Some Understandings and New Designs of Recurrent and Convolutional Networks

In this talk, I will talk about some of our recent work in understanding and reforming the well-known recurrent and convolutional networks. In the first part, I will talk about our experience utilizing LSTM in multi-target tracking and show some intuitions about why the current LSTM may be insufficient for long-term multi-object tracking. A novel bilinear LSTM model suitable for multi-target tracking problems will be proposed, motivated by the classic recursive least squares formulation. Results on the MOT 2016 and MOT 2017 challenges will be shown that significantly outperform traditional LSTMs in terms of identity switches. In the second part, I will talk about some of our recent work in understanding and redesigning CNN, including the explanation neural network (XNN) that projects a deep network into several human-understandable visual concepts, and the PointConv approach we recently developed for fully implementing CNN on irregularly spaced point cloud data. State-of-the-art results on the CIFAR-10 and ScanNet benchmarks will be shown that significantly outperform other recent attempts building deep networks on point clouds.
Fuxin Li is currently an assistant professor in the School of Electrical Engineering and Computer Science at Oregon State University. Before that, he has held research positions in University of Bonn and Georgia Institute of Technology. He had obtained a Ph.D. degree in the Institute of Automation, Chinese Academy of Sciences in 2009. He has won an NSF CAREER award, (co-)won the PASCAL VOC semantic segmentation challenges from 2009-2012, and led a team to the 4th place finish in the DAVIS Video Segmentation challenge 2017. He has published more than 40 papers in computer vision, machine learning and natural language processing. His main research interests are deep learning, video object segmentation, multi-target tracking, point cloud deep networks, adversarial deep learning and human understanding of deep learning.