Broad learning system (BLS) as a novel training framework derived from the random vector functional link neural network (RVFLNN). Unlike RVFLNN, which directly applies raw data to network learning, BLS first transforms input data into feature nodes through feature mapping. These feature nodes are then nonlinearly transformed into enhanced nodes. Both feature nodes and enhanced nodes are concatenated and connected to the output layer, and the corresponding output weights are derived via pseudo-inverse. Due to its shallow architecture and the need to train only the output weights, BLS achieves highly efficient learning capabilities.Moreover, when adding new nodes, BLS does not require retraining from scratch; it only adjusts the weights associated with the new nodes. Different from deep learning, BLS aims to expand the broad rather than the depth of the neural network to solve complex problems. It not only overcomes the drawback of the long-training process in deep learning but also enables the rapid incremental construction of network models. Therefore, BLS has received extensive attention from both the academic and industrial communities. In view of this, this paper conducts a systematic review of BLS. First, we outline the research background of the BLS. Then, we elaborate on the relevant concepts and definitions of BLS. Furthermore, recent advances in BLS are introduced. In addition, we present the extensive applications of BLS in various fields. Finally, several possible development directions for BLS are proposed. © 2025 Elsevier B.V.

The conventional broad learning system (BLS) struggles to represent the complex nonlinear features of hyperspectral images (HSI) due to its reliance on linear sparse feature extraction methods. Additionally, traditional BLS models focus primarily on class separability, ignoring the manifold structure that characterizes relationships between samples. To address these issues, previous research has incorporated graph convolutional networks (GCNs) and manifold learning into the BLS framework, but these methods often emphasize only local manifold structures, overlooking global structural information. In this paper, we propose a Global-Local Graph Convolutional Broad Network (GLGBN) for HSI classification. GLGBN addresses both global and local manifold structures, optimizing the classification boundary by minimizing local scatter and maximizing global scatter. It uses linear discriminant analysis (LDA) to preserve global manifold structure and locality preserving projections (LPP) to model local relationships via a Laplacian graph. This dual approach ensures that similar samples remain close while dissimilar samples are separated, enhancing classification accuracy. The proposed GLGBN model demonstrated outstanding overall accuracy across multiple public datasets: 95.31% on Indian Pines, 97.67% on Pavia University and 98.37% on Salinas, surpassing several classical and state-of-the-art approaches. © 2025

As an extension of random vector functional link neural network (RVFLNN), the broad learning system (BLS) has gradually been applied in various fields due to its advantages of simple network structure, rapid training, and fast model updating. Despite its remarkable achievements, there are still some difficulties and challenges that restrict its further application in practical tasks. In BLS, all samples equally contribute to constructing the optimal classification boundary. However, when training samples contain noise or outliers, it will result in the obtained classification boundary not being the true optimal classification boundary. In other words, BLS is sensitive to noise points or outliers. To address it, we propose a robust rough fuzzy broad learning system with local neighborhood structure (RFBLS). In RFBLS, fuzzy sets are utilized by associating a membership degree with each sample. This incorporation ensures diverse contributions of different samples during the minimization of the objective function. Samples with noise or outliers are assigned smaller weights, effectively eliminating the influence of noise and outlier samples in the learning process. We employ the rough set theory to explore the influence-decision degree of each conditional attribute on the decision attribute, aiming to eliminate the impact of redundant or interfering features. Subsequently, we comprehensively consider the distance between samples and class centers, as well as the local neighborhood structure relationship between samples. Initially, we calculate the initial membership degree of each sample based on the distance to class centers. Then, utilizing fuzzy K-nearest neighbors, we calculate and rank the membership degrees of K nearest neighbors for each sample. These rankings are then weighted with the initial membership degree, achieving the fusion of relationships and membership degree calculations between samples. Finally, extensive comparative experiments are conducted on various UCI datasets and image datasets, and the results indicate that our method exhibits superior noise resistance and classification performance. © 2025

Hyperspectral image (HSI) classification methods based on graph convolutional networks (GCNs) have gained attention due to their ability to process irregular regions using graph encoding techniques. Most existing GCN-based HSI classification methods use multilayer perceptrons (MLPs) for classification, relying primarily on aligning predicted values with actual sample values. However, these methods often overlook inter-class separability and intra-class compactness, limiting their ability to achieve effective inter-class separation and intra-class aggregation, which compromises feature discrimination. Additionally, HSIs exhibit complex spectral characteristics where different substances can share similar spectra, and identical substances may present varying spectra, creating sample uncertainty. Addressing these challenges, we propose a fuzzy discriminative broad fusion network (FDBFN) for HSI classification. FDBFN leverages fuzzy set theory and manifold learning to calculate a membership matrix, capturing the global structure and discriminative information of samples. This matrix enables samples to be classified across categories, capturing their distributional uncertainty. Using this information, we construct inter-class and intra-class scatter matrices and design a loss function that minimizes intra-class distances while maximizing inter-class distances to enhance feature discrimination. FDBFN further employs broad learning in the classification layer, integrating features through feature nodes and enhanced nodes layers for full utilization of network-extracted features. Experimental results show that FDBFN achieves classification accuracies of 97.45%, 98.34%, and 99.50% on the Indian Pines, Pavia University, and Salinas datasets, respectively. Compared to several state-of-the-art methods, FDBFN enhances classification accuracy, robustness, and efficiency in HSI, demonstrating its superiority and adaptability across diverse datasets. © 2024 Elsevier Ltd

In recent years, graph convolutional networks (GCNs) have shown strong performance in hyperspectral image (HSI) classification. However, traditional GCN methods often use superpixel-based nodes to reduce computational complexity, which fails to capture pixel-level spectral-spatial features. Additionally, these methods typically focus on matching predicted labels with ground truth, neglecting the relationships between inter-class and intra-class distances, leading to less discriminative features. To address these issues, we propose a feature fusion fuzzy graph broad network (F3GBN) for HSI classification. Our method extracts pixel-level attribute contour features using attribute filters and fuses them with superpixel features through canonical correlation analysis. We employ a broad learning system (BLS) as the classifier, which fully utilizes spectral-spatial information via nonlinear transformations. Furthermore, we construct intra-class and inter-class graphs based on fuzzy set and manifold learning theories to ensure better clustering of samples within the same class and separation between different classes. A novel loss function is introduced in BLS to minimize intra-class distances and maximize inter-class distances, enhancing feature discriminability. The proposed F3GBN model achieved impressive overall accuracy on public datasets: 96.73% on Indian Pines, 98.29% on Pavia University, 98.69% on Salinas, and 99.43% on Kennedy Space Center, outperforming several classical and state-of-the-art methods, thereby demonstrating its effectiveness and feasibility. © 2024

目前,许多基于深度学习和神经网络的算法被应用于脑电(electroencephalogram, EEG)信号情感识别.然而,现有研究大多采用提取单维脑电信号特征的方法.随着多传感技术的更新,更具全面性和系统性的多维信号特征提取需求出现.本文尝试将复杂网络研究应用到多维情感脑电识别中,提出一种基于水平可视图多元联合模体熵的情感识别算法,该方法可以有效避免人工选取特征对实验结果的影响,保持原始序列的非线性动力学特征.首先利用水平可视图算法将多维情感脑电信号分别转换为多路可视图网络,提取模体熵特征识别情感脑电研究中的关键频带和关键通道.在此基础上,将水平可视图网络两两联合,提取多元水平联合模体熵向量,作为输入参数对情感脑电信号进行识别.由于情感脑电序列长度会对识别效果产生影响,我们将脑电信号切割成大小不一的窗口,对比不同窗口大小对分类准确率的影响.实验结果表明,当切割窗口大小为10 s时,多元水平联合模体熵对情感脑电信号的识别效果最佳,对积极脑电/消极脑电、积极脑电/中性脑电、消极脑电/中性脑电的分类准确率分别达到95.07%, 97.73%, 90.26%,优于其他二维连接参数.同时,三分类的准确率为93.67%,本文算法无论在识别复杂度和准确率上,与已有算法相比均有较大提高.

图像分类作为计算机视觉识别领域的基本研究主题之一,其研究目标是对于一给定的图像和分类标签集合,通过分类方法进行识别后,可以预测出其输入图像的类别标签。整个分类过程中,特征提取和选择概括起来统称为特征表达,而良好的特征表达对提升图像分类精度起到了至关重要的作用。因此,针对图像分类任务,深度学习以其强大的特征提取能力已获得了广泛的关注和应用。而卷积神经网络作为深度学习重要的模型之一,它的权值共享网络结构大大降低了网络模型的复杂度,减少了参数的数量,从而避免了传统识别算法中复杂的特征提取和数据重建过程。但是,仍有一些问题需要解决,例如:（1）梯度弥散。现有的深度神经网络通常采用梯度下降法进行参数训练。但是,随着网络层数的不断增加,训练过程中反向传播梯度逐渐消失,准确率很快达到饱和并迅速下降,从而导致网络性能的退化。（2）网络冗余。通过简单堆叠更多的网络层次可构造出超深层网络。然而,随着结构规模的增大,网络中产生了大量的参数,增加了网络的冗余性,进而导致网络性能的退化。针对上述问题,以提高图像分类精度为目标,设计了更为高效的深度卷积网络模型。主要研究内容如下:第一,在卷积神经网络的学习过程中,卷积核的初始值通常是随机赋值的,这将导致学习过程陷入局部最优。另外,随着网络深度的增加,基于梯度下降法的网络参数学习法则通常会导致梯度弥散现象。针对上述问题,提出一种基于反卷积特征提取的深度卷积网络模型。首先,采用无监督的两层堆叠反卷积网络从原始图像中学习得到特征映射矩阵;然后,将该特征映射矩阵作为深度卷积网络的卷积核对原始图像进行逐层卷积与池化操作;最后,采用附加动量系数的小批次随机梯度下降法对深度卷积网络微调以避免梯度弥散。第二,通过增加深度卷积网络的层数可有效提升分类精度,但随着网络深度的增加,训练过程中梯度逐渐消失,准确率很快达到饱和并迅速下降,从而导致网络性能的退化。针对上述问题,提出一种基于金字塔结构的深度卷积网络模型。首先,各网络层通过逐渐增加特征图维数,以分散集中在受下采样影响的结构单元压力,使其在所有单元中均匀分布;然后,通过探讨结构单元内部堆叠元素之间的顺序,设计出一种金字塔结构单元;最后,通过使用小批次随机梯度下降法进行参数训练,进一步避免梯度弥散。第三,深度融合网络由于具有学习多尺度表达和优化信息流动的能力,可以改善深层网络的训练过程。但是,深度融合网络中的深度并没有最大限度地提升网络整体性能。随着网络深度的增加,准确率很快达到饱和并迅速下降,从而导致网络性能的退化。针对上述问题,提出一种基于多路融合的深度卷积网络模型。首先,以拼接融合的方式组合两种网络结构,生成一个深度融合网络集成结构;然后,与嵌入的学习机制组成一个集成单元,从而改善模型的特征表达能力;最后,在无损性能的前提下,通过引入分组卷积来提升模型的计算效率。第四,通过增加深度卷积网络的深度和宽度可有效提升分类精度。然而,随着网络结构规模的增大,网络中产生了大量的参数,增加了网络的冗余性。针对上述问题,提出一种基于交错融合分组的深度卷积网络模型。首先,使用相同的网络拓扑结构以及拆分-变换-拼接策略构建出一个模板模块,并通过堆叠模板模块形成深度网络;然后,在深度网络中引入小的卷积核和分组卷积来构建出更高效的卷积核;最后,将结构化稀疏卷积核和深度网络组合形成一种高度模块化和轻量级的网络结构。在图像分类标准数据集上进行的对比实验结果表明,所提出的深度卷积网络具有较好的泛化能力,可有效提高图像的分类精度。

零样本图像分类利用已知类数据进行训练,借助辅助信息将已知类的知识迁移到未知类,从而实现对未知类的识别,有效解决了现实中新物种标注不足的问题,促进了图像分类领域的发展,具有巨大的研究和应用价值。本文针对零样本图像分类中存在的领域偏移、属性描述能力不足和枢纽点问题,在语义自编码模型的基础上,结合对比学习、语义约束和视觉约束进行研究,提出了两种基于自编码器的零样本图像分类模型,主要的研究内容如下:1.针对零样本图像分类中的领域偏移问题,提出一种基于未知类语义约束自编码的零样本图像分类模型。首先,利用预训练的Res Net101网络进行特征提取,通过计算得到已知类的视觉中心,并采用无监督聚类算法得到未知类的视觉中心;其次,利用编码器将已知类的视觉中心映射到语义空间,与语义类原型进行对齐;然后,利用解码器将已知类的映射语义向量重构为视觉特征,与视觉中心进行对齐;同时,利用未知类语义约束项对自编码器的训练进行约束;最后,在语义空间中计算待测样本的语义向量与各测试类原型之间的相似性,采用最近邻算法实现零样本图像分类。2.针对零样本图像分类中属性描述能力不足和枢纽点问题,提出一种基于潜在属性扩展与未知类视觉约束自编码的零样本图像分类模型。首先,利用预训练的对比学习模型提取所有对象类的潜在属性,并采用全连接网络将潜在属性与已有的语义属性加权拼接组成混合属性;其次,将视觉空间作为嵌入空间,利用自编码器学习语义空间到视觉空间的映射模型,同时利用未知类视觉约束项对自编码器的训练进行约束;然后,将未知类的混合属性送入训练好的编码器中,得到未知类的预测视觉中心,并将其与已知类的真实视觉中心共同构成视觉空间;最后,在视觉空间中计算待测样本的视觉特征和各测试类视觉中心之间的相似性,采用最近邻算法实现零样本图像分类。本文分别在Aw A2数据集和CUB数据集上进行了实验,并对结果进行了对比分析,验证了本文所提出的两种模型均能有效提升零样本图像分类的性能。论文共包含图30幅,表8个,参考文献106篇。

图像作为信息和数据的重要载体,已广泛渗透于现代生产生活的每个环节。利用机器学习方法对海量图像进行分类等加工处理已成为当今各行业领域生产力提升的迫切需求。零样本图像分类是指在训练集有标签样本类别无法涵盖测试集所有类别的情况下,利用属性等辅助信息实现对测试样本的正确分类。零样本图像分类在当前图像涉及类别及场景极速增长、新类别层出不穷和分类精细度不断提升等背景下,具有广泛应用前景。本文针对当前零样本图像分类研究中特征-属性关系构建不全面、属性描述不充分等问题,利用弹性网约束、宽度学习和属性关系有向图等方法充分挖掘属性-特征关系、属性-属性关系、属性空间结构和属性-类别关系,提出了四种基于属性挖掘的零样本图像分类方法,主要工作如下:1.针对零样本图像分类中属性和特征表达能力不足的问题,提出基于增强属性-特征的宽度属性预测模型。首先,利用弹性网约束学习二值化的稀疏增强属性,并与手动标定的语义属性共同构成混合属性;其次,通过宽度学习的增强节点获得增强特征,对已有图像特征进行扩展;同时,采用宽度学习中岭回归的伪逆矩阵投影同步得到所有属性的预测结果;最后,通过曼哈顿距离计算预测属性与各测试类属性的相似性,实现零样本图像分类。2.针对属性描述不充分和属性与特征之间映射不全面的问题,提出基于加权重构混合属性组的零样本图像分类模型。首先,利用层次聚类对语义属性进行自动分组,然后通过宽度结构对属性分别按组进行增强,共同构成混合属性;其次,考虑属性组之间的权重关系,通过加权自动编码器实现属性空间和特征空间之间的映射;同时,在目标函数中引入结构化稀疏L21范数,去除属性冗余;最后,在特征空间计算测试样本特征和各预测类别特征的相似性,实现零样本图像分类。3.针对属性和特征之间映射不全面的问题,同时考虑合理挖掘属性空间结构,提出基于关系有向图正则化的属性三因子分解模型。首先,利用属性的矩阵三因子分解实现属性空间和特征空间的映射,将投影矩阵作为训练和测试阶段的共享因子;其次,通过加权属性之间的相似性定义权值矩阵,构建属性关系有向图;最后,在属性空间或特征空间计算测试样本和各测试类别的相似性,实现图像分类。针对投影领域偏移问题,通过同时考虑测试类别关系和测试样本分布进一步提出直推式模型。4.针对属性与特征关系缺乏考虑样本特征分布的问题,同时考虑现有生成模型中类别属性表示过于相似、测试集中生成样本和真实样本分布不一致的缺点,提出基于属性核矩阵的生成特征领域自适应模型。首先,利用核方法在语义空间计算核函数,进而构造属性核矩阵;其次,将语义属性-类别关系矩阵与属性核矩阵合并作为条件,通过条件Wasserstein生成对抗网络得到伪样本特征;然后,采用联合分布自适应方法缩小测试集有标签生成样本与无标签真实样本的边缘分布和条件分布差异;最后,利用测试集生成样本,通过有监督学习实现零样本图像分类。在公共属性数据集上的对比实验结果表明,所提算法均有效提高了不同设置情形下的零样本图像分类精度。本文研究成果不仅可以丰富现有的机器学习理论和方法,而且能够广泛推广应用到诸多相关领域,具有重要理论意义和实用价值。论文共包含图58幅,表12个,参考文献216篇。