IEEE transactions on neural networks and learning systems最新文献

Accurate segmentation of intracranial arteries in digital subtraction angiography (DSA) sequences is critical for cerebrovascular diagnosis but remains challenging due to limited annotations and complex vascular structures. We propose the temporal-adaptive fusion (TAF)-Net, a semisupervised dual-path framework that integrates a vision foundation model, MedSAM, and a task-specific UNet to leverage anatomical priors and fine-grained vascular features. To address interframe inconsistency and vessel discontinuity, we introduce a TAF strategy that dynamically fuses model predictions based on framewise confidence and temporal priors. In addition, we design a spatiotemporal topology-aware loss to enforce structural continuity by penalizing critical disconnection components across frames. Extensive experiments on two public multiframe DSA datasets (DIAS and DSCA) demonstrate that the TAF-Net consistently outperforms state-of-the-art methods in both overlap accuracy (DSC, IoU) and topological integrity (Cost, 95HD), especially under low-label regimes.

Although existing bipartite graph-based multiview clustering (MVC) methods effectively exploit the structural relationships within multiview data, they exhibit three major limitations: 1) they primarily focus on direct similarities between data points and anchors, neglecting underlying neighborhood structures; 2) most existing methods fail to capture high-order correlations across bipartite graphs from different views; and 3) they overlook the relationships among anchor points, limiting the discriminative power of the learned graph. To address these challenges, we propose a unified framework, termed enhanced multiorder bipartite graph learning (EMOBGL) for MVC. The proposed EMOBGL method first constructs a second-order bipartite graph (SOBG) to capture both local and neighboring structural relationships between data points and anchors through first-order similarity (FOS) and second-order similarity (SOS). Then, the tensor Schatten- $p$ regularizer is incorporated to construct a multiorder bipartite graph (MOBG) to capture third-order similarity (TOS) across views. Meanwhile, the anchor structure regularization (ASR) is introduced to model anchor-anchor interactions, further enhancing the structural expressiveness and discriminability of the bipartite graph. The resulting EMOBGL model effectively integrates multiorder and multiview relationships within a unified framework, achieving robust and discriminative clustering performance. An efficient alternating direction method of multipliers (ADMMs) is developed to optimize the model, and we theoretically prove that the solution converges to a Karush-Kuhn-Tucker (KKT) stationary point. Extensive comparative experiments on 13 benchmark datasets demonstrate that the proposed EMOBGL consistently outperforms 13 state-of-the-art methods in both clustering accuracy and robustness. The source code is available at https://github.com/DongHuangTaiYi871/EMOBGL.

Multi-view clustering (MVC) has attracted significant attention in recent years due to its ability to leverage heterogeneous features from multiple views. However, existing methods often lack the ability to jointly model first-order and topological relationships, which is crucial for capturing a more comprehensive clustering structure. In this article, we propose a novel multi-view clustering method with hybrid-order similarity learning (MCHL), which integrates multiple view-specific graphs while considering their first-order and topological correlations, and iteratively learns the view weights and the consensus graph within a unified framework. In addition, we impose a connectivity constraint on the consensus graph to ensure that data points belonging to the same cluster are properly connected within the same component. Extensive experiments on multiple benchmark datasets demonstrate the superior clustering performance of MCHL over the state-of-the-art methods.

Directed hypergraphs have gained increasing attention for modeling group interactions while preserving directionality. However, link prediction in directed hypergraphs has rarely been studied despite its practical significance in complex systems analysis. Existing models perform poorly due to three major challenges in directed hypergraphs: 1) lacking effective feature initialization methods; 2) neglecting to detect higher order substructures; and 3) failing to capture long-range dependencies among vertices. To address these challenges, we propose a novel directed hypergraph motif-based neural network (DHMNN) for directed hyperlink prediction, which simultaneously captures higher order structural and connectivity information from the directed hypergraph topology. First, we introduce directed hypergraph motifs (DH-motifs) to explore higher order neighborhoods, analyzing vertex structural equivalence and generating structural features. Secondly, we utilize hypergraph incidence matrices to measure local connectivity, quantifying vertex co-occurrences and producing connectivity features. Then, we employ hypergraph attention to refine the vertex features at both global and local levels, further capturing long- and short-range dependencies. Finally, a new scoring layer is designed to assess the reliability of each link, considering its local properties, feature variance, and directionality. Extensive experiments on seven metabolic networks and three social networks demonstrate that DHMNN significantly and consistently outperforms state-of-the-art models, achieving a 3.40%-9.90% increase in accuracy. Our code is available at: https://github.com/XihangMeng/DHMNN.

Echo state networks (ESNs) are a particular type of untrained recurrent neural networks (RNNs) within the reservoir computing (RC) framework, popular for their fast and efficient learning. However, traditional ESNs often struggle with long-term information processing. In this article, we introduce a novel class of deep untrained RNNs based on temporal residual connections, called deep residual ESNs (DeepResESNs). We show that leveraging a hierarchy of untrained residual recurrent layers significantly boosts memory capacity and long-term temporal modeling. For the temporal residual connections, we consider different orthogonal configurations, including randomly generated and fixed-structure, and study their effect on network dynamics. Athorough mathematical analysis outlines necessary and sufficient conditions to ensure stable dynamics within DeepResESN. Empirically, the proposed approach consistently outperforms traditional shallow and deep RC on a variety of time-series tasks. Overall, DeepResESN offers a promising approach for designing hierarchical ESNs with better prediction accuracy on long sequences, without sacrificing the computational advantages that make RC attractive.