深圳大学西丽校区

学术讲座

【医学部】学术讲座——From Model Optimization to Interpretable and Coll

2018年12月10日

主讲嘉宾:樊鑫  教授  大连理工大学

时间:20181213日周四上午10:00-11:30

地点:医学院A6-811会议室

主持人:雷柏英

内容简介:Model optimization plays the key role in many learning and vision  tasks. However, designing numerical schemes always need high mathematical skills  and rich domain knowledge. Moreover, it is always challenging to apply the  generally designed iterations in specific real-world scenarios. In this talk, we  introduce a series of paradigms to design task-specific optimization schemes  based on inexact learnable architectures. The theoretical properties of these  deeply trained propagations are carefully investigated. We demonstrate that we  actually provide a new way to establish interpretable and collaborative deep  learning models for different real-world applications. Comparisons to  adversarial mechanisms in GAN will also be covered. Finally, we demonstrate how  to apply the proposed framework to address MRI reconstruction. By optimizing a  compressed sensing energy minimization formulation, we design a coupled deep  model to simultaneously reconstruct MRI images and remove Rician noise,  resulting in a robust medical imaging  reconstruction.

专家简历:Dr. Xin Fan is currently Professor of Dalian University of  Technology, and Dean of School of DUT-RU Information Science and Engineering. He  received the B.E. and Ph.D. degrees in information and communication engineering  from Xian Jiaotong University, Xian, China, in 1998 and 2004, respectively. He spent three years for  postdoc training in US. He joined the School of Software, Dalian University of  University, in 2009. His current research interests include computational  geometry and machine learning, and their applications to image processing and  analysis. He has published over 100 papers in top journals and conferences  including IEEE TIP, TMM, NIPS, ICCV, ECCV, ACM MM, AAAI and IJCAI. He has been  the PI of four grants from NSF of China including one key project. He won the  best student paper award at ICME2015 as the corresponding author and was among  the shortlist of the best paper awards at ICIP2013, ICIP2015 and  ICME2017.