching with stage-conditioned Feature modulation), which decouples high- and low-level policies and incorporates a target estimation algorithm to enhance adaptability and robustness in complex tasks. Simulation results demonstrate that ASF-HDRL achieves superior cooperative tracking performance in challenging simulation environments, outperforming several mainstream baseline methods in terms of convergence speed and tracking accuracy.

Large-scale neural models face significant deployment challenges in resource-constrained AIoT environments that demand computational efficiency and collaborative adaptability. Modular Neural Networks (MNNs) offer promising solutions but struggle with automatic construction of functionally differentiated modules and lack efficient decentralized collaboration mechanisms for distributed AIoT deployments. To address these challenges, we propose BrickNN, a self-organizing modular neural network framework for distributed AIoT systems. BrickNN consists of three key components. First, a constraint-guided module construction mechanism enables functional differentiation through Specialization-Biased Module Design, Routing-Induced Specialization Learning, and Partitioned Static Signature Initialization. Second, Input-Driven Hierarchical Routing generates routing decisions from input features and decouples semantic module selection from execution constraints, enabling efficient and adaptive routing under heterogeneous deployment conditions. Third, Task-aware Decentralized Collaboration performs reliability-aware module collaboration and selective aggregation, enabling efficient knowledge sharing without central coordination. Extensive experiments on visual recognition and IoT sensing benchmarks demonstrate that BrickNN consistently improves system efficiency, achieving up to 10.38%–497.27% gains over baseline methods under varying communication conditions. The code is available at https://github.com/limmodular/BrickNN .

Federated Learning (FL) has gained prominence for its ability to enable collaborative model training while preserving data privacy in distributed environments. Despite its potential, two critical challenges persist. First, non-Independent and Identically Distributed (non-IID) data across clients induce severe parameter divergence, which conventional linear aggregation fails to reconcile, leading to suboptimal convergence. Second, FL often encounters a performance plateau where frequent client-server synchronization yields diminishing returns. While existing adaptive communication schemes attempt to mitigate this, they often struggle to balance the optimal model accuracy with the stringent resource constraints of edge devices. To address these challenges, we propose FedPDR, a two-stage server-side optimization framework designed to improve both efficiency and convergence in FL. An adaptive plateau detection mechanism is employed to determine the onset of the post-optimization stage. Once a performance plateau is identified, the refinement tasks are offloaded to the server, eliminating further client involvement and mitigating post-plateau inefficiencies. Central to this approach is the use of a latent diffusion model to capture the distribution of global model parameters and generate optimized updates, thereby alleviating suboptimal convergence caused by data heterogeneity. Extensive experiments across multiple benchmarks and real-world settings demonstrate that FedPDR effectively shifts computational burdens, reduces client-side overhead, and consistently improves global model performance while maintaining broad compatibility with existing FL frameworks.

Inter-agent communication serves as an effective mechanism for enhancing performance in collaborative multi-agent reinforcement learning (MARL) systems. However, the inherent communication latency in practical systems induces both action decision delays and outdated information sharing, impeding MARL performance gains, particularly in time-critical applications like autonomous driving. In this work, we propose a Value-of-Information aware Low-latency Communication (VIL2C) scheme that proactively adjusts the latency distribution to mitigate its effects in MARL systems. Specifically, we define a Value of Information (VoI) metric to quantify the importance of delayed messages on the recipient agent's decision. We then design a VoI aware resource allocation method that dynamically prioritizes message transmission based on each delayed message's importance. Moreover, we propose a progressive message reception mechanism to adaptively adjust the reception duration based on received messages. We derive the optimized VoI aware resource allocation and theoretically prove the performance advantage of the proposed VIL2C scheme. Extensive experiments demonstrate that VIL2C outperforms existing approaches under various communication conditions. These gains are attributed to the low-latency transmission of high-VoI messages via resource allocation and the elimination of unnecessary waiting periods via adaptive reception duration.

Over the past two decades, deep learning (DL) has achieved unprecedented breakthroughs across diverse application domains spanning computer vision (CV) to natural language processing (NLP). However, despite significant advances in computational resources and algorithmic frameworks, the training of deep neural networks continues to present formidable challenges due to persistent issues of training inefficiency and inherent data distribution biases. Recent years have witnessed the emergence of hard sample mining (HSM) as a promising paradigm to mitigate training inefficiencies and enhance model robustness through representative sample selection. Although HSM is reshaping contemporary AI research, its critical role in enabling efficient and robust model training has not yet been systematically explored. This article presents a comprehensive survey of HSM methodologies by: 1) establishing unified definitions of hard samples through rigorous sample complexity quantification criteria; 2) proposing a systematic taxonomy of HSM approaches with in-depth technical analysis; and 3) identifying pivotal research frontiers in this evolving field. This survey not only consolidates the foundations of HSM but also provides a roadmap for advancing efficient, robust, and generalizable deep learning models.

Recommendation systems alleviate the issue of information overload via modeling user preferences from interaction sequences. Although self-attention based sequential models effectively capture long-range dependencies, they are susceptible to noise amplification in sparse sequences and over-smoothing of item representations, which obscures true user intent and reduces sensitivity to fine-grained behavioral changes. To overcome these challenges, we propose SAFA, a sparse sequential recommendation framework comprising: (1) an adaptive sparse attention mechanism that suppresses noisy interactions while preserving embedding diversity; (2) a frequency-aware encoder that decomposes interaction sequences into low-frequency components for long-term preference modeling and high-frequency components for short-term intent dynamics; and (3) a simplified focal loss that removes the class-balancing term while preserving the focusing factor, emphasizing hard-to-predict samples rather than class priors. Experiments on seven benchmark datasets demonstrate that SAFA consistently achieve state-of-the-art performance with average improvements of up to 3.77%, 4.10% and 4.25% in terms of HR@5, HR@10 and HR@20, respectively, and 4.30%, 4.78% and 4.58% in terms of NDCG@5, NDCG@10 and NDCG@20, respectively, over the best competing model. Ablation studies verify the importance of each component, with notable performance degradation upon removing the sparse attention or frequency-aware encoder. Overall, SAFA enhances sequential recommendation by improving robustness and discriminative learning under noisy and sparse conditions.

Multidevice cooperative wireless sensing offers a promising solution for human motion recognition, owing to its superior privacy preservation and robustness. In the sensing process, devices continuously extract features from channel echoes and transmit them to a fusion center for motion recognition over successive time slots. The intertwined sub-processes of sensing and communication jointly determine recognition accuracy, yet simultaneously compete for limited radio resources. Moreover, the dynamic nature of practical environments further complicates this interplay due to the presence of moving interference sources and time-varying number of cellular users sharing the available bandwidth. Therefore, it is of paramount importance to jointly optimize sensing and communication resource allocation among devices and across time slots, while meticulously accounting for the impacts of dynamic environment to maximize recognition accuracy. In this paper, we propose a task-oriented integrated sensing and communication (ISAC) system for multidevice cooperative wireless motion recognition in dynamic environments. Specifically, we formulate a joint sensing and communication resource allocation problem to maximize recognition accuracy, represented by a discriminant gain metric that explicitly accounts for both sensing quality and communication constraints. Since this problem is a fractional program, we transform the original sum-of-ratios objective function into an equivalently subtractive form that facilities the development of a two-step iterative offline optimization (TSIO) algorithm to achieve the benchmark performance. Furthermore, to effectively cope with dynamic environmental influences, we further design a multi-agent reinforcement learning (MARL)-based online optimization (MRLO) scheme, which predicts environmental conditions at the subsequent time slot and adaptively optimizes resource allocation. Extensive numerical results illustrate that the proposed algorithm significantly enhances the recognition accuracy with dynamic environment influences, compared to existing benchmark algorithms. It is also observed from results that the sensing performance primarily drives recognition accuracy when energy is limited, whereas communication performance becomes the dominant factor under bandwidth constraints.

Semi-supervised multiview clustering has garnered considerable attention for its ability to integrate multiview data with limited labeled information. However, existing methods predominantly focus on labeled samples, neglecting abundant unlabeled data, which leads to suboptimal utilization of available prior knowledge. Moreover, existing pairwise constraint propagation-based methods typically follow a two-stage procedure, resulting in unstable clustering outcomes. To address these limitations, we introduce a unified framework that integrates multiview subspace clustering with pairwise constraint propagation, proposing a tensor-based semi-supervised multiview subspace clustering (TSMSC) method with pairwise constraint propagation. Specifically, each view’s subspace representation is decomposed into a consensus part and private parts, enabling the consensus representation to better approximate the low-rank structure. Then, a pairwise constraint propagation method is developed for multiview data, which propagates the initial pairwise constraint through a low-rank matrix completion approach. Finally, observing that the ideal multiview consensus representation and the propagated pairwise constraint matrix share the same low-rank structure, we naturally construct them into a third-order tensor to capture high-order correlations via tensor low-rank representation, allowing for joint optimization within a unified framework. Extensive experiments on eight real-world datasets demonstrate the superiority of our method over state-of-the-art approaches.

Active learning can effectively reduce the cost of labeling while enhancing model classification performance. However, prior studies have indicated that imbalanced class distributions adversely impact active learning, leading to diminished model effectiveness. Existing approaches to unbalanced active learning often neglect the multi-class imbalance problem and suffer from low performance and high time consumption. To address these issues, this paper introduces a hybrid active learning with online weighted broad learning system (HAL-OWBLS). Its main advantages include: (1) We optimize the initial labeled instance selection through an approximate query strategy to avoid the cold-start problem and introduce a sample selection strategy based on double uncertainty to enhance the rationality of active learning iterations. (2) A weighted broad learning system (WBLS) is chosen as the classifier, and an improved weighting strategy is adopted for multi-class imbalanced data. (3) We theoretically derive an efficient online updating model for WBLS, which reduces the time cost of active learning iterations by using only newly labeled samples for fast updating. The proposed HAL-OWBLS algorithm has better performance and robustness compared with existing related algorithms on various multi-class imbalanced data sets.

Semi-supervised learning (SSL) based on pseudo-label and consistency has achieved significant success. The core idea behind these methods is to assign sample weights based on pseudo-label probabilities, thereby guiding the model toward biased learning. However, existing research still faces two major challenges in guiding learning: (1) how to evaluate learning states across different classes in the absence of labels, and (2) how to construct an effective sample weight space that provides precise guidance throughout training. To address these challenges, we propose the Bi-Dimensional Sample Weight Guidance algorithm, BidMatch. BidMatch introduces Class Information Entropy (CIE), which captures the learning relationships between classes and reflects the model’s learning state for each class. Additionally, Pseudo-label Probability Redistribution (PPR) is proposed to maintain distribution invariance and sparsity during training, thereby emphasizing differences in sample importance. By leveraging CIE and PPR, BidMatch generates sample weights that account for both class and instance dimensions, effectively guiding the model toward balanced and efficient learning across classes. BidMatch has demonstrated state-of-the-art performance on various SSL datasets. Notably, it achieved a 6.45% error rate on CIFAR-10 with only one label per class, significantly outperforming baseline methods.

Due to its exceptional feature representation capabilities and high computational efficiency, the broad learning system (BLS) has been widely employed in various classification tasks. Nevertheless, BLS encounters considerable challenges in semi-supervised classification tasks involving complex heterogeneous data, given the data’s high-dimensional and noisy nature, coupled with a limited number of available labeled samples. To tackle these challenges, this article introduces a semi-supervised BLS based on distance constraint regularization (DRBLS) and a semi-supervised broad ensemble method (E-DRBLS) for high-dimensional data. Specifically, we present a distance constraint regularization (DR) that utilizes both labeled and unlabeled data to derive an optimal projection matrix, which maximizes the preservation of the original data’s intrinsic distribution structure. DR is designed to minimize intraclass distance, maximize interclass distance, and minimize the distance between neighboring samples. To boost the performance of BLS in semi-supervised classification, we integrate DR and BLS to construct the semi-supervised classifier DRBLS. Finally, we propose a mixed dimensionality reduction space generation (MDRSG) method that generates multiple high-quality and diverse mixed dimensionality reduction spaces (MDRSs). Based on MDRS, an ensemble framework, E-DRBLS, is developed for semi-supervised classification tasks targeting high-dimensional data. Comprehensive experiments confirm the superiority of the proposed methods.

Knowledge distillation (KD) is a widely utilized technique for transferring knowledge from a large model (the teacher) to a smaller model (the student). Owing to its flexibility and broad applicability, KD has been extensively applied in the compression of server-side models to meet the Quality of Service (QoS) requirements of client users. Despite significant advancements, the performance of distillation is substantially compromised when a large disparity exists between the capabilities of the server and the requirements of the client. To alleviate this problem, we propose a novel distillation approach, named Progressive^2, which operates through the combination of a progressively stronger teacher and a progressively smaller student. On the side of the teacher, rather than involving all layers simultaneously, we progressively select additional layers for distillation following a raw-to-rich semantic progression, establishing a systematic learning curriculum. Furthermore, we design a teacher-side multi-feature fusion adapter for the teacher to improve training stability, which is theoretically supported by the framework of Lipschitz continuity. On the side of the student, rather than directly training a tiny model, we gradually reduce the size of the network to facilitate an iterative co-evolution with the teacher. Progressive^2 serves as a flexible framework; the progressive strategy of the teacher can be deployed independently to achieve an optimal balance between accuracy and training efficiency, while the joint integration of the teacher and the student yields further improvements in overall performance.

Environment sensing and fusion via onboard sensors are envisioned to be widely applied in future autonomous driving networks. This paper considers a vehicular system with multiple self-driving vehicles that is assisted by multi-access edge computing (MEC), where image data collected by the sensors is offloaded from cellular vehicles to the MEC server using vehicle-to-infrastructure (V2I) links. Sensory data can also be shared among surrounding vehicles via vehicle-to-vehicle (V2V) communication links. To improve spectrum utilization, the V2V links may reuse the same frequency spectrum as the V2I links, which may cause severe interference. To tackle this issue, we leverage reconfigurable intelligent computational surfaces (RICSs) to jointly enable V2I reflective links and mitigate interference appearing at the V2V links. Considering the limitations of traditional algorithms in addressing this problem, such as the assumption of quasi-static channel state information, which restricts their ability to adapt to dynamic environmental changes and leads to poor performance under frequently varying channel conditions, in this paper, we formulate the problem at hand as a Markov game. Our novel formulation is applied to time-varying channels subject to multi-user interference and introduces a collaborative learning mechanism among users. The considered optimization problem is solved via a driving safety-enabled multi-agent deep reinforcement learning (DS-MADRL) approach that capitalizes on the RICS presence. Our extensive numerical investigations showcase that the proposed reinforcement learning approach achieves faster convergence and significant enhancements in both data rate and driving safety, as compared to various state-of-the-art benchmarks.

The emergence of the concept of heterogeneous crowd sensing and the development of multi-agent reinforcement learning technology have brought new ideas for achieving low-cost, high-coverage, and real-time target tracking in complex spatiotemporal environments. In order to deal with the challenges faced by collaborative scheduling of heterogeneous crowd agents in complex environments, such as the lack of correlation between the crowd agents and the environment, and the complexity of the crowd agent collaboration modes caused by heterogeneous functional attributes, we propose a collaborative scheduling framework for cooperative competitive games based on hybrid Markov games. Firstly, we construct a correlation modeling framework for complex spatiotemporal regional environments and heterogeneous crowd agents. Secondly, we design a multi-level Markov collaborative scheduling algorithm assisted by environmental information enhancement. Crowd sensing information enhances the information transparency in incomplete information games, and the alliance formation and dissolution mechanism helps to give full play to the functional characteristics of heterogeneous crowd agents. In addition, we design a reward function containing multi-dimensional reward information. After comprehensive evaluation, the proposed method achieves higher cumulative reward, faster loss convergence, and shorter tracking time compared with baseline methods. The correctness and effectiveness of the model are verified through heterogeneous proportion analysis of crowd agents and comparative analysis of core module ablation. Experimental results show that the model can reduce tracking time and improve resource utilization efficiency in complex environments.

The synthesis of cross-modal medical images plays a vital role in bridging diagnostic gaps between imaging modalities such as CT, MRI, and PET. This integration enables a more comprehensive evaluation of a patient's condition, improving diagnostic accuracy and aiding clinical decision-making. However, the performance of conditional denoising diffusion probabilistic models is often hindered by pronounced structural and intensity discrepancies between modalities, as well as the inherently slow nature of the diffusion process. To address these challenges, this paper proposes Wavelet-Based Diffusion in the Difference Domain for Cross-Modality Medical Image Generation (Med-D3CG), a novel framework that transforms the synthesis process by emphasizing the difference domain. Instead of directly generating target images like conventional methods, Med-D3CG models the residual information between conditioned and target images. This strategy allows the framework to accurately capture essential structural and intensity variations between modalities, leading to more precise and realistic image synthesis. Additionally, Med-D3CG integrates the Discrete Wavelet Transform (DWT) to improve efficiency, accelerating the diffusion process while maintaining high image fidelity. On SynthRAD2023 and HMIFD datasets, state-of-the-art performance is achieved on pelvis and HMIFD using Med-D3CG, with the best Learned Perceptual Image Patch Similarity (LPIPS) and competitive FID observed on brain. Code and pretrained models are provided at https://github.com/ZgzTTTer/Med-D3CG.

Semi-supervised learning can leverage both labeled and unlabeled samples simultaneously to improve performance. However, existing methods often present the following issues: (1) The emphasis of learning is put on either the similarity structures or the regression losses of data, neglecting the interaction between them. (2) The similarity structures among boundary samples might be unreliable, which misleads label propagation and impairs the performance of models on out-of-sample data. (3) They often involve the inverses of high-order matrices, making them inefficient in computation. To overcome these issues, we propose a scalable semi-supervised learning framework with Discriminative Label Propagation and Correction (DLPC), which collaboratively exploits the regression losses and similarity structures of data. Particularly, each sample is projected onto the independent class labels associated with nonnegative adjustment vectors rather than the propagated labels, such that the distances between samples from different classes are naturally enlarged, making regression losses more effective for boundary samples. Benefiting from this, the regression losses can guide the propagation of labels in boundary areas. Thus, the label information is first propagated through dynamically optimized graph structures and then corrected by the regression losses, effectively improving the quality of labels and facilitating feature projection learning. Furthermore, an accelerated solution has been developed to reduce the computational costs of DLPC on sample scales, thereby making it scalable to relatively large-scale problems. Moreover, the proposed DLPC can not only be applied to single-view scenarios but also extended to multi-view tasks. Additionally, an optimization strategy with fast convergence has been presented for DLPC, and extensive experiments demonstrate the effectiveness and superiority of DLPC over state-of-the-art competitors.

The Internet of Vehicles (IoV) is transforming the transportation ecosystem, enabling pervasive connectivity and data-driven approaches. Deep learning and generative artificial intelligence (GAI) have the potential to significantly enhance IoV application performance by enabling efficient decision- making and predictive capabilities, including intelligent navigation, vehicle safety monitoring, and intelligent traffic management. Nevertheless, efficiently transmitting and processing the massive volumes of data generated by the IoV in real time remains a significant challenge, particularly in dynamic, unpredictable wireless channel conditions. To address these challenges, this paper proposes a semantic communication (SC) framework based on a CNN-Transformer architecture that provides Channel estimation via Diffusion (CTCD) to improve the accuracy and efficiency of data transmission. The SC model extracts and compresses the information to be transmitted. The wireless channel is then estimated using a generative diffusion model to predict dynamic channel states, thereby improving the quality of IoV services. In dynamic scenarios, however, channel estimation performance may degrade when substantially new scenarios occur, adversely affecting the user experience. To mitigate this limitation, we fine-tune the channel generation model to enhance its adaptability across varying scenarios. The performance and reliability of the proposed framework are evaluated on two public image datasets, and the framework outperforms the comparative methods in image reconstruction metrics.

Human mobility predictability concerns the best prediction performance attainable from a given target and input information, but its ground truth is not directly observable on real mobility data. We present BER-PEF, a Bayes-error-rate-based framework that converts BER estimation into mobility predictability estimation and provides a unified protocol for comparing estimators without observable ground truth. The framework maps symbolic sequences, numeric trajectories, contextual features, and learned representations into a common feature–label space, then evaluates estimator outputs along controlled perturbation curves against a shared predictability reference interval by measuring deviations below the interval, above the interval, and across the full interval. Experiments on Foursquare NYC and TKY, GeoLife, and T-Drive show that several BER-based estimators achieve lower reference discrepancy than existing predictability methods on symbolic sequences and numeric trajectories, while their estimates track changes in empirical prediction performance under perturbation. Additional analyses show that contextual inputs and multiple structured representations can be evaluated under the same protocol, and that aggregating evidence across multiple perturbation levels provides a more reliable basis for estimator selection than relying on a single unperturbed observation. BER-PEF therefore offers a unified and verifiable path for evaluating predictability estimators on heterogeneous mobility data when ground-truth predictability is unavailable.