Visible-thermal cross-modality learning enhances downstream task performance by integrating information from multiple sources. In real-world scenarios such as autonomous driving, new classes continually emerge, and data is often incomplete due to sensor occlusions. This raises a key question about how to incrementally update a model with incomplete cross-modality data. To address this problem, we propose a practical task termed incomplete cross-modality class-incremental learning (ICMCIL), which aims to effectively leverage incomplete cross-modality information to learn new knowledge without forgetting the old. We construct a benchmark for ICMCIL and thoroughly analyze its challenges, revealing that (1) different modalities experience varying degrees of forgetting, (2) conventional cross-modality fusion only partially alleviates forgetting, and (3) missing data exacerbates the forgetting of previous classes. To address these issues, we propose Hybrid Fusion via Completion (HFC), a unified framework that integrates completion, fusion, and forgetting prevention. Additionally, we enhance information fusion by introducing a feature interchange mechanism, wherein features are shuffled and channels are reordered to improve information flow. Extensive experiments demonstrate that HFC effectively addresses ICMCIL, significantly mitigating modality forgetting.

Clustering data with anisotropic distributions and complex local structures remains a significant challenge, particularly for models that rely on homogeneity assumptions. Such methods often struggle to capture directional variations, leading to blurred cluster boundaries and degraded performance on non-uniform or large-scale datasets. To address these limitations, we introduce EB-DP, a structure-adaptive clustering algorithm that integrates ellipsoidal ball modeling with a multi-scale hybrid splitting strategy. EB-DP constructs ellipsoidal granules through eigendecomposition of local covariance matrices, enabling alignment with the principal directions of data variation. A hierarchical splitting scheme combining axis-guided projection and two-means refinement ensures coarse-to-fine partitioning while preserving structural consistency. To improve cluster center identification and label assignment, during the clustering decision-making stage, a triple constraint mechanism incorporating ellipsoidal density, inter-ball distance, and axis-angle is introduced. Extensive experiments on synthetic and real-world datasets demonstrate that EB-DP generally achieves better ACC and NMI than representative baselines, while also reducing computational overhead.

Single-domain generalization object detection (S-DGOD) in remote sensing trains a detector using labeled data from a single source domain and deploys it to unseen domains, where performance often drops under cross-region distribution shift. Existing methods mainly enforce domain-invariant constraints or enlarge source diversity, but they rarely model object-background co-occurrences that become shortcuts when regional context changes. This letter proposes a stability-enhanced S-DGOD framework that integrates spatially aware stable learning with a perturbation-and-averaging training mechanism. The Spatially-Aware Stable Learning (SASL) module constructs paired object and surrounding-background features for each RoI and learns instance weights to reduce dependencies within object features and between object and background features, thereby suppressing region-specific shortcut correlations. To stabilize the high-variance reweighting process, the QAT-driven Perturbation and Ensemble (QPE) mechanism uses quantization-aware training as a structured training-time perturbation to explore neighboring decorrelated solutions, and then applies stochastic weight averaging to aggregate them into a single inference-time detector. Experiments on the RWDS dataset show consistent gains over representative S-DGOD baselines; the proposed method improves the average harmonic mean by 2.3 mAP50 over the best competing method on RWDS-FR. Codes are available at https://github.com/JunhongLu0704/SASL-DGOD.

In the era of deep learning and foundation models, constraints such as limited labels, domain shift, and privacy make learning from scratch impractical, systems must reuse and reorganize prior competence efficiently while integrating heterogeneous evidence robustly under uncertainty. To achieve this, dynamic knowledge transmission across models and scenario-adaptive knowledge fusion from multi-source signals have been widely used to achieve robust generalization. Existing research predominantly focuses on isolated aspects of knowledge transmission or fusion, or adopts fragmented, task-specific perspectives, such as federated learning and knowledge distillation, leading to weak connections and transferability across different fields. We therefore propose a unified framework that provides a high-level perspective of knowledge transmission and fusion across different learning paradigms. We first distinguish two coupled yet conceptually different processes: knowledge transmission, which improves efficiency by transferring sample, feature, or model-level carriers; and knowledge fusion, which improves robustness and coverage via pre-fusion, in-training fusion, or post-fusion. We then provide a task-aligned lookup template that maps practical constraints and requirements to suitable strategies, what/who/how-to-transmit, and why/when/how-to-fuse, further illustrating how these mechanisms manifest in six general tasks. We finally unified evaluation enabling clear comparisons and trade-offs among efficiency, robustness, consistency, and specialization.

Multi-teacher knowledge distillation (MTKD) enhances student model performance by integrating knowledge from multiple teachers. However, existing methods often aggregate teacher knowledge without adequately assessing the reliability of individual teacher contributions or effectively resolving inherent conflicts, which can lead to the student model inheriting low-quality or conflicting knowledge, particularly in challenging fine-grained categories or overlapping concepts. To address this, we propose the Consensus, Conflict, and Coordination Multi-Teacher Knowledge Distillation (C3-MKD) Framework. The C3-MKD framework leverages three core mechanisms: at the feature level, we introduce Prototype-Based Guidance to adaptively select high-quality teacher features, thereby building reliable consensus knowledge; at the logits level, we employ Dempster-Shafer Theory to resolve conflicts; and finally, by intelligently integrating these consensus-building and conflict-resolution mechanisms, C3-MKD meticulously coordinates the contributions of multiple teachers. Extensive experiments on widely adopted benchmarks demonstrate that our proposed framework consistently outperforms existing state-of-the-art methods across various teacher-student architectures.

Class Incremental Learning (CIL) aims to enable models to continually learn new classes while retaining previously learned knowledge. The principal challenge in CIL is catastrophic forgetting, which prior approaches typically address by distilling knowledge from previous model. However, such way is often limited to pairwise alignment, failing to preserve the underlying global manifold structure of feature space—ultimately resulting in semantic drift over time. To capture multi-scale structural patterns in the feature space, we propose a topology-aware distillation framework that leverages persistent homology. Specifically, by enforcing topological alignment across incremental stages, our method ensures structure-consistent knowledge transfer and robust preservation of old classes. Furthermore, we still devise a dual-branch architecture with an inverse sampling and dynamic reweighting mechanism that addresses the inherent data imbalance in standard replay-based frameworks. These innovations coalesce into TaKP (Topology-aware Knowledge Preservation), a unified framework designed to enhance knowledge preservation in CIL. Extensive experiments demonstrate that TaKP achieves state-of-the-art performance on multiple benchmarks, significantly improving old-class preservation and average accuracy.

Few-shot segmentation (FSS) has gained significant attention for its ability to generalize to novel classes with limited supervision, yet remains challenged by structural misalignment and cross-view inconsistency under large appearance or viewpoint variations. This paper tackles these challenges by introducing VINE (View-Informed NEtwork), a unified framework that jointly models structural consistency and foreground discrimination to refine class-specific prototypes. Specifically, VINE introduces a spatial-view graph on backbone features, where the spatial graph captures local geometric topology and the view graph connects features from different perspectives to propagate view-invariant structural semantics. To further alleviate foreground ambiguity, we derive a discriminative prior from the support-query feature discrepancy to capture category-specific contrast, which reweights SAM features by emphasizing salient regions and recalibrates backbone activations for improved structural focus. The foreground-enhanced SAM features and structurally enriched ResNet features are progressively integrated through masked cross-attention, yielding class-consistent prototypes used as adaptive prompts for the SAM decoder to generate accurate masks. Extensive experiments on multiple FSS benchmarks validate the effectiveness and robustness of VINE, particularly under challenging scenarios with viewpoint shifts and complex structures. The code is available at https://github.com/HongliLiu1/VINE-main.

Recent advances in deep learning, particularly generative adversarial networks (GANs), have significantly advanced image inpainting. However, reconstructing structural details and accurate chromaticity distributions in complex scenes remains challenging. This paper proposes a GAN-based method via dual-domain perception and cross-domain collaboration for image inpainting, including spatial-frequency guided attention module, frequency-spatial collaborative attention module and adaptive spatial-channel feature aggregation module. The spatial-frequency guided attention module leverages spatial features to refine frequency-domain components from Fast Fourier Transform, highlighting structural and textural information. The frequency-spatial collaborative attention module optimizes chromaticity and edge details through amplitude and phase guidance from Inverse Fast Fourier Transform. Finally, the adaptive spatial-channel feature aggregation module fuses spatial and channel features, then dynamically modulates chromatic-aware and edge-aware features. Evaluation results on public datasets demonstrate superior performance in reducing chromaticity deviations and edge blurring compared with the state-of-the-art methods. The code will be available at https://github.com/RRChen001/SFNet .

Weakly supervised 3D semantic segmentation has proven effective in alleviating the heavy dependence on dense annotations by generating high-quality pseudo-labels. However, due to the scene complexity and disorder of the point cloud, merely applying the model semantic prediction or hand-crafted feature similarity for pseudo labeling is inefficient and biased. This limitation inevitably results in incorrect pseudo labels. To tackle this challenge, we propose a new method called Multi-granularity Superpoint Graph Learning (MSGL) that leverages the multi-scale local features of point clouds to improve the quality of pseudo labels. We first design a multi-granularity local representation learning module on the superpoint graph to capture the neighboring structure information of each superpoint within complex scenes. Subsequently, the generated structural embedding is utilized to enhance the affinity matrix of label propagation, thereby yielding high-quality pseudo labels. To further enforce the generalization of the structural representation module under scenario changes or data fluctuations, we present a multi-granularity consistency loss in MSGL. This loss is applied across different views of the superpoint graph within each scene to ensure a robust and consistent learning process. Our experiments conducted on three benchmarks show that the proposed method outperforms existing weakly supervised methods under several sparse label settings, and improves the baseline by an average of 7.7% with only 1% extra computation cost. Moreover, our approach even compares favorably to some fully supervised methods with only one point labeled for each thing.

Graph neural networks (GNNs) have recently emerged as a promising approach for solving few-shot learning (FSL) problems, enabling generalization to new categories by establishing associations among limited samples. However, most previous methods focus on local interactions within a layer or between successive layers, overlooking the continuous evolution of node states across deeper layers. Especially in FSL scenarios with scarce data, this oversight leads to excessive compression or loss of node features in deep networks. To address this challenge, we propose a fuzzy cognitive GNN (FCGNN), which models the evolution of node fuzzy state representation across layers by integrating principles from fuzzy cognitive maps. First, we construct a fuzzy cognitive graph (FCG) based on Gaussian similarity to enhance the information flow within the layer. FCG establishes fuzzy relationships between nodes to enhance the connections between different samples within the layer. Then, we design a fuzzy concept evolution (FCE) module based on dual fuzzy state representation, incorporating the fuzzy activation state to illustrate the continuous changes in node states. FCE combines explicit and implicit fuzzy state representation to achieve the dynamic evolution of node concepts between layers. Finally, we map the node states after multiple graph iterations to the category space for FSL image classification. In experiments on four benchmark datasets, FCGNN achieves superior classification performance and outperforms several baseline methods by up to 2.73% on average accuracy.

Multi-label learning (MLL) involves assigning multiple labels to each instance, necessitating feature selection methods that capture both global label dependencies and local label combinations. Existing multi-label feature selection methods often focus on global relevance, overlooking the effects of local label disturbances. Moreover, they typically assess redundancy through inter-feature similarity rather than semantic overlap. To overcome these limitations, we propose the Structure-Aware Multi-Label Feature Selection (SAMLFS) method. First, SAMLFS ranks features based on both global label coupling strength and local disturbance sensitivity, reflecting reductions in joint-label uncertainty and shifts in label distributions. It then partitions the data into label-induced subcontexts, constructs a concept lattice within each subspace, and prunes low-contribution concepts using a contribution function that combines feature participation with label-consistency entropy. Additionally, SAMLFS computes structural coverage to evaluate each feature’s stability across pruned concept families and integrates the retained structural signals with the original feature space via a lightweight matching vector, generating a compact and interpretable representation. Experiments across 10 public multi-label benchmark datasets show that SAMLFS achieves competitive results in Average Precision, Coverage, Ranking Loss, One Error, and Micro-F1. Statistical tests at the 0.05 significance level confirm the superiority of SAMLFS over competing methods. By grounding feature selection in concept-level coverage, SAMLFS enables interpretable and structure-aware multi-label learning.

Multi-sequence magnetic resonance imaging (MRI) is essential for clinical diagnosis because it enables comprehensive characterization of complex anatomy. However, its substantial heterogeneity limits the generalizability of deep learning models and hinders clinical translation. Here we present MARS, a large-scale MRI foundation model with a novel pretraining strategy that disentangles anatomy-invariant features from sequence-specific variations to learn robust and generalizable representations for diverse clinical applications. We collected 64 datasets spanning 10 anatomical structures and multiple MRI sequences. Among these, 336,476 volumetric scans from 34 datasets (8 public and 26 private) were curated to build a large multi-organ, multi-sequence MRI pretraining corpus. We further established a benchmark of 44 downstream tasks covering diagnosis, segmentation, registration, progression prediction and report generation. MARS ranked first in 41 of 44 benchmarks, with statistically significant improvements. Its strong performance on heterogeneous and external datasets underscores MARS as a scalable foundation for versatile real-world multi-sequence MRI analysis.

Graph-structured data typically exhibits complex topological heterogeneity, making it difficult to model accurately within a single Riemannian manifold. While emerging mixed-curvature methods attempt to capture such diversity, they often rely on implicit, task-driven routing that lacks fundamental geometric grounding. To address this challenge, we propose a Geometric Mixture-of-Experts framework (GeoMoE) that adaptively fuses node representations across diverse Riemannian spaces to better accommodate multi-scale topological structures. At its core, GeoMoE leverages Ollivier-Ricci Curvature (ORC) as an intrinsic geometric prior to orchestrate the collaboration of specialized experts. Specifically, we design a graph-aware gating network that assigns node-specific fusion weights, regularized by a curvature-guided alignment loss to ensure interpretable and geometry-consistent routing. Additionally, we introduce a curvature-aware contrastive objective that promotes geometric discriminability by constructing positive and negative pairs according to curvature consistency. Extensive experiments on six benchmark datasets demonstrate that GeoMoE outperforms state-of-the-art baselines across diverse graph types.

Class incremental learning (CIL) aims to enable models to continuously learn new classes without catastrophically forgetting old ones. A promising direction is to learn and use prototypes of classes during incremental updates. Despite simplicity and intuition, we find that such methods suffer from inadequate representation capability and unsatisfied feature overlap. These two factors cause class-wise confusion and limited performance. In this paper, we develop a Confusion-REduced AuTo-Encoder classifier (CREATE) for CIL. Specifically, our method employs a lightweight auto-encoder module to learn compact manifold for each class in the latent subspace, constraining samples to be well reconstructed only on the semantically correct auto-encoder. Thus, the representation stability and capability of class distributions are enhanced, alleviating the potential class-wise confusion problem. To further distinguish the overlapped features, we propose a confusion-aware latent space separation loss that ensures samples are closely distributed in their corresponding low-dimensional manifold while keeping away from the distributions of features from other classes. Our method demonstrates stronger representational capacity and discrimination ability by learning disentangled manifolds and reduces class confusion. Extensive experiments on multiple datasets and settings show that CREATE outperforms other state-of-the-art methods up to 5.41%. The code is available at https://github.com/lilyht/CREATE.

Weakly supervised 3-D semantic segmentation has successfully mitigated the labor-intensive and time-consuming task of annotating 3-D point clouds. However, reliably utilizing the minimal point-wise annotations for unlabeled data in complex and large-scale scenes is still challenging, such as only 20 points labeled in 2 million points. To tackle this challenge, we propose a new Uncertainty-aware Superpoint Graph Transformer (UaSGT) framework that utilizes minimal annotations for unlabeled data learning through reliable long-range supervision propagation from labeled superpoints to unlabeled superpoints. First, we propose a superpoint graph transformer to achieve long-range supervision propagation along the attention-based fuzzy subsets defined on superpoints. The attention-based fuzzy subset measures the membership of unlabeled superpoints to clusters centered on labeled superpoints. Second, we employ an uncertainty-aware membership rectification technique on the fuzzy subset to ensure reliable propagation among superpoints within the same category. This technique integrates an uncertainty prediction module to mask the influence of unreliable membership and a spatial prior refinement module to reduce uncertainty in intraclass membership degrees. Finally, experimental results on two large-scale benchmarks S3DIS and ScanNet-V2 demonstrate the superiority of our approach compared to the state-of-the-art with at least 90% annotation reduction, and our method also achieves comparable performance to fully supervised methods with less than 0.1% labeled points.

Visible-thermal cross-modality learning has gained significant success by utilizing information from different sources to improve performance in downstream tasks. To make cross-modality learning methods more practical in open-ended environments, such as autonomous driving, a straightforward question arises on how we can incrementally update the model with such cross-modality data. In response, we propose a practical task, namely cross-modality class-incremental learning (CIL). We first provide a cross-modality CIL benchmark, which encompasses datasets abundant in scene diversity and environmental variability, including urban, residential, and diverse temporal settings. Then we thoroughly analyze its challenges and find that: (1) simply using cross-modality learning or CIL methods cannot achieve satisfactory performance in cross-modality CIL; (2) combining existing cross-modality learning methods and CIL strategies benefits cross-modality CIL to a very limited extent; (3) modality forgetting, i.e., neurons favoring a dominant modality and diminishing the weaker, is the significant factor that leads to forgetting in cross-modality CIL. To address the challenges of cross-modality CIL, we design a simple yet effective method, termed Hybrid-fused Forgetting Prevention (HFP). Specifically, HFP first introduces a block-diagonal structured sparsity on the channel dimension and implements the information interchange mechanism by shuffling the feature maps of different groups and changing the sequence of channels, thus fully exploiting the complementarity of cross-modality data. Then, with representation fusion and forgetting prevention strategies, more complementary features can be learned and exploited in retaining old knowledge, which can alleviate modality forgetting. Extensive experiments validate that HFP contributes a favorable framework to cross-modality CIL and alleviates forgetting effectively.

Heterogeneous graphs organize data with nodes and edges, and have been widely used in various graph-centric applications. Often, some data are omitted during manual construction, leading to data reduction and performance degeneration on downstream tasks. Existing methods recover the missing data based on the data already within a single graph, neglecting the fact that graphs from different sources share some common nodes due to scope overlap. In this paper, we concentrate on the missing data recovery task on multi-source heterogeneous graphs under the incremental scenario and design a novel framework to recover the missing data by fusing multi-source complementary data from previously appeared graphs. Our model, namely SIKE, is present with a pre-trained language model and graph-specific adapters. To take advantage of the complementary data of multi-source graphs, we propose an embedding-based data fusion method to gather data among graphs. To evaluate the proposed model, we build two new datasets consisting of multi-source heterogeneous graphs. The experimental results show that our model SIKE achieves significant improvements compared with competitive baseline models, demonstrating the effectiveness of our model and shedding light on multi-source data fusion for data governance.

High-dimensional data is a key study object for both machine learning (ML) and information visualization. In the field of visualization, dimensionality reduction (DR) methods, also known as projections, are one of the most frequently used classes of techniques for visually exploring large and high-dimensional datasets. In ML, high-dimensional data is generated and processed by classifiers and regressors, which increasingly require visualization for explanation and exploration. This thesis focuses on a recent visualization technique called decision maps. A decision map is a 2D image that visualizes the decision boundaries of a classifier in the data space, and can be used to explain and improve the behavior of ML classifiers. Constructing decision maps essentially involves a DR method and its inverse process (inverse projection). As such, ML techniques can help to create decision maps by providing improved (inverse) projections. We begin with a case study applying decision maps to explain the classification of mineral deposit genesis. Our findings show that decision maps provide extra insights into the mineral classification model, aiding geologists in interpreting the model. However, we also identified gaps in current decision map techniques that present opportunities for improvement. Following this case study, we conducted a comprehensive evaluation of three notable decision map techniques. Our evaluation shows that each technique has unique advantages and disadvantages. Our results can guide users in selecting the most suitable technique for specific tasks. A particularly salient finding of this evaluation was that all tested decision maps exhibit a surface-like behavior when applied to a 3D dataset. We explored the aforementioned surface-like behavior of decision maps across more scenarios. By estimating the intrinsic dimensionality of the maps, we found that existing decision map methods cover only a small portion of the intrinsic dimensionality of high-dimensional data spaces. This finding highlights fundamental limitations in all current approaches to constructing decision maps. To address these limitations, we propose a novel approach for computing inverse projections with the support of ML. Our method allows users to interactively control the location of the inversely projected points, thus also of visualizations like decision maps, within the high-dimensional space. In this way, users can explore larger parts of the data space, bypassing the practical limitations imposed by the aforementioned surface-like behavior of decision maps. We demonstrate the effectiveness of our approach through its application to a style transfer task. Finally, we introduce an accelerated computation method for decision maps. Our method significantly reduces the computation time for both basic decision maps and enhanced variations thereof such as gradient maps. The acceleration facilitates the further deployment of such visualizations in interactive visual analytics workflows for classifier engineering.

Class incremental learning (CIL) aims to mitigate catastrophic forgetting of previously learned classes when integrating new knowledge. A primary challenge contributing to forgetting is the absence of data from earlier classes. Researchers have designed a variety of methods to solve the problem, among which topology-preserving methods show tremendous potential. However, two problems remain: first, a large hyperparameter search space for constructing and utilizing a complex topology hinders efficient performance optimization, and second, constraining the network to preserve the topology in the objective makes it difficult to optimize. This article proposes SPARK, a simple and parameter-free method by embedding fuzzy cognitive maps, to address the problems. First, we construct a fuzzy cognitive map with nodes representing class prototypes and edges representing interclass similarities. Then, we exploit the fuzzy cognitive map to obtain class-level embedding by aggregating features of other classes for each class. Finally, the class- and sample-level embeddings are fused and fed to the classifier. The proposed method can be easily optimized without introducing additional loss terms and hyperparameters. We theoretically prove that such a simple fuzzy cognitive map embedding can efficiently preserve the structural information of the fuzzy cognitive map. Experimental results indicate that SPARK achieves up to 5.85% higher average accuracy and 8.69% reduction in forgetting compared to baseline methods.