基于混合高斯先验变分自编码器的深度多球支持向量数据描述  

Deep Multiple-sphere Support Vector Data Description Based on Variational Autoencoder with Mixture-of-Gaussians Prior

作　　者：武慧囡 邢红杰[1] 李刚[2,3] WU Huinan;XING Hongjie;LI Gang(Hebei Key Laboratory of Machine Learning and Computational Intelligence,College of Mathematics and Information Science,Hebei University,Baoding,Hebei 071002,China;Department of Computer,North China Electric Power University,Baoding,Hebei 071003,China;Engineering Research Center of Intelligent Computing for Complex Energy Systems,Baoding,Hebei 071003,China)

机构地区：[1]河北大学数学与信息科学学院河北省机器学习与计算智能重点实验室,河北保定071002 [2]华北电力大学计算机系,河北保定071003 [3]复杂能源系统智能计算教育部工程研究中心,河北保定071003

出　　处：《计算机科学》2024年第6期135-143,共9页Computer Science

基　　金：国家自然科学基金(61672205);河北省自然科学基金(F2017201020);河北大学高层次人才科研启动项目(521100222002);复杂能源系统智能计算教育部工程研究中心开放基金(ESIC202101)。

摘　　要：随着数据维度和规模的不断增加,基于深度学习的异常检测方法取得了优异的检测性能,其中深度支持向量数据描述(Deep SVDD)得到了广泛应用。然而,要缓解超球崩溃问题,就需要对Deep SVDD中映射网络的各种参数施加约束。为了进一步提高Deep SVDD中映射网络的特征学习能力,同时解决超球崩溃问题,提出了基于混合高斯先验变分自编码器的深度多球支持向量数据描述(Deep Multiple-Sphere Support Vector Data Description Based on Variational Autoencoder with Mixture-of-Gaussians Prior,DMSVDD-VAE-MoG)。首先,通过预训练初始化网络参数和多个超球中心;其次,利用映射网络获得训练数据的潜在特征,对VAE损失、多个超球的平均半径和潜在特征到所对应超球中心的平均距离进行联合优化,以获得最优网络连接权重和多个最小超球。实验结果表明,所提DMSVDD-VAE-MoG在MNIST,Fashion-MNIST和CIFAR-10上均取得了优于其他8种相关方法的检测性能。With the continuous increase of data dimension and scale,anomaly detection methods based on deep learning have achieved excellent detection performance,among which deep support vector data description(Deep SVDD)has been widely used.However,it is necessary to impose constraints on various parameters of the mapping network in Deep SVDD to alleviate the hypersphere collapse problem.In order to further improve the feature learning ability of the mapping network in Deep SVDD and solve the hypersphere collapse problem,deep multiple-sphere support vector data description based on variational autoencoder with mixture-of-gaussians prior(DMSVDD-VAE-MoG)is proposed.First,the network parameters and multiple hypersphere centers are initialized by pre-training.Second,the latent features of the training data are obtained by mapping network.The VAE loss,the average radius of multiple hyperspheres together with the average distance between the latent features and their corres-ponding hypersphere centers are jointly optimized to obtain the optimal network connection weights and multiple minimum hyperspheres.In comparison with the other eight related methods,the experimental results show that the proposed DMSVDD-VAE-MoG achieves better detection performance upon MNIST,Fashion-MNIST and CIFAR-10.

关 键 词：深度支持向量数据描述 混合高斯先验 变分自编码器 异常检测 超球崩溃 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]