Initialization Method of Batch Uniformization Auto Encoder by Principal Component Analysis
Initialization Method of Batch Uniformization Auto Encoder by Principal Component Analysis
复制标题
基于主成分分析的批量均匀化自动编码器初始化方法
DOI：
10.1109/ssci50451.2021.9660064
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Shogo Takaoka; Takuya Kitamura; Aiga Suzuki; Masahiro Murakawa
中科院分区：
文献类型：
--
作者：
Shogo Takaoka; Takuya Kitamura; Aiga Suzuki; Masahiro Murakawa
如果网络的初始权值是由基于随机数的方法确定的，则批量均匀化自编码器（BU-AE）在异常检测问题中的性能往往很差。提出了一种基于异常数据的BU-AE主成分初始化方法（PCI-AD），通过主成分分析获得初始权值，训练PCI-AD以减小正常数据的重构误差，增大异常数据的重构误差。由于这些权重在概念上等同于BU-AE的目标函数，因此具有PCI-AD的BU-AE提供稳定的性能。为了证明PCI-AD的有效性，通过几个计算实验，与何的初始化，这是一种基于随机数的初始化方法，所提出的方法进行了比较。在使用CIFAR-10的实验中，证实了与He初始化相比，使用PCI-AD将AUC值提高了高达17%。
A batch uniformization autoencoder (BU-AE) often performs poorly in anomaly detection problems if the initial weights of the network are determined by a random number-based method. This paper proposes a principal component initialization method with anomalous data (PCI-AD) for BU-AE, where the initial weights are obtained by principal component analysis; the PCI-AD is trained to reduce the reconstruction error of normal data and increase that of the anomalous data. Since these weights are equivalent in concept to the objective function of BU-AE, BU-AE with PCI-AD provides stable performance. To demonstrate the effectiveness of PCI-AD, the proposed method was compared through several computational experiments with He's initialization, which is a random number-based initialization method. In the experiment using CIFAR-10, it was confirmed that the use of PCI-AD improved the value of AUC by up to 17% compared to He's initialization.