基于VAE的伪样本重排练实现的类增量学习  

Class incremental learning based on pseudo-rehearsal with variational autoencoder

作　　者：莫建文[1] 陈瑶嘉 MO Jian-wen;CHEN Yao-jia(Key Laboratory of Cognitive Radio and Information Processing of Ministry of Education,School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China;School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,认知无线电与信息处理省部共建教育部重点实验室,广西桂林541004 [2]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《计算机工程与设计》2021年第8期2284-2290,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61661017、61967005、U1501252);广西自然科学基金项目(2017GXNSFBA198212);广西科技基地和人才专项基金项目(桂科AD19110060);中国博士后科学基金面上基金项目(2016M602923XB);认知无线电教育部重点实验室基金项目(CRKL150103、CRKL190107、CRKL160104);桂林电子科技大学研究生创新基金项目(2019YCXS020)。

摘　　要：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,基于分类器训练样本的类均值特征,使筛选出来的伪样本更具代表性。在MNIST、FASHION数据集上的实验结果表明,所提方法能有效减少灾难性遗忘的影响,提高图像分类精度。Aiming at the catastrophic forgetting problem generated by the neural network model during class incremental lear-ning,a kind of incremental learning method based on pseudo-rehearsal with variational autoencoder was proposed.VAE was used to generate pseudo samples to achieve rehearsal.Pseudo images similar to real images were generated.The large-margin Softmax loss was introduced to train the classifier,which enabled the classifier to learn the features of greater distance between classes and smaller distance within classes.A pseudo-sample selection strategy based on class-average features was used,which was based on the class-average features of the classifier training samples,making the filtered pseudo-samples more representative.Experimental results on the MNIST and FASHION datasets show that the proposed method can effectively reduce the impact of catastrophic forgetting and improve the accuracy of image classification.

关 键 词：类增量学习 灾难性遗忘 伪样本重排练 变分自编码器 伪样本选择 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]