Neural combinatorial optimization methods based on deep reinforcement learning (DRL) have shown promising results recently in solving complex combinatorial optimization problems (COPs). However, existing approaches usually require training separate models for each COP. When facing multiobjective COPs (MOCOPs), a large number of models still need to be trained to solve the single-objective subproblems generated by the weight scalarization strategy. To address the two key challenges of cross-task learning and multiobjective optimization, we propose an effective cross-task unified multiobjective neural metaframework (CUMNM), which uses the novel attribute-sharing attention model (AM) as the base model. It is the first general end-to-end framework for solving various MOCOPs with an arbitrary number of objectives using a single model. Initially, a metamodel is trained on multiple MOCOPs through metalearning. For a specific MOCOP, only a few fine-tuning steps for the metamodel are required to derive a new task-adapted model. Subsequently, the scalarized subproblems of the task can be solved directly by updating only a small subset of parameters in the new model. This enables the construction of an approximate Pareto front (PF) without retraining from scratch for each subproblem. During training, we exploit the symmetry of MOCOP solutions and introduce a low-variance baseline method to stabilize and accelerate the training process. For inference, performance is enhanced via instance augmentation and a novel regret-based multistart search strategy. Comprehensive experiments on four typical MOCOPs demonstrate that our method significantly outperforms other state-of-the-art algorithms in solution quality and model efficiency. It also exhibits strong performance on unseen but structurally similar tasks, demonstrating zero-shot generalization.
Recent advances in contrastive self-supervised learning (SSL) have pursued increasingly complex architectures, such as momentum encoders and deep projection heads, adding significant computational overhead. This paradigm of growing complexity motivates a systematic deconstruction of modern frameworks to identify their truly essential components. Our analysis reveals that a precisely defined “Cross-View Symmetric InfoNCE” loss, which purifies the contrastive objective by exclusively using interview negative samples, is a primary driver of both superior performance and accelerated convergence. We encapsulate this principle within contrastive representation with essential components (CORE), a minimalist framework built upon a shared encoder and an optimized three-layer projection head. On ImageNet, CORE achieves a competitive 73.7% top-1 accuracy with just 100 training epochs and demonstrates remarkable efficiency, outperforming key baselines by up to 1.5% on ImageNet-100. Furthermore, this efficient framework reveals a counter-intuitive finding: the simpler shared encoder consistently outperforms the more complex momentum encoder under large-batch settings (e.g., 1024), challenging its conventional necessity for stable training. Our results, supported by extensive ablations and theoretical insights from an information bottleneck perspective, provide a clear, validated design paradigm for building highly efficient and powerful SSL systems. Code and training logs are available at https://github.com/Jaywalk18/CORE.
Existing medical image segmentation (MedISeg) models predominantly rely on convolutional neural networks (CNNs) and Transformer architectures. However, the limited receptive fields of CNNs and the quadratic computational cost of Transformers hinder their scalability and efficiency. Recently, receptance-weighted key-value (RWKV) has emerged as a promising linear-complexity alternative for global context modeling. In this article, we propose SRWKV, a shape-guided RWKV (SGR) model for parameter-efficient MedISeg. SRWKV introduces an SGR block that uses a shape prior predicted from the deepest encoder feature to guide token traversal during decoding, reducing foreground-background interleaving and improving structural coherence during sequence formation. In addition, we develop a deformable adaptive shift (DA-Shift) module that dynamically adjusts token interactions according to local context, enabling flexible receptive field adaptation for diverse anatomical structures. Extensive experiments across six MedISeg tasks on 11 datasets demonstrate that SRWKV achieves strong segmentation performance with a compact parameter footprint. Our code is available at https://github.com/ukeLin/SRWKV.
Accurate prediction of protein-protein interaction sites (PPISs) plays a crucial role in understanding protein function, elucidating disease mechanisms, and facilitating drug target discovery. Although conventional approaches based on sequence or structural features have shown promising results, they still face several challenges. These challenges include oversmoothing in deep graph neural networks (GNNs) and poor generalization to domain-specific data. To address these issues, we propose RGLLA-PPIS, a novel multimodal prediction model that integrates retrieval-augmented learning and residual GNNs for PPIS identification. In RGLLA-PPIS, protein graphs are constructed by combining AlphaFold3 (AF3)-predicted protein structures with multiple sequence-derived features. To effectively capture both local and global spatial dependencies, the model employs equivariant GNN (EGNN) and GCN modules with residual connections, which help alleviate the oversmoothing problem and preserve node-level variability. Moreover, during prediction, we used the retrieval-augmented knowledge provided by the pretrained protein language model (PLM) Evolla and ChatGPT-4o to construct semantic priors to supplement potential functional site information and enhance the generalization capacity of the prediction model. Extensive experiments on benchmark datasets show that RGLLA-PPIS outperforms several state-of-the-art baselines in both accuracy and robustness. Furthermore, comparison with wet-lab results on a domain-specific protein system reveals a strong correspondence between experimental functional sites and the high-probability regions predicted by RGLLA-PPIS. This demonstrates the model's potential to guide real-world protein engineering tasks. The source code can be found at: https://github.com/MiJia-ID/RGLLA-PPIS.
Adversarial attacks pose significant challenges to the security and robustness of deep-learning models. Stochastic neural networks (SNNs) have shown promising effectiveness in improving robustness by injecting stochastic noise into model activations, features, or weights. However, most existing SNN-based defenses rely on predefined distributional forms, such as Gaussian or Uniform. In real-world scenarios, data distributions are often non-Gaussian, skewed, or multimodal, which cannot be adequately captured by such fixed assumptions, thereby limiting the robustness of existing methods. To address this limitation, we propose a novel SNN named arbitrary distribution injection (ADI), which enables distribution modeling from nonpredefined, data-dependent distributions. In particular, we introduce a conditional stochastic feature mapping mechanism to model feature distributions, together with a theoretically grounded variance-regularization loss. Extensive experiments across diverse attack methods, datasets, modalities, and network architectures show that ADI achieves robustness improvements and promising generalization across the evaluated settings. Furthermore, detailed parameter analyses and feature-distribution visualizations provide deeper insights into the underlying mechanisms of ADI.
This article proposes a novel learning-based Shrink-Horizon Economic Model Predictive Control (SHEMPC) algorithm for constrained nonlinear systems with unknown dynamics and additive disturbances. By combining data clustering and kernel techniques, a componentwise Hölder regression algorithm is customized to construct an accurate and sparse dynamics predictor. The kernel weights and Hölder matrices within each cluster are determined using a tailored Block Coordinate Descent (BCD) optimization method. Then, two error bounds of the regression-based predictor are estimated a priori for the design of SHEMPC. Next, by integrating the customized constraint-tightening method of the predictor, the SHEMPC algorithm is formulated as two finite-horizon receding optimization problems solved iteratively. The algorithm enables a flexible trade-off among economic performance, stability, and online computational burden via simple parameter adjustments. Moreover, some mild assumptions are derived to ensure the feasibility and Input-to-State Stability (ISS) of the closed-loop system as well as smoothness and prediction complexity of the predictor. The performance and merits of the proposed algorithm are illustrated through a numerical example and a Continuously Stirred Tank Reactor (CSTR) simulation.
Motor imagery (MI) is a core paradigm in Brain-computer interface (BCI) research, but its practical application remains limited by intersubject variability and the scarcity of labeled target-domain data. Existing methods usually focus on a single adaptation level, such as domain alignment, feature interaction, or model fine-tuning, which limits comprehensive cross-domain adaptation (DA). To address this issue, this study proposes U2Multi-UDA, a unified multilevel multisource unsupervised DA framework for MI decoding. U2Multi-UDA integrates these three adaptation levels into a single pipeline. First, optimal transport (OT) aligns source and target distributions, while mutual information estimates source-domain relevance weights to characterize the contribution of each source domain to the target domain. Second, spatio-temporal electroencephalography (EEG) features are extracted and fused through multisource cross-attention, where the source-domain relevance weights guide cross-domain feature fusion, and pseudolabels enhance target-domain feature learning. Finally, segmented weight-decomposed low-rank adaptation (DoRA) enables parameter-efficient target-domain fine-tuning while reducing overfitting. Experiments on BCI Competition IV 2a, BCI Competition IV 2b, and the self-constructed MI-GS dataset show that U2Multi-UDA improves mean accuracy by 2.69, 1.89, and 3.83 percentage points, respectively, over the best-performing baselines, with consistent gains in Kappa values. Ablation and sensitivity analyses further confirm the effectiveness, robustness, and physiological plausibility of the proposed framework.
In this article, we propose trajectory-constrained human-guided reinforcement learning (TCHug-RL), a new framework that 1) produces guidance at the full-trajectory level to suit the planning horizon of autonomous driving and 2) provides a formal criterion for deciding when and how to inject human input into the policy-update loop. First, we propose a trajectory-level similarity of human guidance, which is defined such that the satisfaction of trajectory-level similarity implies the satisfaction of single-step similarity. Consequently, trajectory-level similarity is strictly stronger than the single-step one. Then, the trajectory-level similarity is modeled as a constraint to guide the subsequent policy optimization. In this way, not only is a clear criterion for human guidance provided, but the influence of such guidance on the policy is also transparent and interpretable. Finally, we leverage the Lagrangian method to solve the constrained optimization problem and provide a strong duality analysis for it in the unparameterized policy space. Furthermore, a practical algorithm implementation of TCHug-RL and experiments are provided, demonstrating that TCHug-RL effectively leverages human guidance and achieves improvements of 22.1% in learning efficiency and 18.7% in overall performance in autonomous driving tasks, compared to the state-of-the-art methods.
Multistage cable routing requires a robot to successfully navigate a cable through a series of clips, which is a challenging task. Due to the unpredictability of cable deformation and the complexity of aligning cables with clip openings, traditional model-based or imitation learning methods are difficult to achieve satisfactory results. To address the difficulty, we propose a novel reinforcement learning-based routing framework guided by a vision-language model (VLM). The core of the proposed framework is to establish the description of the spatial position constraints between cables and clips in VLMs to produce a high-level plan that incorporates grasping and routing points, which then guides the local routing strategy. The local routing strategy is trained by human-in-the-loop reinforcement learning, enabling the model to learn a strategy that reduces deformation when interacting with clips. Our method is validated on a Franka robot across multistage cable routing tasks, and it outperforms some baselines in terms of reliability and generalization.
Conditional latent diffusion has become a promising paradigm for synthesizing paired image-mask data for segmentation, yet preserving structural topology remains challenging for curvilinear objects such as vessels and cracks. Existing methods mainly focus on how to inject spatial conditions into generative backbones, while paying limited attention to whether the low-resolution condition representations themselves remain topologically reliable. As a result, mask-conditioned diffusion often suffers from two coupled failure modes: topology-breaking low-resolution projection of thin structural masks and progressive attenuation of structural cues during denoising. To address these issues, we propose TopoSegDiff, a topology-preserving conditioning framework for latent diffusion built on an adaptive teacher-student-adapter pipeline. Specifically, an adaptive topology constructor (ATC) first generates multiscale topology-preserving teacher conditions from the input mask. A lightweight topology distillation encoder (TDE) then amortizes these teacher conditions into feed-forward soft topology cues together with scale-wise reliability estimates. Finally, a reliability-aware topology adapter (RTA) injects the predicted conditions into the diffusion decoder, while a reliability-aware topology alignment objective regularizes intermediate features against topology-preserving targets. In this way, topology-preserving supervision is transformed into scalable and differentiable guidance for structure-aware generation. Experiments on six public curvilinear-segmentation benchmarks across three application domains show that TopoSegDiff produces more realistic and structurally faithful image-mask pairs, reduces topology inconsistency, and consistently improves downstream segmentation performance when used for synthetic augmentation. These results support topology-preserving conditioning as an effective route toward more reliable structure-aware generation and learning for curvilinear objects.
Multivariate time series classification (MTSC) poses significant challenges in various domains. Existing traditional machine learning (TML) approaches for MTSC offer interpretability, but may include nondiscriminative features, leading to reduced accuracy. Conversely, deep learning (DL) methods achieve high performance but often sacrifice interpretability. To address these limitations, we propose RLDFS, a deep reinforcement learning-based discriminative feature selection framework, designed to identify distinctive features and enhance accuracy. To ensure the reliability of candidate feature sets and mitigate model variance, we leverage an interval-based ensemble method to capture features effectively. Furthermore, we illustrate how features selected through reinforcement learning (RL) can be intuitively interpreted, demonstrating their importance in classification tasks. Extensive experimental evaluations showcase that our framework outperforms state-of-the-art MTSC algorithms.
To address the modeling challenges posed by multiscale temporal dependencies and sensor spatial correlations in multisensor systems, this article proposes a reinforcement learning-enhanced multiscale spatiotemporal deep network (MSTSDN-RL). Multiscale temporal convolutional networks with causal dilated convolutions are used to capture hierarchical temporal features with enlarged receptive fields, while adaptive graph structure learning and a multibranch graph convolution model capture dynamic sensor correlations. In addition, prediction correction is formulated as a Markov decision process, where reinforcement learning is introduced to sequentially optimize prediction results. Experiments on four C-MAPSS subsets achieve root mean square error (RMSE) values of 12.49, 13.11, 13.58, and 16.37, and Score values of 203.03, 601.25, 205.57, and 1156.88, respectively. On the XJTU-SY dataset, MSTSDN-RL obtains an average RMSE of 29.49 and mean absolute error (MAE) of 25.87. Comparisons with recent state-of-the-art methods, ablation studies, and robustness tests under noise and missing sensor data demonstrate the competitive accuracy, robustness, and generalization capability of the proposed method for aero-engine and rolling bearing prognostics.
This article addresses the optimal control problem of nonlinear multiagent pursuit-evasion (PE) games and proposes a novel adaptive dynamic programming (ADP) framework that simultaneously accounts for intrateam cooperation and interteam antagonism. Subsequently, for the continuous-time, nonaffine, and nonlinear structure induced by the PE error dynamics, a $Q$ -function is defined, and a continuous-time $Q$ -learning recursion is derived via integral reinforcement learning. An actor-critic neural network architecture is then employed to approximate the $Q$ -function and the optimal policy, respectively. To address a key difficulty in Lyapunov stability analysis for the nonaffine case, where the actor-network weight-update rate is hard to construct as a negative-definite quadratic form in the weight error and thus does not readily yield uniformly ultimately bounded (UUB), a structured weight-update law is proposed along with a closed-loop stability proof. Under the stated compact-domain, persistent-excitation, and small-gain conditions, the closed-loop analysis establishes uniform ultimate boundedness of the system errors and the weight-estimation errors. Simulation results for a scenario with two pursuers against three evaders demonstrate that the pursuers can achieve capture.
In recent years, deep learning (DL) has achieved excellent performance on the joint classification of hyperspectral images (HSIs) and light detection and ranging (LiDAR) data. However, the existing DL methods often struggle to capture both global and local spatial features simultaneously, resulting in inadequate modeling of long- and short-range spatial dependencies. Moreover, the inherent heterogeneity of multisource data complicates effective fusion. To address these issues, we propose a multifrequency Mamba (MF-Mamba) network for the HSI and LiDAR joint classification. First, a Toeplitz fusion module (TFM) is designed to enable multilevel feature interaction between modalities, thereby reducing discrepancies. Second, a global-local Mamba (GL-Mamba) module is presented, which employs a frequency-space processing strategy. Specifically, the discrete wavelet transform (DWT) is first applied to decompose features into low- and high-frequency components. These components are then efficiently enhanced by Mamba before being reconstructed via inverse DWT (IDWT). Finally, to overcome the limitations of conventional fusion methods in capturing complex feature interdependencies, a multisource cross fusion (MSCF) module is proposed to integrate high-level semantic features from heterogeneous data. Extensive experiments demonstrate that MF-Mamba outperforms the state-of-the-art joint HSI-LiDAR classification networks, highlighting its superiority and potential. The source code will be made available at https://github.com/Dilingliao/MF-Mamba.
Deep learning has been used for object detection in remote sensing images successfully. However, how to detect objects precisely and efficiently with proper architecture complexity is an important issue when there are limited computing resources. Furthermore, current object detection neural networks are usually static and cannot perform differentiated training on samples of varying difficulty. In this article, we propose a method combining hierarchical category decoupling (HCD) and structural continual learning (SCL), named HCD and SCL network (HCD-SCL Net), for object detection in remote sensing images, which can generate the optimal architecture adaptively according to the difficulty level of different categories, overcome overfitting and catastrophic forgetting, and thus the detection performance is improved. First, the hierarchical category decoupling method is proposed to select categories of samples for object detection according to the difficulties of samples and class probability. Then, the SCL method is proposed to generate a proper architecture with a different number of feature extraction layers for different object classes. Thereby, the current stage of the detection model can learn the difficult categories of objects without forgetting the learned detection capabilities for easy categories in previous stages. By doing so, the dynamic tradeoff between the complexity of different classes of object detection and the computational cost is reduced. The experimental results on the remote sensing object detection datasets RSOD, DIOR, and NWPU VHR-10 show that the proposed method had excellent performance compared with state-of-the-art deep learning object detection methods.
The interpretable geometric constructive neural network (IGCNet) has demonstrated significant potential for developing rapid learning models, making it particularly well-suited for data analysis tasks. However, its geometric control strategy is difficult to adapt to the dynamic transformation of network residuals, resulting in limited generalization ability. To address this issue, this article proposes a greedy variant of IGCNet, termed IGCNet-G, which enhances generalization. First, an adaptive relaxation parameter is constructed by analyzing the convergence behavior of network residuals within a function class. This approach eliminates the excessive computational cost associated with passive hyperparameter tuning. Second, the relaxed parameters and greedy theory are incorporated into the geometric control strategy of IGCNet, leading to the development of a relaxed greedy control strategy. This strategy enables the adaptive relaxation of constraints on candidate node parameters according to the convergence state of residuals, thereby accelerating residual convergence. Finally, the universal approximation property of IGCNet-G is substantiated through theoretical analysis. Experimental evaluations conducted on a nonlinear function example and a real coal flotation dataset confirm that IGCNet-G achieves a high level of predictive performance.
Visual degradation caused by adverse meteorological conditions, such as low light and rain, significantly hinders the deployment of salient object detection (SOD) on edge devices. Existing methods often rely on computationally expensive restoration preprocessing or complex feature stacking, making real-time inference difficult. To address these challenges, this article proposes a novel lightweight model, termed the robust elastic adaptive difference network (READNet). The core innovation lies in a novel elastic pixel differential convolution operator, which flexibly captures microscopic gradient cues to effectively coordinate high-level semantics with low-level details. This operator is further embedded into an inverted residual block through a multibranch structural reparameterization, enabling structure-aware feature enhancement during inference with zero additional computational overhead. Furthermore, to mitigate nonuniformly distributed environmental noise, a condition-adaptive dual gate is introduced. The module innovatively integrates second-order variance statistics and contextual difference mechanisms to adaptively recalibrate features across both channel and spatial dimensions. Experimental results demonstrate that READNet achieves state-of-the-art performance on challenging benchmarks, validating its superior parameter efficiency and suitability for real-time applications. The source code is publicly available at https://github.com/TurnHug/READNet.git
Linear models are extensively employed in tomographic imaging to model the imaging system, with tomographic reconstruction from measurement data typically formulated as a linear inverse problem. However, linear models fail to accurately characterize the imaging system under nonideal imaging conditions. Nonlinear models provide a superior characterization of tomographic imaging systems in such conditions, albeit posing substantial challenges in directly solving the associated nonlinear inverse problem. This article introduces a physics-inspired dual regression network (Tomo-Net), which integrates dual learning with the physical mechanisms of tomographic imaging to solve the nonlinear inverse problem. Tomo-Net constructs a dual regression network by incorporating both primal and dual mappings. Primal mapping facilitates the transformation of degraded measurement data into high-quality reconstructed images, while dual mapping enables the transformation of reconstructed images back to the original measurement data. Tomo-Net incorporates the physical mechanisms of tomographic imaging in both mappings to constrain the solution of the nonlinear inverse problem, and solves the issue of nonuniqueness in the solution space by enforcing data fidelity constraints, dual cycle consistency constraints, and physical mechanism consistency constraints. In addition, Tomo-Net applies consistency constraints during the training phase rather than inference, thereby avoiding additional inference costs. While adaptable to various imaging modalities with suitable modifications, this article focuses on X-ray computed tomography (CT) imaging as a specific example and evaluates its effectiveness across various reconstruction tasks. The extensive experimental results demonstrate that Tomo-Net outperforms traditional machine learning models in solving nonlinear inverse problems in CT. The Tomo-Net code is publicly available at https://github.com/guoyii/Tomo-Net
Timely and accurate detection of sensor drift faults is critical for ensuring operational safety and compliance with effluent quality standards in the wastewater treatment process (WWTP). However, the slight and elusive nature of drift faults, coupled with the dynamic, nonlinear, and noisy characteristics of WWTP, poses a significant challenge to the detection of incipient drift faults. To address this problem, a novel feature-augmented broad neural network (FABNN) is proposed for sensor incipient drift fault detection. FABNN is constructed as a unified broad learning architecture, in which the feature layer, enhancement layer, and output layer are jointly redesigned to improve fault detection and faulty sensor localization accuracy. Specifically, a weighted feature filter (WFF) is devised for the feature layer to highlight low-frequency faults and suppress high-frequency noise. In the enhancement layer, a time-gated attention (TGA) mechanism is devised to capture the temporal characteristics of incipient drift faults, thereby enhancing fault feature extraction capability. In addition, an incremental learning version of FABNN is developed to dynamically extend the model structure without time-consuming retraining. To assess the performance of FABNN, it is evaluated on both benchmark and actual datasets from WWTP. The experimental results demonstrate that FABNN outperforms representative comparative methods in detecting incipient drift faults and locating faulty sensors at minimal sampling intervals after fault occurrence while achieving the highest accuracy and the lowest false and missed alarm rates (MARs). Moreover, FABNN shows strong robustness to various noise types and maintains stable detection performance for incipient drift faults.
Poisson noise and Gaussian noise are widely present in high-dimensional data, which poses a challenging problem. Moreover, Poisson noise is signal-dependent and influenced by sensor parameters and the imaging environment; therefore, methods based on the additive Gaussian model are limited in handling such real-world scenarios. Meanwhile, self-supervised learning-based network frameworks can construct labels from noisy images for network training, addressing the issue of matching between noisy and clean images. However, while existing frameworks are very effective in handling Gaussian noise, the presence of a mixture of Poisson noise and Gaussian noise suppresses their performance. To address this issue, we propose a Poisson-Gaussian mixed noise removal framework, SPGM-Net, which is the first method for high-dimensional data Poisson-Gaussian mixture noise removal based on a self-supervised learning framework. First, we propose a self-supervised noise estimation network based on the generalized Anscombe transformation (GAT) noise transformation formula, which transforms noise into approximately Gaussian noise. Second, based on subspace representation theory, we build a series of training samples for self-supervised network training using a neighbor pixel sampler on the transformed noisy eigenimages. No matter in the noise estimation stage or the self-supervised denoising stage, only noisy images are used, which is suitable for real-world scenes. To verify the effectiveness of the proposed method, we test it on various data, including indoor hyperspectral images (HSIs) and real satellite images. The results of both the simulated data and the real data show that the proposed method is superior to state-of-the-art methods.
