Deep Contextual Learning and Regional Gating Neural Networks

报告题目：Deep Contextual Learning and Regional Gating Neural Networks

报告人：李建国 博士，Intel中国研究院高级主任研究员、研发总监

报告时间：2016年9月22日上午10:00-11:00

报告地点：中北校区理科大楼A510报告厅

报告内容：

We study the problem on how to utilize semantic contextual information with deep neural networks for better image classification. The motivation is two folds.  First, global image features (including CNN based features) ignore the underlying context information among different semantic objects in an image. Consequently, some people attempted to use information from objectness regions. However, current objectness region proposal algorithms usually produce several thousands of region candidates, which include many irrelevant or even noisy regions. This leads to the second problem: how to select useful contextual regions for image classification. We propose two solutions from two different perspective. First, we propose to aggregating region-level attributes from deep neural networks as a contextual representation for further classification purpose. We name this method as deep attributes. Second, we propose regional-gating neural networks, which is an end-to-end deep learning framework that can automatically select contextual region features with specially designed gate units during training procedure. We show that the proposed two methods can bring significant classification accuracy improvement on several benchmarks.

报告人简介：

李建国是英特尔中国研究院高级主任研究员和研发总监。他2001年本科毕业于华中科技大学，2006年在清华大学自动化 系获得博士学位。研究兴趣包括计算机视觉，机器学习(特别是深度学习),以及它们的实际应用。他获得40余项美国专利授权，发表了30多篇学术文章，其中 包括10余篇顶级会议如CVPR, ICCV, ICML, IJCAI, ACM MM的文章。他的诸多研究成果转换为英特尔公司产品特征。特别的，他贡献了来自中国区的第一项CPU/GPU硬件特征. 他获得多项奖励，包括英特尔研究 院最高奖，戈登摩尔奖。他是IEEE /ACM/CCF会员，CCF计算机视觉专委委员。