Biograph: Dr. Guiguang Ding received his Ph.D degree in electronic engineering from Xidian University, China, in 2004. He joined the faculty of School of Software at Tsinghua University (THSS) in July 2006. He is currently a tenured associated professor and vice dean of the school of software, at Tsinghua University. His research is dedicated to tackling significant problems in the field of computer vision and multimedia retrieval. With his students and collaborators, he has published about 100 papers in international journal (e.g., IEEE TIP, IEEE TMM) and the international top conferences (e.g., CVPR, ICCV, ICML). There are five Essential Science Indicators (ESI) highly cited papers. These papers have about 4000 citations (Google Scholar). Except for the theoretical research, he was also dedicated to the transformation and engineering application of new technologies. he has applied for 40 patents for invention, among which 30 were authorized. Based on the feature learning and indexing technologies, he developed the video content analysis systems, which were deployed to the monitoring and management of the Internet video content and Skynet Project of China.

Title: Filter Pruning Methods for Deep Neural Network Compression for Computer Vision Applications

Abstract: Deep Neural Network (DNN) is powerful but computationally expensive and memory intensive, thus impeding its practical usage on resource-constrained frontend devices. Therefore, many compression and acceleration technologies of deep learning models are proposed in the industry and academia domains. Filter pruning is an approach for deep model compression, which aims at eliminating some parameters with tolerable performance degradation. The lecture will summarize the progress of the Filter Pruning technology in the past three years, and introduce a new approximated Oracle Filter Pruning framework, which features high quality of importance estimation, reasonable time complexity and no need for heuristic knowledge. We empirically found out that the structural change in CNNs can be analyzed with local information only, which may inspire further theoretical researches.