20180926-30 张磊：迁移学习和领域自适应方法及应用

报告嘉宾：张磊（重庆大学）

报告时间：2018年09月26日（星期三）晚上20:00（北京时间）

报告题目：迁移学习和领域自适应方法及应用（Transfer learning/Domain adaptation: Methods and Applications）

主持人：任传贤（中山大学）

报告人简介：

张磊，现任重庆大学微电子与通信工程学院研究员，博士生导师，重庆市生物感知与智能信息处理重点实验室副主任，LiVE智能学习与视觉团队负责人（www.leizhang.tk），CCF计算机视觉专委会委员，CAAI智能服务专委会委员。2013年博士毕业于重庆大学，2013年6月至9月在清华大学深圳研究院、哈尔滨工业大学深圳研究院访问交流，2013年10月至2015年9月为香港理工大学计算机系博士后研究员，2017年12月至2018年5月为澳门大学计算机与信息科学系访问学者。主要聚焦于迁移学习、深度学习、跨领域视觉大数据建模与理解和机器嗅觉与味觉方面的智能理论与应用研究等，部分仿生技术和算法已产品化。主持国家自然科学基金（青年和面上）、重庆市人工智能重大专项、重庆市留学生创新创业支持计划等研究项目14项；目前已在IEEE TNNLS、TIP、TMM、TCYB、TIM、TSMCA、TCBB以及ICCV、ACM MM、IJCNN等会议上发表论文70余篇，中科院一区论文12篇，3篇论文入选ESI高被引论文，2篇入选ESI热点论文，撰写英文专著1部，发明专利15项。曾担任WRJB期刊主编和JECE期刊首席客座编辑；担任20余个国际学术会议IEEE TENCON/ICONIP/SSCI/ICCT等会议的最佳论文奖评审主席、荣誉主席、分会主席、地区主席和特邀讲者；担任IEEE TIP/ TNNLS/ TIE/ TCYB/ TMM/ TIM/ TCSVT/ TSMCA/ TCAS/ PR等50余个国际期刊审稿人。曾获CCBR最佳论文奖、香江学者奖、重庆市优秀博士学位论文奖、杰出审稿人奖、重庆市自然科学优秀学术论文奖、重庆大学黄尚廉院士青年创新奖等，并于2013年入选香江学者计划，于2015年入选重庆大学“百人计划”。

Dr Lei Zhang is currently a distinguished research fellow/professor at the Collage of Microelectronics and Communication Engineering of Chongqing University, Ph.D supervisor, deputy director of the Chongqing Key Laboratory of Bio-perception & Intelligent Information Processing, LiVE (Learning Intelligence & Visual Essential) Group Leader (www.leizhang.tk), member of CCF Computer Vision Committee and CAAI Intelligent Service Committee. In 2013, he graduated from Chongqing University. From June to September 2013, he visited the Tsinghua University (Shenzhen) and the Harbin Institute of Technology (Shenzhen). From October 2013 to September 2015, he was a postdoctoral researcher at the Department of Computing, Hong Kong Polytechnic University. From December 2017 to May 2018, he was a visiting professor in the Department of Computer and Information Science at the University of Macau. His research interests focus on transfer learning, deep learning, cross-domain visual big data modeling and understanding, and intelligent theory and application research in machine olfaction and taste. Some biomimetic technologies and algorithms have been commercialized. He has hosted over 14 research projects including the National Natural Science Foundation of China, Chongqing Artificial Intelligence Project, etc.; He has currently published more than 70 papers in IEEE Transactions including IEEE TNNLS, TIP, TMM, TCYB, TIM, TSMCA, TCBB, etc. and international conferences such as ICCV, ACM MM, IJCNN etc. 3 papers were selected as ESI highly-cited papers, 2 papers were selected as ESI hot papers, 1 English monograph and 14 invention patents. He has served as editor-in-chief of WRJB journals and leading guest editor of JECE journals; served as chairs, honorary chair, session chair, regional chair and keynote speaker for more than 20 international academic conferences such as IEEE TENCON/ICONIP/SSCI/ICCT, etc. He has served as reviewers for more than 50 international journals such as IEEE TIP, TNNLS, TIE, TCYB, TMM, TIM, TCSVT, TSMCA, TCAS, PR, etc. Prof Zhang was awarded the CCBR Best Paper Award, Hong Kong Scholar Award, Chongqing Excellent Doctoral Dissertation Award, Outstanding Reviewer Award, Chongqing Natural Science Excellent Academic Paper Award, Huang Shanglian Academician Youth Innovation Award, etc. Prof Zhang was selected as “Hong Kong Scholar Program” in 2013 and the “Hundred Talents Program” of Chongqing University in 2015.

个人主页：

http://www.leizhang.tk

报告摘要：

统计机器学习在人工智能研究和应用中起着至关重要的作用。然而，建立有效的机器学习模型需要大量标记的训练数据，而特定目标领域的数据标记过程需要耗费大量的人力物力，且十分耗时。因此，通过利用相关的且已标记的源域数据辅助模型的训练和学习，用于标记样本稀缺或完全无标记的弱标记目标域数据的精确识别已成为主流。随着大数据的爆炸式增长，数据的异质性日益突出。随之而来的问题是多域数据的独立同分布（i.i.d.）假设不成立，从而导致传统的机器学习性能下降或失效。近年来，用于解决标记数据稀缺和域间分布差异化问题的迁移学习（TL）和领域自适应（DA）作为典型的弱监督学习范式受到了极大的关注，并在很多智能化应用中包括图像识别、人脸识别、自然语言处理、遥感数据分析、医学图像分析、虚拟现实等领域取得了巨大的成功。本报告中，将介绍迁移学习/领域自适应的概念基础和过去10年来最新研究进展。此外，将介绍我们近年来在TL / DA方面的研究工作，包括分类器自适应模型（EDA，TIP16），子空间重建迁移学习（LSDT，TIP16; CRTL，ICCV17），公共子空间迁移学习（CDSL，TIM17），流形准则迁移模型（MCTL，TNNLS18）和自我对抗迁移网络（AdvNet，ACM MM17）。最后，将简要介绍我们提出的一种新的学习范式，即引导式学习，并用于异构视觉识别。

Statistical machine learning plays a critical role in artificial intelligence research and applications. However, establishing an effective machine learning model needs a number of labeled training data, whereas the labeling process in a specific target domain is labor cost expensive and time consuming. Therefore, training a learning model on the target domain where labeled patterns are scarce or unavailable by exploiting other labeled, related, auxiliary source domains becomes a mainstream. With the explosive growth of big data, the heterogeneity of data becomes increasingly prominent. The problem that comes with it is that the independent identical distribution (i.i.d.) assumption of multi-domain data does not hold, as a result, conventional machine learning cannot work well without i.i.d. condition. Recently, transfer learning (TL) and domain adaptation (DA), as representative weakly-supervised learning paradigm, have been proposed and received great attention, which are prevailingly needed and have demonstrated a great success in various applications such as image classification, face recognition, remote sensing, medical image, AR/VR etc. In this talk, I will present the concept basis and the recent progress of transfer learning/domain adaptation during the past 10 years. Further, I will introduce our recent works in TL/DA methodologies including the classifier adaptation models (EDA, TIP16), the subspace reconstruction transfer models (LSDT, TIP16; CRTL, ICCV17), the common subspace transfer model (CDSL, TIM17), the manifold criterion transfer models (MCTL, TNNLS18), the self-adversarial transfer (AdvNet, ACM MM17). Finally, I may briefly introduce our proposed new learning paradigm, Guide Learning (TNNLS18 in review), for heterogeneous visual recognition. 

参考文献：

[1]  Lei Zhang, Wangmeng Zuo, and David Zhang, "LSDT: Latent Sparse Domain Transfer Learning for Visual Adaptation," IEEE Transactions on Image Processing, vol. 25, no. 3, pp. 1177-1191, 2016.

[2] Lei Zhang and David Zhang, "Robust Visual Knowledge Transfer via Extreme Learning Machine based Domain Adaptation," IEEE Transactions on Image Processing, vol. 25, no. 10, pp. 4959-4973, 2016.

[3] Lei Zhang and David Zhang, "Visual Understanding via Multi-Feature Shared Learning with Global Consistency," IEEE Transactions on Multimedia, vol. 18, no. 2, pp. 247-259, 2016.

[4] Lei Zhang, Yan Liu and Pingling Deng, "Odor Recognition in Multiple E-nose Systems with Cross-domain Discriminative Subspace Learning," IEEE Transactions on Instrumentation and Measurement, vol. 66, no. 7, pp. 1679-1692, July 2017.

[5] Lei Zhang, Jian Yang, and David Zhang, "Domain Class Consistency based Transfer Learning for Image Classification Across Domains," Information Sciences, vol. 418-419, pp. 242-257, 2017.

[6] Shanshan Wang, Lei Zhang, and Wangmeng Zuo, "Class-specific Reconstruction Transfer Learning via Sparse Low-rank Constraint," IEEE Int'Conf. Computer Vision (ICCVW), pp. 949-957, Oct 2017.

[7] Qingyan Duan and Lei Zhang, "AdvNet: Adversarial Contrastive Residual Net for 1 Million Kinship Recognition," ACM Conf. Multimedia (ACM MM), pp. 21-29, Oct 2017.

[8] Jingru Fu, Lei Zhang, Bob Zhang, and Wei Jia, "Guided Learning: A New Paradigm For Multi-task Classification," CCBR, pp. 239-246, 2018 [oral].

18-30期VALSE在线学术报告参与方式：

长按或扫描下方二维码，关注“VALSE”微信公众号（valse_wechat），后台回复“30期”，获取直播地址。

特别鸣谢本次Webinar主要组织者：

VOOC责任委员：任传贤（中山大学）

活动参与方式：

1、VALSE Webinar活动依托在线直播平台进行，活动时讲者会上传PPT或共享屏幕，听众可以看到Slides，听到讲者的语音，并通过聊天功能与讲者交互；

2、为参加活动，请关注VALSE微信公众号：valse_wechat 或加入VALSE QQ群（目前A、B、C、D、E、F、G群已满，除讲者等嘉宾外，只能申请加入VALSE H群，群号：701662399）；

*注：申请加入VALSE QQ群时需验证姓名、单位和身份，缺一不可。入群后，请实名，姓名身份单位。身份：学校及科研单位人员T；企业研发I；博士D；硕士M。

3、在活动开始前5分钟左右，讲者会开启直播，听众点击直播链接即可参加活动，支持安装Windows系统的电脑、MAC电脑、手机等设备；

4、活动过程中，请不要说无关话语，以免影响活动正常进行；

5、活动过程中，如出现听不到或看不到视频等问题，建议退出再重新进入，一般都能解决问题；

6、建议务必在速度较快的网络上参加活动，优先采用有线网络连接

7、VALSE微信公众号会在每周一推送上一周Webinar报告的总结及视频（经讲者允许后），每周四发布下一周Webinar报告的通知及直播链接。

[slides]