The superior performance of Convolutional Neural Networks (CNNs) has been demonstrated in many applications such as image classification, detection and processing. Yet, CNN’s working principle remains a mystery. In this talk, I will shed light on the underlying operational principle of CNNs using an interpretable initialization. The roles of convolutional operations, the bias term, max pooling, convolutional layers and fully connected layers can be well explained through this initialization. Furthermore, it gives a very good result without backpropagation, and the backpropagation can be used to finetune the initial result. To go further, I will present a new methodology based on the second-order statistics (i.e. the covariance matrix) of input image pixels. It has a good potential for computer vision and image processing tasks without being configured into the network form. Some preliminary results will be presented. Finally, the CNN approach and the new second-order statistics approaches will be compared.
Dr. C.-C. Jay Kuo received his Ph.D. degree from the Massachusetts Institute of Technology in 1987. He is now with the University of Southern California (USC) as Director of the Media Communications Laboratory and Dean’s Professor in Electrical Engineering-Systems. His research interests are in the areas of media processing, compression and understanding. Dr. Kuo was the Editor-in-Chief for the IEEE Trans. on Information Forensics and Security in 2012-2014. He was the Editor-in-Chief for the Journal of Visual Communication and Image Representation in 1997-2011, and served as Editor for 10 other international journals. Dr. Kuo received the 1992 National Science Foundation Young Investigator (NYI) Award, the 1993 National Science Foundation Presidential Faculty Fellow (PFF) Award, the 2010 Electronic Imaging Scientist of the Year Award, the 2010-11 Fulbright-Nokia Distinguished Chair in Information and Communications Technologies, the 2011 Pan Wen-Yuan Outstanding Research Award, the 2014 USC Northrop Grumman Excellence in Teaching Award, the 2016 USC Associates Award for Excellence in Teaching, the 2016 IEEE Computer Society Taylor L. Booth Education Award, the 2016 IEEE Circuits and Systems Society John Choma Education Award, the 2016 IS&T Raymond C. Bowman Award, and the 2017 IEEE Leon K. Kirchmayer Graduate Teaching Award. Dr. Kuo is a Fellow of AAAS, IEEE and SPIE. He has guided 140 students to their Ph.D. degrees and supervised 27 postdoctoral research fellows. Dr. Kuo is a co-author of 260 journal papers, 900 conference papers and 14 books.
Marc Pollefeys is Director of Science leading a team of scientist and engineers to develop advanced perception capabilities for HoloLens. He is also a Professor of Computer Science at ETH Zurich and was elected Fellow of the IEEE in 2012.
He is best known for his work in 3D computer vision, having been the first to develop a software pipeline to automatically turn photographs into 3D models, but also works on robotics, graphics and machine learning problems. Other noteworthy projects he worked on with collaborators at UNC Chapel Hill and ETH Zurich are real-time 3D scanning with mobile devices, a real-time pipeline for 3D reconstruction of cities from vehicle mounted-cameras, camera-based self-driving cars and the first fully autonomous vision-based drone. Most recently his academic research has focused on combining 3D reconstruction with semantic scene understanding. He has published over 300 peer-reviewed publications and holds several patents. Marc was program chair for CVPR 2009, general chair for ECCV 2014 and is general chair for ICCV 2019. He is also heavily involved in the 3D Vision conference series. His lab at ETH Zurich also developed the PixHawk auto-pilot which can be found in over half a million drones and he has co-founded several computer vision start-ups.
Visual sensing and processing have experienced a significant growth in the past several decades. Complementing the technology R&D dealing with visible signals and patterns are those exploiting a variety of nearly invisible “micro-signals,” which are typically an order of magnitude lower in strength or scale than the dominant ones. These micro-signals are ubiquitous but traditionally removed or ignored as nuances. Increasingly, it has been found that many of these micro-signals help us connect the physical world with the digital and cyber space, and harnessing them may bring about beneficial information that would otherwise harder to obtain. This talk will provide an overview on visualizing and analyzing several representative types of micro-signals, discussing the recent research advances and novel applications ranging from security to digital humanity to fitness and health.
Min Wu is a Professor of Electrical and Computer Engineering and a Distinguished Scholar-Teacher at the University of Maryland, College Park. She received her Ph.D. degree in electrical engineering from Princeton University in 2001. At UMD, she leads the Media and Security Team (MAST), with main research interests on information security and forensics and multimedia signal processing. Her research and education have been recognized by a U.S. NSF CAREER award, a TR100 Young Innovator Award from the MIT Technology Review, an U.S. ONR Young Investigator Award, a Computer World "40 Under 40" IT Innovator Award, University of Maryland Invention of the Year Awards, IEEE Distinguished Lecturer recognition, and several paper awards from IEEE SPS, ACM, and EURASIP. She was elected IEEE Fellow and AAAS Fellow for contributions to multimedia security and forensics. Dr. Wu chaired the IEEE Technical Committee on Information Forensics and Security (2012-2013), and has served as Vice President - Finance of the IEEE Signal Processing Society (2010-2012), Founding Chief Editor of the IEEE SigPort initiative (2013-2014), and Editor-in-Chief of the IEEE Signal Processing Magazine (2015-2017).