Multimodal sentiment analysis (MSA) aims to accurately predict the user's emotional tendency by integrating multimodal data such as text, audio and visual information posted by users. However, in most prior studies, the information differences between modalities are not fully considered as these modalities are treated equally. In fact, text modality usually contains more richer semantics to exhibit emotion, while audio and visual ones cover more redundant and even noisy information. Though some works have highlighted the key role of text modality, they still fail to fully utilize the non-text modalities, leading to sub-optimal performance for MSA. To this end, we propose a novel Text-centric Bidirectional Modality Enhancement Network (TB-MEN) in this paper, which recognizes the core role of text and fully exploits the audio and visual modalities at the same time. First, in light of the text features extracted by BERT and non-text features extracted by LSTM respectively, we develop a trimodal multi-scale bottleneck fusion (TMBF) module to capture semantically enhanced text features. Specifically, this is achieved by bottleneck fusion mechanism and transferring the audio and visual modalities to text one. Furthermore, for non-text modalities enhancement, a sparse subspace alignment strategy is employed in text-dominant subspace alignment (TDSA) module, which sparsely maps non-text features into the text subspace to aggregate semantic information. Experimental results show that on several public datasets, including CMU-MOSI, CMU-MOSEI and CH-SIMS, our model has achieved significant gains over the state-of-the-arts. For example, on the CMU-MOSI dataset, TB-MEN improves Acc-2, Acc-5, Acc-7, and F1 by 0.45%, 2.04%, 2.62%, and 0.4%, respectively, compared to the second-best one.

Audio-visual scene classification (AVSC) aims at classifying a video recording into one of the predefined scene categories, using both audio and visual modalities, which is a fundamental yet challenging problem in scene understanding. Although some impressive results for AVSC have been achieved recently, the modeling of object semantics and long-term dependencies from various modalities remains fully unexplored in the literature. To this end, in this paper, a novel Multimodal Multi-scale Temporal Enhancement Network, namely M2TE-Net, is proposed for audio-visual scene classification. In particular, to handle the temporal information of audio-visual modalities, the Multi-Scale Temporal Enhancement Modules(MSTEM) are devised from global and local perspectives. Then, a U-shape Multimodal Interactive Module (UMIM) is designed to mine the interaction between two modalities, where the shared semantic information from different modalities is fused by exploiting the consistency and complementary ones. Experimental results on two benchmark datasets demonstrate that the proposed M2TE-Net outperforms state-of-the-art models, with competitive performance.

Existing Retinex-based training-free low-light image enhancement (LLIE) methods often rely on complex architectures or lack support for text-controlled personalization. In this paper, we propose RetinexGDP, a training-free and text-controllable LLIE framework that uniquely integrates Retinex-based image modeling with generative diffusion priors. First, we introduce a simplified Retinex decomposition by embedding weighted total variation optimization into a single Gaussian convolutional layer, enabling robust illumination estimation without the need for training. Next, we guide the diffusion denoising process using the estimated reflectance map, employing patch-wise inversion and reflectance-conditioned sampling to effectively suppress noise while preserving structural details. Finally, unlike previous diffusion-based LLIE methods that perform only monotonous global brightness enhancement, we incorporate text guidance into the sampling process, enabling controllable enhancement that aligns with user-specific stylistic preferences. RetinexGDP thus provides a modular, interpretable, and text-controllable solution for low-light image enhancement. Experimental results show that RetinexGDP achieves state-of-the-art performance in terms of NIQMC and CPCQI metrics across seven real-world datasets. Code will be available at: https://github.com/zhaozunjin/PLIE

Physics-Informed Neural Networks (PINNs) combine deep learning with physical constraints for solving partial differential equations (PDEs), and are widely applied in fluid mechanics, heat transfer, and solid mechanics. However, PINN training still suffers from high-dimensional non-convex loss landscapes, imbalanced multiobjective constraints, and ineffective information propagation. Existing curriculum learning and causality-guided strategies improve training stability, but mainly focus on temporal or parametric progression, lacking explicit treatment of spatial information propagation and inter-region consistency. Moreover, they are not directly applicable to boundary value problems (BVPs) with strong spatial coupling. To address this issue, we propose a spatially correlated curriculum learning framework for PINNs. To the best of our knowledge, this is the first work to address PINN training difficulties from the perspective of spatial coupling among subregions. First, spatial causal weights guide information from near-boundary regions inward, reducing optimization failures and spurious convergence. Second, a low-frequency information bridge enforces pseudo-label-based consistency across spatially separated regions, suppressing global low-frequency drift. Third, a region-adaptive reweighting strategy adjusts subregion losses to reduce local residuals and recover high-frequency details. Experiments on PDE benchmarks show that, under comparable computational cost, the proposed method alleviates training failures and improves solution accuracy. The code is available at https://github.com/pigofmomo/CurriculumLearningPINN.

Mixed integer linear programs (MILPs) are widely used to model large-scale or time-sensitive real-world optimization problems, where high-performance computing (HPC) techniques are often required. Modern solvers rely on the branch-and-bound (B B) search algorithm, where branching decisions critically determine search efficiency. Traditional branching heuristics rely on mathematical computations at each search step, which can be very expensive when high-quality decisions are required. This has motivated learning-to-branch approaches to shift computational burden from online search to offline training. Many existing methods adopt imitation learning (IL) to mimic traditional heuristics, but incur high offline costs for expert data generation and remain fundamentally limited by expert quality. Reinforcement learning (RL) offers a promising alternative, yet learning effective branching policies from scratch is difficult and still results in prohibitively expensive offline cost. Therefore, this work proposes PEBSI, an efficient RL-based branching policy that realizes this computational burden shift while addressing the cost of offline training. PEBSI learns without expert demonstrations or restrictive solver settings. It is guided by a decision-aware reward that provides informative learning signals and a quality-aware exploration strategy that improves training sample quality. Additionally, it employs a fully parallelizable training scheme that leverages HPC resources for scalable data generation and efficient policy learning. Extensive experiments on diverse MILP benchmarks show that PEBSI consistently outperforms RL baselines trained without expert guidance and, in several settings, matches or surpasses the IL-based counterpart.

Recent real-time semantic segmentation models have achieved remarkable performance and speed, yet certain challenges persist. Existing methods like multi-branch architectures can capture spatial details to enhance performance, but they often compromise inference speed; conversely, single-branch models such as SCTNet, though efficient, tend to overlook the learning of spatial details and still rely on third-party teacher models for performance enhancement. This paper proposes S2Net, a single-branch network with self-coaching that enhances both semantic information in low-resolution features and spatial details in high-resolution features without relying on third-party high-performance models, through the Self-Coaching Mechanism (SCM) and the Feature Alignment Module (FAM). Experimental results demonstrate that S2Net achieves an optimal balance between performance and speed on datasets like Cityscapes, delivering an impressive 77.3% mIoU and 211.2 FPS without the use of ImageNet pre-training. The code will be uploaded to https://github.com/gyyang23/ S2Net.

Physics-informed neural networks (PINNs) are a class of machine learning methods for solving partial differential equations (PDEs). Integrating domain decomposition with PINNs can alleviate issues such as the spectral bias of neural networks and gradient flow pathology during optimization, thereby improving convergence speed and accuracy. However, existing domain decomposition PINNs depend on manually decomposing the domain and are hindered by complex interface condition configurations and insufficient dynamic adaptability. This research introduces Adaptive Domain Decomposition based Physics-Informed Neural Networks (ADD-PINN) via Spatial Clustering. Firstly, spatial clustering is employed to automatically generate polygonal blocks, enabling shape manipulation of the computational domain. Secondly, the interface conditions and loss function are improved, and an algorithm to control information exchange based on gradient analysis is introduced. Thirdly, the tasks of subdomain initialization and dynamic adjustment are formulated as a graph optimization problem, which can be addressed using heuristic algorithms. Experimental results on PDE cases demonstrate that the proposed method attains superior accuracy and convergence efficiency compared with the baseline method XPINN. Through an analysis of the algorithmic complexity and runtime records, we have also identified the time bottlenecks associated with domain-decomposed PINNs and observed notable improvements in time efficiency achieved by our proposed methods

Facial expression recognition (FER) remains challenging when faces are partially occluded, as occlusions obscure expression-related regions and introduce irrelevant visual patterns, leading to substantial performance degradation in real-world scenarios. In this work, we propose an occlusion-aware low-rank learning framework for robust FER under occlusion. Unlike previous occlusion-robust FER methods that often require manually annotated occluded facial images, our method employs synthetically occluded images for training, thereby reducing the reliance on costly real-world occlusion annotations. The proposed framework integrates occlusion detection, masking, and feature-level low-rank recovery into a unified pipeline. Specifically, the model first detects occluded regions and replaces them with mask values to suppress the influence of diverse occlusion appearances, producing a cleaner but incomplete facial representation. Since this masked representation may still lose semantic information from occluded facial components, we introduce a low-rank learning module to recover missing expression-related features in the nonlinear feature space rather than reconstructing occluded pixels in the image space. Furthermore, because the corresponding non-occluded images in the synthetic training process provide complete facial semantics, we design a dual-branch guidance structure in which the non-occluded branch assists the occluded branch in feature recovery. By jointly optimizing these components, the proposed method achieves competitive performance on widely used in-the-wild FER benchmarks, including RAF-DB and AffectNet, and obtains state-of-the-art or highly competitive results on occluded FER datasets, including Occlusion-RAF-DB, Syn-RAF-DB, Syn-AffectNet, and FED-RO. These results demonstrate the robustness and generalization ability of the proposed framework under occluded conditions.

Anomalous Sound Detection (ASD) plays a critical role in practical scenarios ranging from audio surveillance to industrial machine condition monitoring. A big challenge of current methods is their poor generalization ability in detecting unknown anomalous sounds. Existing dominant solutions primarily adopt Two-Dimensional Convolutional Neural Networks (2D-CNNs) to extract features from the time-frequency spectrograms of audio signals. Nevertheless, the inherent limited receptive field of 2D-CNNs fails to fully capture and utilize comprehensive time-frequency information from spectrograms, leading to insufficient feature representation and degraded detection performance. To this end, in this paper, a novel Dual-path Channel-attention WaveNet (DC-WaveNet) is proposed to learn discriminative, robust audio feature representations for ASD tasks. The core novelty of the proposed DC-WaveNet lies in its dual-path feature learning paradigm equipped with channel-attention WaveNet encoders. Specifically, the network first excavates fine-grained latent feature information from the frequency dimension of audio signals, and further aggregates and models temporal dimension output features to capture global overall audio characteristics, ultimately generating highly discriminative feature representations suitable for anomalous sound identification.

The emergence of multi-agent reinforcement learning (MARL) is significantly transforming various fields like autonomous vehicle networks. However, real-world multi-agent systems typically contain multiple roles, and the scale of these systems dynamically fluctuates. Consequently, in order to achieve zero-shot scalable collaboration, it is essential that strategies for different roles can be updated flexibly according to the scales, which is still a challenge for current MARL frameworks. To address this, we propose a novel MARL framework named Scalable and Heterogeneous Proximal Policy Optimization (SHPPO), integrating heterogeneity into parameter-shared PPO-based MARL networks. We first leverage a latent network to learn strategy patterns for each agent adaptively. Second, we introduce a heterogeneous layer to be inserted into decision-making networks, whose parameters are specifically generated by the learned latent variables. Our approach is scalable as all the parameters are shared except for the heterogeneous layer, and gains both inter-individual and temporal heterogeneity, allowing SHPPO to adapt effectively to varying scales. SHPPO exhibits superior performance in classic MARL environments like Starcraft Multi-Agent Challenge (SMAC) and Google Research Football (GRF), showcasing enhanced zero-shot scalability, and offering insights into the learned latent variables' impact on team performance by visualization.

Recent real-time semantic segmentation models, whether single-branch or multi-branch, achieve good performance and speed. However, their speed is limited by multi-path blocks, and some depend on high-performance teacher models for training. To overcome these issues, we propose Golden Cudgel Network (GCNet). Specifically, GCNet uses vertical multi-convolutions and horizontal multi-paths for training, which are reparameterized into a single convolution for inference, optimizing both performance and speed. This design allows GCNet to self-enlarge during training and self-contract during inference, effectively becoming a "teacher model" without needing external ones. Experimental results show that GCNet outperforms existing state-of-the-art models in terms of performance and speed on the Cityscapes, CamVid, and Pascal VOC 2012 datasets. The code is available at https://github.com/gyyang23/GCNet.

Temporal moment localization in videos using natural language (TMLVNL) is challenging problem in computer vision. TMLVNL aims to determine the correct moment in a lengthy, untrimmed video that corresponds to the input query. In addition to its inherent complexity, TMLVNL encounters several additional difficulties that can degrade performance, such as rare object positions, occlusions, camera focus issues, and motion blurriness. To address these issues, this study proposes a novel solution called the Temporal Ziggurat Transformer Network. First, we developed a novel method rather than relying on various combinational approaches. Second, we proposed complicated scenarios, such as unusual object postures, object occlusions, camera focus issues, and motion blurriness, by incorporating specialized blocks into our Customized Ziggurat Transformer (ZT) to thoroughly explore visual features. Third, to facilitate the understanding of visual features associated with query words, we proposed a query word-specific transformer (QWST) as a submodule of ZT. QWST integrates query word feature representations with extensively investigated visual features. Fourth, in our module named STDF, we managed query-sentence representations along with query word attributes to extract semantic context from video chunks. The moment was then localized, with its start and end borders identified using the moment localization module. Comprehensive experiments on the Charades-STA, TACos, and Activity-Netcaption datasets demonstrated that our strategy outperformed existing state-of-the-art methods.

Recent advances in approximate model predictive control (MPC) have leveraged deep neural networks (DNNs) to imitate optimal control laws, significantly reducing the computational cost of real-time optimization. However, these methods often struggle to adapt efficiently to previously unseen system dynamics and suffer from poor data efficiency in real time. In this paper, we address these challenges by treating control problems with different system parameters as separate tasks within a meta-learning scope. Under this scope, we propose OdML, an Optimization-driven Meta-Learning approach for approximate MPC that enables rapid online fine-tuning of DNNbased controllers. We introduce an optimization-driven fine-tuning mechanism that allows for fast adaptation to new tasks using limited online data, without requiring full knowledge of the system parameters. Furthermore, to ensure safety during adaptation, we incorporate control barrier functions into the optimization process, allowing the controller to satisfy safety constraints even under previously unseen conditions. We demonstrate the effectiveness of OdML through various simulation scenarios, highlighting its ability to achieve safe and efficient control.

Few-shot learning has achieved great success in recent years, thanks to its requirement of limited number of labeled data. However, most of the state-of-the-art techniques of few-shot learning employ transfer learning, which still requires massive labeled data to train a meta-learning system. To simulate the human learning mechanism, a deep model of few-shot learning is proposed to learn from one, or a few examples. First of all in this paper, we analyze and note that the problem with representative semi-supervised few-shot learning methods is the negligence of intra-class and inter-class scatter. To address this issue, we propose a new semi-supervised few-shot learning method with triplet metric learning and Kullback-Leibler optimization, in which KL divergence is employed to dynamically determine the inter-class margin; whereas the triplet metric learning is employed to achieve intra-class clustering. In training, the deep learning and expectation-maximization algorithm are used to optimize models. Intensive experiments have been conducted on three popular benchmark datasets, and the experimental results show that this method significantly improves the classification ability of few-shot learning tasks and obtains the most advanced performance.

Semi-supervised few-shot learning aims to exploit limited number of labeled samples and massive unlabeled samples to improve the model classification capability. One of the mainstream methods is to obtain class means as prototypes to characterize each class. Nonetheless, these mean-of-class prototypes can exhibit bias due to sparse or outlier samples. To mitigate this, we propose Critical-Descriptor Prototypes (CDP), a novel framework that refines prototype learning through dictionary-driven feature sparsification and supervised contrastive optimization. Specifically, CDP introduces two key innovations: firstly, Dictionary-based Descriptor Sparsification: Instead of directly averaging embeddings, CDP decomposes each prototype into 25 discriminative descriptors and reconstructs unlabeled samples via sparse coding. This forces the model to prioritize high-attribution local features by minimizing reconstruction errors. Secondly, Supervised Contrastive Learning: Leveraging pseudo-labels from CDP-refined prototypes, we design a supervised contrastive loss that explicitly clusters samples from the same class while dispersing inter-class features. Extensive experiments on mini-ImageNet and tiered-ImageNet demonstrate state-of-the-art results. Notably, CDP outperforms our prior work (SS-VPE) by 13.94% in 1-shot scenarios, validating its superiority in mitigating prototype bias through critical feature localization.

Few-shot learning has achieved great success in recent years, thanks to its requirement of limited number of labeled data. However, most of the state-of-the-art techniques of few-shot learning employ transfer learning, which still requires massive labeled data to train. To simulate the human learning mechanism, a deep model of few-shot learning is proposed to learn from one, or a few examples. First of all in this paper, we analyze and note that the problem with representative semi-supervised few-shot learning methods is getting stuck in local optimization and prototype bias problems. To address these challenges, we propose a new semi-supervised few-shot learning method with Convex Kullback-Leibler and critical descriptor prototypes, hereafter referred to as CKL. Specifically, CKL optimizes joint probability density via KL divergence, subsequently deriving a strictly convex function to facilitate global optimization in semi-supervised clustering. In addition, by incorporating dictionary learning, the critical descriptor facilitates the extraction of more prototypical features, thereby capturing more distinct feature information and avoiding the problem of prototype bias caused by limited labeled samples. Intensive experiments have been conducted on three popular benchmark datasets, and the experimental results show that this method significantly improves the classification ability of few-shot learning and obtains the most advanced performance. In the future, we will explore additional methods that can be integrated with deep learning to further uncover essential features within samples.

Phase images often contain more comprehensive information compared to amplitude images. The precise visualization of phase images generated from light scattering by submicron particles is paramount for investigating the underlying scattering mechanisms and exploring applications centered around these submicron particles. This study showcases the efficacy of neural networks in acquiring the capacity to conduct phase reconstruction and recover singularities through appropriate training. Our deep learning-based approach introduces an entirely novel framework for phase recovery, accomplished by learning the distribution of scattered light fields from submicron particles, as obtained through a polarized indirect microscopic imaging (PIMI) system, within a custom Poincaré sphere (phase space). We validate this method by successfully reconstructing phase images of polystyrene (PS) balls with varying radii, and comparing diverse network structures and data flows in the process. Furthermore, we substantiate our experimental results by corroborating them with finite difference time domain (FDTD) simulations. These findings underscore the remarkable regularity within the distribution of phase space data, opening up new avenues for advancing the study of nanostructures.

In current Retinex-based low-light image enhancement (LLIE) methods, fine-tuning regularization parameters for Retinex decomposition and illumination estimation can be cumbersome. To address this, we present a novel non-regularization self-supervised Retinex approach for illumination estimation. Our contributions are twofold: First, we introduce a self-supervised method that incorporates edge-aware smoothness properties in bilateral learning, eliminating the need for regularization terms and simplifying parameter adjustments. Second, to enforce smoothness constraints on the estimated bilateral grid, we propose a bilateral grid parameterization network. This network employs a generative encoder to parameterize the bilateral grid of illumination and a trainable slicing layer guided by a map, reconstructing the grid into an illumination map. Despite the absence of regularization terms, our model excels in generating piece-wise smooth illumination, resulting in enhanced naturalness and improved contrast in images. Our model offers exceptional flexibility by eliminating the need for additional regularization terms and parameter fine-tuning. Moreover, it does not depend on external datasets for training, overcoming dataset collection challenges. Extensive experiments, comparing our model with eight state-of-the-art methods across five public available datasets, unequivocally demonstrate our model's state-of-the-art performance based on key metrics such as NIQE, NIQMC, and CPCQI. These results reaffirm the effectiveness of our approach in low-light image enhancement. Code will be available at: https://github.com/zhaozunjin/NeurBR.

Few-shot learning has achieved great success in many fields, thanks to its requirement of limited number of labeled data. However, most of the state-of-the-art techniques of few-shot learning employ transfer learning, which still requires massive labeled data to train a meta-learning system. To simulate the human learning mechanism, a deep model of few-shot learning is proposed to learn from one, or a few examples. First of all in this paper, we analyze and note that the problem with representative semi-supervised few-shot learning methods is getting stuck in local optimization and the negligence of intra-class compactness problem. To address these issue, we propose a novel semi-supervised few-shot learning method with Convex Kullback-Leibler, hereafter referred to as CKL, in which KL divergence is employed to achieve global optimum solution by optimizing a strictly convex functions to perform clustering; whereas sample selection strategy is employed to achieve intra-class compactness. In training, the CKL is optimized iteratively via deep learning and expectation-maximization algorithm. Intensive experiments have been conducted on three popular benchmark data sets, take miniImagenet data set for example, our proposed CKL achieved 76.83% and 85.78% under 5-way 1-shot and 5-way 5-shot, the experimental results show that this method significantly improves the classification ability of few-shot learning tasks and obtains the start-of-the-art performance.

Semantic segmentation plays a key role in applications such as autonomous driving and medical image. Although existing real-time semantic segmentation models achieve a commendable balance between accuracy and speed, their multi-path blocks still affect overall speed. To address this issue, this study proposes a Reparameterizable Dual-Resolution Network (RDRNet) dedicated to real-time semantic segmentation. Specifically, RDRNet employs a two-branch architecture, utilizing multi-path blocks during training and reparameterizing them into single-path blocks during inference, thereby enhancing both accuracy and inference speed simultaneously. Furthermore, we propose the Reparameterizable Pyramid Pooling Module (RPPM) to enhance the feature representation of the pyramid pooling module without increasing its inference time. Experimental results on the Cityscapes, CamVid, and Pascal VOC 2012 datasets demonstrate that RDRNet outperforms existing state-of-the-art models in terms of both performance and speed. The code is available at https://github.com/gyyang23/RDRNet.