摘要
为解决图像分类任务中模型结构固化、产生巨大内存消耗、时间消耗的问题,提出一种增量式深度神经网络(IDNN)。输入样本通过聚类算法激活不同簇并被分别处理:如果新样本激活已有簇,则更新该簇参数;否则为新簇开辟分支,并训练独立特征集。在Caltech-101、ORL Face、ETH-80数据库的验证结果表明,该系统能自动调整网络结构,适用于轮廓、纹理、视角等不同环境的增量式学习,例如在Caltech-101库分类任务中准确率超出VGGNet 5.08%、Alex Net 3.44%。
In order to solve the problem of model structure fossilization,generating huge memory consumption and time consumption in image classification task,an incremental depth neural network(IDNN)is proposed.IDNN divides input samples into branches and trains them separately.If the input sample activates one of the existing clusters,IDNN updates parameters of the corresponding cluster;otherwise,IDNN creates new branch and learns an independent feature set.Experiments reported on Caltech-101,ORL Face and ETH-80 dataset show that IDNN can automatically adjust the network structure and is suitable for incremental learning in different environments such as contour variation,texture variation and viewpoint diversity.Also,we prove experimentally that its classification performance outperforms the VGGNet by 5.08%and AlexNet by 3.44%on Caltech-101 dataset.
作者
胡晶晶
王冉
詹伶俐
王元庆
HU Jingjing;WANG Ran;ZHAN Lingli;WANG Yuanqing(School of Electronic Science and Engineering,Nanjing University,Nanjing 210046,China;Tandon School of Engineering,New York University,New York 11201,USA)
出处
《电子器件》
CAS
北大核心
2018年第1期72-78,共7页
Chinese Journal of Electron Devices
基金
国家科技重大专项项目(AHJ2011Z001)
