基于公共空间嵌入的端到端深度零样本学习  

End-to-end Deep Zero-shot Learning Based on Co-space Embedding

作　　者：秦牧轩 荆晓远[1] 吴飞[1] QIN Mu-xuan;JING Xiao-yuan;WU Fei(School of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学自动化学院,江苏南京210003

出　　处：《计算机技术与发展》2018年第11期44-47,共4页Computer Technology and Development

基　　金：国家自然科学基金(61702280);江苏省自然科学基金(BK20170900);南京邮电大学引进人才科研启动基金(NY217009)

摘　　要：零样本学习(ZSL)是一种特殊的机器学习问题,目的是在测试阶段识别出训练集中未曾出现的类别样本并进行分类。目前主流技术手段有两种:一种是基于属性学习,一种是基于词嵌入模型。两种方法各有优缺点。属性学习可以看作人工特征标记,分类效果往往取决于人工设定的属性好坏。而利用文本特征(词向量)的词嵌入模型无需人为参与也可得到不错的特征以替代属性。文中提出了一种基于公共空间嵌入的深度零样本学习方法。通过图像和文本分别建立深度神经网络并连接两个网络,在顶层学习一个联合嵌入的公共空间。基于深度学习端到端模型的设计可以同时利用属性特征和文本特征实现图像的零样本学习。实验结果表明,该方法达到了较好的识别效果。Zero-shot learning(ZSL)is a special machine learning problem and aims to classify the samples without appearing in the training set at test stage.At present,there are two main technical methods:one is based on attribute learning and the other is based word embedding.They have their own advantages and disadvantages.Attribute learning can be regarded as artificial feature and its classification effect often depends on the quality of artificial mark.Word embedding model using text feature(word2vector)can get great feature to replace attribute without human participation.In this paper,we propose a deep zero-shot learning based on co-space embedding.The deep neural networks are established respectively by image and text and connected to learn a jointly embedding co-space at the top.The design of the end-to-end model based on deep learning can use both attribute and text feature in image ZSL.Experiment shows that this method achieves better recognition results.

关 键 词：零样本学习 嵌入模型 属性学习 深度神经网络 公共空间 词向量 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]