面向移动互联应用数据的特征学习技术

In this project, we aim to investigate the following four issues of data classification and retrieval in mobile applications, from the aspect of feature learning: 1. Aiming at the problem of high feature extraction costs, how to efficiently extract features or feature groups for mobile data objects sequentially, to reduce the extraction costs for each particular mobile data sample? 2. The auxiliary features collected by mobile devices are not directly related to the main target of mobile data processing. Aiming at the problem caused by the weak auxiliary features, how to expand and make full use of the auxiliary weak view features from different modalities? 3. Aiming at the large number of target concepts of mobile data objects, how to learn proper latent feature representations with multi-instance feature learning techniques, to exploit the relationship between the low-level visual representations and the large number of target concepts? 4. Aiming at the quasi real time requirements on mobile data processing, how to utilize the special properties of data and the relationships of data objects in mobile applications, to extract transitional feature representations for efficient classification and retrieval on mobile devices? As an application of the solutions obtained to these four questions, we will also build a prototype system which focuses on media objects classification and retrieval based on mobile devices. During this project, it is expected to publish 8-12 high quality papers on important international journals, conferences and top native journals, apply 2-5 patents, and supervise 5-8 graduate students.

本项目将针对移动互联应用数据对象处理中所面临的四个问题，从特征表示学习的角度进行研究：首先，移动应用数据特征抽取代价大，如何对数据对象进行高效的特征按序抽取，以降低特定样本的特征抽取代价？其次，移动设备采集的辅助特征弱，也即辅助特征包含的信息与移动数据对象处理目标间接相关，如何对多种形式的辅助信息构成的弱视图特征进行扩展并加以利用？第三，移动应用数据目标类别多，如何在多示例表示、学习框架中引入中间特征表示机制，以应对潜在目标概念数量巨大的问题？第四，用户对移动应用数据处理实时要求高，如何利用数据特性以及数据对象之间的关系抽取过渡特征，以进行适于多种移动应用任务的快速移动应用数据分类、检索？本项目将为上述问题提供基于特征表示学习的解决方案并研制原型系统，发表国际期刊、会议和国内一级学报高质量论文8-12 篇，申请专利2-5项，培养5-8名研究生。

本项目针对移动互联应用数据对象处理中所面临的四个问题，从特征表示学习的角度进行研究，也即针对移动应用数据特征抽取代价大，移动设备采集的辅助特征弱，移动应用数据目标类别多，用户对移动应用数据处理实时要求高四个方面分别展开深入研究。项目截至2021.12月，提出一类对移动应用数据进行特征按序抽取的特征学习方法;提出一类能够利用辅助特征信息的弱视图特征学习方法；提出一类能够处理大量目标类别的多示例特征学习的方法；提出一类能够在移动设备和特定服务器上进行准实时特征学习方法；在重要国际刊物、国际会议及国内一级学报上发表32篇高质量论文；获得国家发明专利7项；研制一个基于移动设备的图像视频分类、检索原型系统；培养研究生20余名，其中毕业9名（博士2名，硕士7名）。