基于VAE的伪样本重排练实现的类增量学习

针对神经网络模型进行类增量学习时产生的灾难性遗忘问题,提出一种基于VAE的伪样本重排练实现的类增量学习方法.采用VAE生成伪样本的方法,实现重排练,生成与真实图像相似的伪图像;引入大边界Softmax损失训练分类器,使分类器学习到类间距离更大、类内距离更小的特征;采用一种基于类均值特征的伪样本选择策略,基于分类器训练样...     

基于脑启发的类增量学习

现有的类增量学习方法多是采用存储数据或者扩展网络结构,但受内存资源限制不能有效缓解灾难性遗忘问题。针对这一问题,创新地提出基于脑启发生成式重放方法。首先,通过VAE-ACGAN模拟记忆自组织系统,提高生成伪样本的质量;再引入共享参数模块和私有参数模块,保护已提取的特征;最后,针对生成器中的潜在变量使用高斯混合模型,采样特定重放伪样本。在MNIST、Permuted MNIST和CIFAR-10数据集上的实验结果表明,所提方法的分类准确率分别为92.91%、91.44%和40.58%,显著优于其他类增量学习方法。此外,在MNIST数据集上,反向迁移和正向迁移指标达到了3.32%和0.83%,证明该方法实现任务的稳定性和可塑性之间的权衡,有效地防止了灾难性遗忘。     

基于未知类语义约束自编码的零样本图像分类

为缓解传统零样本图像分类模型中存在的领域偏移问题,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ResNet101网络提取所有已知类和未知类图像的视觉特征;其次,通过编码器将提取的图像深度视觉特征从视觉空间映射到语义空间;然后,通过解码器将映射后得到的语义向量重构为视觉特征向量,在语义自编码器的训练过程中,利用未知类图像的聚类视觉中心和未知类语义类原型的分布对齐施加约束,以缓解领域偏移问题;最后,基于经编码器预测得到的测试图像语义向量和各测试类语义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.在AwA2和CUB数据集上的实验结果表明,所提出模型具有较高的分类准确度.     

结合知识图谱的变分自编码器零样本图像识别

近年来,结合生成模型的零样本算法得到了广泛的研究,但此类方法通常仅使用属性注释,缺少类别语义,而单一信息对类别表征能力不够强,容易产生域偏移,影响知识迁移的效果,进而降低分类结果的准确率。为了解决此问题,提出一种结合知识图谱变分自编码器零样本识别算法（KG-VAE）,通过构建联合类别分级结构,类别文本描述和词向量的层次结构化知识图谱作为语义信息库,将知识图谱中丰富的语义知识结合到以变分自编码器为基础的生成模型中,使生成的潜在特征更好保留有效的判定性信息,减小域偏移,促进知识迁移。在四个公开的零样本数据集上进行了实验,对比基准方法 CADA-VAE,分类平均准确率有一定的提高;同时利用消融实验证明了知识图谱作为语义辅助信息的有效性。     

基于任务相似度的增量学习优化方法

针对增量学习存在的灾难性遗忘和新任务数据逐步积累问题,提出了基于新旧任务之间相似度的样本重放优化学习方法,相似度越高,重放样本越少。并选择MINIST数据集在卷积神经网络上进行了实验研究,验证了该方法的可行性和有效性。     

基于孪生变分自编码器的小样本图像分类方法

当前深度学习大都基于大量数据通过构建深层次的网络实现自动识别,但在很多场景中难以获得大量的样本数据.针对这一问题,提出一种基于孪生变分自编码器(siamese variational auto-encoder,S-VAE)的小样本图像分类方法.通过变分自编码器提取原始训练数据的高层语义特征,然后由两个训练好的变分自编码...     

视觉特征对比解耦的广义零样本学习

广义零样本学习通常利用在ImageNet上预训练的深度模型来提取相应的视觉特征,然而预训练模型提取到的视觉特征不可避免地包含和语义无关的信息,这将导致语义—视觉对齐的偏差以及对不可见类的负迁移,从而影响分类结果。为解决上述问题,提出了视觉特征对比解耦的广义零样本学习模型(visual feature contrast decoupling for generalized zero-shot learning, VFCD-GZSL),通过解耦出视觉特征中的语义相关表示来降低冗余信息对分类结果的影响。具体来说,首先用条件变分自编码器生成不可见类的视觉特征。然后通过解耦模块将视觉特征解耦语义相关和语义无关的潜层表示,同时添加总相关惩罚和对比损失来鼓励两者间的相互独立,并用语义关系匹配模型衡量其语义一致性,从而指导模型学习语义相关表示。最后使用特征细化模块细化后的特征和语义相关表示联合学习一个广义零样本学习分类器。在四个数据集上的实验均取得较优的结果,证实了所提方法的有效性。     

基于融合变分图注意自编码器的深度聚类模型

聚类作为数据挖掘和机器学习中最基本的任务之一,在各种现实世界任务中已得到广泛应用.随着深度学习的发展,深度聚类成为一个研究热点.现有的深度聚类算法主要从节点表征学习或者结构表征学习两个方面入手,较少考虑同时将这两种信息进行融合以完成表征学习.提出一种融合变分图注意自编码器的深度聚类模型FVGTAEDC(Deep Clustering Model Based on Fusion Varitional Graph Attention Self-encoder),此模型通过联合自编码器和变分图注意自编码器进行聚类,模型中自编码器将变分图注意自编码器从网络中学习(低阶和高阶)结构表示进行集成,随后从原始数据中学习特征表示.在两个模块训练的同时,为了适应聚类任务,将自编码器模块融合节点和结构信息的表示特征进行自监督聚类训练.通过综合聚类损失、自编码器重构数据损失、变分图注意自编码器重构邻接矩阵损失、后验概率分布与先验概率分布相对熵损失,该模型可以有效聚合节点的属性和网络的结构,同时优化聚类标签分配和学习适合于聚类的表示特征.综合实验证明,该方法在5个现实数据集上的聚类效果均优于当前先进的深度聚类方法.     

基于双通道变分自编码器的高光谱图像分类

针对现有高光谱图像变分自编码器（variational autoencoder,VAE）分类算法存在空间和光谱特征利用效率低的问题,提出一种基于双通道变分自编码器的高光谱图像深度学习分类算法。通过构建一维条件变分自编码器（conditional variational autoencoder,CVAE）特征提取框架和二维循环通道条件变分自编码（channel-recurrent conditional variational autoencoders,CRCVAE）特征提取框架分别提取高光谱图像的光谱特征和空间特征,将光谱特征向量和空间特征向量叠加形成空谱联合特征向量,将联合特征送入Softmax分类器中进行分类。在Indian pines和Pavia University两种高光谱数据集上进行了分析验证,实验结果显示,与其他算法相比,提出的算法在总分类精度、平均分类精度和Kappa系数等评价指标上至少提高了3.40、2.75和3.57个百分点,结果显示提出的算法得到了最高的分类精度和更好的可视化效果。     

基于重构对比的广义零样本图像分类

广义零样本图像分类中常使用生成模型重构视觉信息或语义信息用于再进一步学习.然而,基于变分自编码器的方法对重构样本利用不够充分,表示性能欠缺.因此,文中提出基于重构对比的广义零样本图像分类模型.首先,使用两个变分自编码器将视觉信息和语义信息编码为同维度的低维隐向量,再将隐向量分别解码到两种模态.然后,使用投影模块投影视觉信息与语义模态的隐向量重构的视觉模态信息.最后,对投影后的特征进行重构对比学习.在保持变分自编码器重构性能的基础上增强编码器重构的判别性能,提高预训练特征在广义零样本图像分类任务上的应用能力.在4个标准数据集上的实验证实文中模型的有效性.     

基于近邻分类的增量学习分类算法研究

为解决分类器学习新样本知识的问题,提出一种基于近邻算法的增量学习算法。该算法以最近邻算法为基础,首先计算新样本与标准样本之间的匹配度,找到最佳匹配样本和次佳匹配样本,然后通过与匹配度阈值进行比较来决定是类内学习还是类别学习。算法采用UCI中的标准数据集进行实验并应用于车辆识别仿真,其结果验证了该算法的有效性。实验进一步研究了匹配度阈值的选择和初始化样本数量选取对分类正确率影响。     

基于增量式学习的数据流实时分类模型

传统数据挖掘方法,主要针对静态数据进行挖掘,而对数据流挖掘往往失效。为了解决数据流的数据挖掘问题,提出一种通过改变传统支持向量机增量式学习方法,利用轮转式结构将多分类器按照数据流时间顺序进行组合,并且通过对分类器的优化,可以提高模型对数据流分类的准确率并减少训练时间消耗。实验结果表明,该模型在保证学习精度和推广能力的同时,提高了训练速度,适合于数据流在线分类和在线学的问题。     

基于图像目标特征空间自学习分类算法

为解决图像分类过程中特征点选择的随机性对分类精度造成的影响,提出一种基于图像目标特征空间自学习分类算法。利用基于颜色和纹理特征的多通道局部主动轮廊模型找到图像的目标区域,在目标区域选取特征并对特征稀疏编码建立图像的目标特征空间。为进一步提高图像分类精度建立投票机制下基于图像目标特征空间的自学习算法。实验结果表明,该方法能避免特征选择的随机性对实验结果的影响,有效地提高图像分类的精度。     

Unsupervised feature selection using clustering ensembles and population based incremental learning algorithm

This paper describes a novel feature selection algorithm for unsupervised clustering, that combines the clustering ensembles method and the population based incremental learning algorithm. The main idea of the proposed unsupervised feature selection algorithm is to search for a subset of all features such that the clustering algorithm trained on this feature subset can achieve the most similar clustering solution to the one obtained by an ensemble learning algorithm. In particular, a clustering solution is firstly achieved by a clustering ensembles method, then the population based incremental learning algorithm is adopted to find the feature subset that best fits the obtained clustering solution. One advantage of the proposed unsupervised feature selection algorithm is that it is dimensionality-unbiased. In addition, the proposed unsupervised feature selection algorithm leverages the consensus across multiple clustering solutions. Experimental results on several real data sets demonstrate that the proposed unsupervised feature selection algorithm is often able to obtain a better feature subset when compared with other existing unsupervised feature selection algorithms.     

An incremental feature selection approach based on scatter matrices for classification of cancer microarray data

Microarray data are often characterized by high dimension and small sample size. There is a need to reduce its dimension for better classification performance and computational efficiency of the learning model. The minimum redundancy and maximum relevance (mRMR), which is widely explored to reduce the dimension of the data, requires discretization and setting of external parameters. We propose an incremental formulation of the trace of ratio of the scatter matrices to determine a relevant set of genes which does not involve discretization and external parameter setting. It is analytically shown that the proposed incremental formulation is computationally efficient in comparison to its batch formulation. Extensive experiments on 14 well-known available microarray cancer datasets demonstrate that the performance of the proposed method is better in comparison to the well-known mRMR method. Statistical tests also show that the proposed method is significantly better when compared to the mRMR method.     

Algorithm learning based neural network integrating feature selection and classification

Feature selection and classification techniques have been studied independently without considering the interaction between both procedures, which leads to a degraded performance. In this paper, we present a new neural network approach, which is called an algorithm learning based neural network (ALBNN), to improve classification accuracy by integrating feature selection and classification procedures. In general, a knowledge-based artificial neural network operates on prior knowledge from domain experience, which provides it with better starting points for the target function and leads to better classification accuracy. However, prior knowledge is usually difficult to identify. Instead of using unknown background resources, the proposed method utilizes prior knowledge that is mathematically calculated from the properties of other learning algorithms such as PCA, LARS, C4.5, and SVM. We employ the extreme learning machine in this study to help obtain better initial points faster and avoid irrelevant time-consuming work, such as determining architecture and manual tuning. ALBNN correctly approximates a target hypothesis by both considering the interaction between two procedures and minimizing individual procedure errors. The approach produces new relevant features and improves the classification accuracy. Experimental results exhibit improved performance in various classification problems. ALBNN can be applied to various fields requiring high classification accuracy.     

Neighborhood based sample and feature selection for SVM classification learning

Support vector machines (SVMs) are a class of popular classification algorithms for their high generalization ability. However, it is time-consuming to train SVMs with a large set of learning samples. Improving learning efficiency is one of most important research tasks on SVMs. It is known that although there are many candidate training samples in some learning tasks, only the samples near decision boundary which are called support vectors have impact on the optimal classification hyper-planes. Finding these samples and training SVMs with them will greatly decrease training time and space complexity. Based on the observation, we introduce neighborhood based rough set model to search boundary samples. Using the model, we firstly divide sample spaces into three subsets: positive region, boundary and noise. Furthermore, we partition the input features into four subsets: strongly relevant features, weakly relevant and indispensable features, weakly relevant and superfluous features, and irrelevant features. Then we train SVMs only with the boundary samples in the relevant and indispensable feature subspaces, thus feature and sample selection is simultaneously conducted with the proposed model. A set of experimental results show the model can select very few features and samples for training; in the mean time the classification performances are preserved or even improved.     

优化强化学习路径特征分类的脉象识别法

脉象识别是中医诊断的重要手段之一。长期以来,依据个人经验进行的脉诊制约了中医的推广与发展。因此,利用传感设备进行脉象识别的研究正在逐步展开。针对神经网络识别脉象的相关研究中,存在需要大量训练数据集,以及存在处理“黑箱”和时间花销较大等问题,在强化学习的框架下,提出了一种采用马尔可夫决策和蒙特卡罗搜索的脉象图分析法。首先依据中医理论对特定的脉象进行路径分类,然后在此基础上为不同的路径选择代表性特征,最终通过对代表性特征的阈值对比完成对脉象的识别。实验结果表明,所提方法可缩减训练时间和所需资源,并可保留完整的经验轨迹;且在提高脉象识别的准确率的同时,还可解决数据处理过程中的“黑箱”问题。