Neurocomputing最新文献

Recently, online learning with imbalanced data streams has aroused wide concern, which reflects an uneven distribution of different classes in data streams. Existing approaches have conventionally been conducted on stationary feature space and they assume that we can obtain the entire labels of data streams in the case of supervised learning. However, in many real scenarios, e.g., the environment monitoring task, new features flood in, and old features are partially lost during the changing environment as the different lifespans of different sensors. Besides, each instance needs to be labeled by experts, resulting in expensive costs and scarcity of labels. To address the above problems, this paper proposes a novel Online Imbalance learning with unpredictable Feature evolution and Label scarcity (OIFL) algorithm. First, we utilize margin-based online active learning to selectively label valuable instances. After obtaining the labels, we handle imbalanced class distribution by optimizing F-measure and transforming F-measure optimization into a weighted surrogate loss minimization. When data streams arrive with augmented features, we combine the online passive-aggressive algorithm and structural risk minimization to update the classifier in the divided feature space. When data streams arrive with incomplete features, we leverage variance to identify the most informative features following the empirical risk minimization principle and continue to update the existing classifier as before. Finally, we obtain a sparse but reliable learner by the strategy of projecting truncation. We derive theoretical analyses of OIFL. Also, experiments on the synthetic datasets and real-world data streams to validate the effectiveness of our method.

In the context of multi-heterogeneous UAV coverage path planning, an effective solution method has been proposed. Firstly, regions are set up as fully connected graphs which are cut into multiple subgraphs by spectral clustering method to assign tasks to multi-heterogeneous UAVs. Additionally, an RL-based hyper-heuristic algorithm is proposed. Heuristic space is parameterized by GNN which is trained with the reward provided by the optimization goal to automate design and enhance the heuristic metrics, avoiding the inefficiency and suboptimality of expert design and manual parameter tuning. Compared with existing methods, the proposed algorithm has a better performance in task completion time, execution time and deviation rate, which shows its potential application in the coverage path planning problem of multi-heterogeneous UAVs.

Medical image segmentation has achieved impressive results thanks to U-Net or its alternatives. Yet, most existing methods perform segmentation by classifying individual pixels, tending to ignore the shape-intensity prior information. This may yield implausible segmentation results. Besides, the segmentation performance often drops greatly on unseen datasets. One possible reason is that the model is biased towards texture information, which varies more than shape information across different datasets. In this paper, we introduce a novel Shape-Intensity-Guided U-Net (SIG-UNet) for improving the generalization ability of variants of U-Net in segmenting medical images. Specifically, we adopt the U-Net architecture to reconstruct class-wisely averaged images that only contain the shape-intensity information. We then add an extra similar decoder branch with the reconstruction decoder for segmentation, and apply skip fusion between them. Since the class-wisely averaged image has no any texture information, the reconstruction decoder focuses more on shape and intensity features than the encoder on the original image. Therefore, the final segmentation decoder has less texture bias. Extensive experiments on three segmentation tasks of medical images with different modalities demonstrate that the proposed SIG-UNet achieves promising intra-dataset results while significantly improving the cross-dataset segmentation performance. The source code will be publicly available after acceptance.

Pseudo-labeling is a well-studied approach in semi-supervised learning. However, unreliable or potentially incorrect pseudo-labels can accumulate training errors during iterative self-training steps, leading to unstable performance. Addressing this challenge typically involves either discarding unreliable pseudo-labels, resulting in the loss of important data, or attempting to refine them, risking the possibility of worsening the pseudo-labels in some cases/pixels. In this paper, we propose a novel method based on pseudo-labeling for semi-supervised segmentation of medical images. Unlike existing approaches, our method neither discards any data nor worsens reliable pseudo-labels. Our approach generates uncertainty masks for the predictions, utilizing reliable pixels without any modification as ground truths and modifying the unreliable ones rather than discarding them. Furthermore, we introduce a novel loss function that incorporates both mentioned parts by multiplying each term by its corresponding uncertainty mask, encompassing reliable and unreliable pixels. The reliable pixels are addressed using a masked cross-entropy loss function, while the modification of the unreliable pixels is performed through a deep-learning-based adaptation of active contours. The entire process is solved within a single loss function without the need to solve traditional active contour equations. We evaluated our approach on three publicly available datasets, including MRI and CT images from cardiac structures and lung tissue. Our proposed method outperforms the state-of-the-art semi-supervised learning methods on all three datasets. Implementation of our work is available at https://github.com/behnam-rahmati/Semi-supervised-medical.

Recently, in the field of Hypergraph Neural Networks (HGNNs), the effectiveness of dynamic hypergraph construction has been validated, which aims to reduce structural noise within the hypergraph through embeddings. However, the existing dynamic construction methods fail to notice the reduction of information contained in the hypergraphs during dynamic updates. This limitation undermines the quality of hypergraphs. Moreover, dynamic hypergraphs are constructed from graphs. Several key nodes play a crucial role in graph, but they are overlooked in hypergraphs. In this paper, we propose a Purity Skeleton Dynamic Hypergraph Neural Network (PS-DHGNN) to address the above issues. Firstly, we leverage purity skeleton method to dynamically construct hypergraphs via the fusion embeddings of features and topology simultaneously. This method effectively reduces structural noise and prevents the loss of information. Secondly, we employ an incremental training strategy, which implements a batch training strategy based on the importance of nodes. The key nodes, as the skeleton of hypergraph, are still highly valued. In addition, we utilize a novel loss function for learning structure information between hypergraph and graph. We conduct extensive experiments on node classification and clustering tasks, which demonstrate that our PS-DHGNN outperforms state-of-the-art methods. Note on real-world traffic flow datasets, PS-DHGNN demonstrates excellent performance, which is highly meaningful in practice.

Real-world data often exhibits a long-tailed distribution, where the number of training samples for head classes far exceeds that of tail classes. This class imbalance phenomenon poses significant challenges for training deep neural networks. Existing class-aware loss methods typically focus only on the numerical relationship between class samples, blindly favoring the optimization of tail classes during the process, while neglecting the difficulty of samples and the similarity between the current class and other classes. To this end, relying only on the number relationship can easily lead to over-fitting of tail classes, thereby failing to fully utilize the potential information in the data. Therefore, we propose the Multi-Factor Margin Loss (MFMLoss), which consists of positive margin loss and negative margin loss. MFMLoss comprehensively considers three factors at three levels: overall, class, and sample: (1) quantitative relationships, (2) inter-class similarity relationships, and (3) sample recognition difficulty. The combined consideration of these three factors enables the model to pay more attention to confusing classes and difficult samples during the training process, rather than solely on tail classes, thus achieving optimization from coarse-grained to fine-grained. To further mitigate the negative impact of the imbalance between head and tail classes on feature learning, we design a new network architecture, called F-BREN. F-BREN consists of two components: the feature balancing network and the feature re-enhancement network. The former is trained with negative margin loss, which reduces the recognizability of easy samples. The latter is trained with positive margin loss, using positive margin to give more attention to hard samples, thus balancing the model’s attention to all samples. We conducted extensive experiments on four long-tailed benchmark datasets: CIFAR10-LT, CIFAR100-LT, ImageNet-LT and iNaturalist 2018, comparing the recognition accuracy of our method with eight state-of-the-art methods. The experimental results demonstrate that our proposed method outperforms the eight compared methods.

Due to the coarse granularity of information extraction in image-level annotation-based weakly supervised semantic segmentation algorithms, there exists a significant gap between the generated pseudo-labels and the real pixel-level labels. In this paper, we propose the DeFB-SV framework, which consists of a dual-branch Siamese network structure. This framework separates the foreground and background of images by generating unified resolution and mixed resolution class activation maps, which are then fused to obtain pseudo-labels. The mixed-resolution class activation maps are produced by a new mixed-resolution patch partition method, where we introduce a semantically heuristic patch scorer to divide the image into patches of different sizes based on semantics. Additionally, a novel multi-confidence region division mechanism is proposed to enable the adaptive extraction of the effective parts of pseudo-labels, further enhancing the accuracy of weakly supervised semantic segmentation algorithms. The proposed semantic segmentation framework, DeFB-SV, is evaluated on the PASCAL VOC 2012 and MS COCO 2014 datasets, demonstrating comparable segmentation performance with state-of-the-art methods.

Existing facial expression recognition (FER) techniques rely primarily on seven coarse-grained emotions as emotional labels, which are insufficient to cover the subtle changes in human emotions in the real world. We use 135 fine-grained emotions as emotional benchmarks to address the problem of highly semantically similar fine-grained emotion recognition. In this work, we propose a robust emotion knowledge-based fine-grained (EK-FG) emotion recognition network that captures inter-class relationships and discriminative representations of fine-grained emotions through two prior-based losses: coarse-grained hierarchical loss and fine-grained semantic loss. Specifically, the coarse-grained hierarchical loss obtains a structured semantic representation of fine-grained emotions based on prior knowledge, and captures inter-class relationships through effective category-level push-pull to obtain discriminative representations. The fine-grained semantic loss provides more accurate measurement information for semantic features based on prior knowledge, and enhances the model’s discriminative ability for subtle facial expression differences through regression constraints. Extensive experimental results on the Emo135 dataset demonstrate that EK-FG can effectively overcome the class ambiguity of fine-grained emotion.

Online education is developing rapidly driven by artificial intelligence technology. The massive learning resources lead to information overload and low resource utilization. Intelligent tutoring system (ITS) plays a vital role in the education platform, providing personalized learning services for students. The data obtained from the online education platform has complex correlations, which can be potentially transformed into multi-level graph structures. In recent years, graph neural networks (GNNs) have been tried to be introduced into intelligent learning services due to their superior performance in processing graph-structured data. This paper aims to provide researchers and engineers with a general overview of modeling processes and techniques for intelligent learning services based on GNNs. Through a careful review of the advanced models published between 2019 and 2023, existing research primarily focuses on four detailed areas within the smart services scenario. The GNN models involved are systematically classified, and the principles, pioneers and variants of various models are summarized in detail. Simultaneously, this paper analyzes the applications, the specific problems to be solved, and the technologies and innovations of graph-based models in the four key areas. In addition, we examine the commonly used datasets and evaluation metrics in the field of education. Finally, the current challenges and future development trends are summarized to provide comprehensive and in-depth guidance for research in related fields.

Cooperative behavior assumes a critical role in resolving conflicts arising between collective and individual interests, while punishment measures serve as a robust deterrent against opportunistic free-riding. Within this context, evolutionary game theory (EGT) emerges as an indispensable paradigm for addressing this multifaceted issue. When it comes to introspection behaviors, reinforcement learning (RL) methods exhibit remarkable capabilities to capture agents’ cognitive processes. Nonetheless, previous research has often focused on a static and time-invariant update rule, neglecting the dynamic nature of real-world scenarios where individuals can flexibly transit between strategies in periodic time-dependent patterns. Here, we propose periodic update rules with Q-learning algorithm and game transition model with a punishment mechanism that grants cooperative agents the autonomy to exercise discretion in deciding whether to initiate punishment actions. The agents display dynamic rules periodically through game model transitions, thus ensuring EGT’s inherent adaptability. By employing Monte Carlo (MC) simulations, we analyze the emergence of cooperation that underscores the substantial enhancement of cooperative behavior through the proposed periodic update rules with Q-learning algorithm and game transitions in the presence of punishment. Our study highlights the indispensable significance of appropriate periodic intervals for updating rules and determining optimal punishment costs in the game transition model as critical elements for fostering the evolution of cooperation in real-world scenarios.