20171129-27：黄伟林 Learn CNNs from Large-scale Web Images without human annotat ...

报告嘉宾：黄伟林（University of Oxford）

报告时间：2017年11月29日（星期三）晚19:00（北京时间）

报告题目：Learn CNNs from Large-scale Web Images without human annotation

主持人：欧阳万里（悉尼大学）

报告摘要：

We present a simple yet efficient approach capable of training deep neural networks on large-scale weaklysupervised web images, which are crawled from the Internet by using text queries, without any human annotation. We develop a principled learning strategy by leveraging curriculum learning, with the goal of handling massive amount of noisy labels and data unbalance effectively. We introduce a new curriculum design method that measures data complexity using cluster densities, and rank it in an unsupervised manner, allowing for an efficient implementation of curriculum learning in our large-scale web images. This results in a high-performance CNN model, where the negative impact of noisy labels is reduced substantially. Importantly, we show by experiments that those images with highly noisy labels surprisingly improve the generalization capability of model, by working as a manner of regularization. Our approaches, with an ensemble of multiple models, obtain an accuracy of 94.75% on the 1000-category image classification task, which won the 1st place in the Webvision Challenge, by outperforming the other submissions by a large margin.

报告人简介：

Weilin Huang is Chief Scientist of Malong Technologies. He was working as a postdoc researcher with Prof. Andrew Zisserman in Visual Geometry Group (VGG),  University of Oxford.  He was an Assistant Professor with the Chinese Academy of Science. He received his Ph.D. degree  from The University of Manchester, U.K. His research interests include scene text detection/recognition, large-scale image classification and medical image analysis. He has served as a PC Member or Reviewer for main computer vision conferences, including ICCV, CVPR, ECCV and AAAI. His team was the first runner-up at the ImageNet 2015 on scene recognition, and was the winner of WebVision Challenge in CVPR 2017.

Personal homepage of the speaker：

http://www.whuang.org/

特别鸣谢本次Webinar主要组织者：

VOOC责任委员：欧阳万里（悉尼大学）

VODB协调理事：卢孝强（中国科学院西安光学精密机械研究所）

活动参与方式：

1、VALSE Webinar活动全部网上依托VALSE QQ群的“群视频”功能在线进行，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，并通过文字或语音与讲者交互；

2、为参加活动，需加入VALSE QQ群，目前A、B、C、D、E、F群已满，除讲者等嘉宾外，只能申请加入VALSE G群，群号：669280237。申请加入时需验证姓名、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、为参加活动，请下载安装Windows QQ最新版，群视频不支持非Windows的系统，如Mac，Linux等，手机QQ可以听语音，但不能看视频slides；

4、在活动开始前10分钟左右，主持人会开启群视频，并发送邀请各群群友加入的链接，参加者直接点击进入即可；

5、活动过程中，请勿送花、棒棒糖等道具，也不要说无关话语，以免影响活动正常进行；

6、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；

7、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。

[slides]