From Feedforward-Designed Convolutional NeuralNetworks (FF CNNs) to Successive Subspace Learning (SSL)
C.-C. Jay Kuo
Prof.
University of Southern California
IEEE/AAAS/SPIE Fellow
IEEE Computer Society Edward J. McCluskey Technical Achievement Award
Abstract: Given a convolutional neural network (CNN) architecture, its network parameters are determined by backpropagation (BP) nowadays. Theunderying mechanism remains to be a black-box alter a large amount of theoretical investigation. In this talk, I describe a new interpretablefeedforward (FF) design with the LeNel-5 as an example. The FF-designed CNN is a data-centric approach that derives network parametersbased on training data stalistics layer by layer in one pass. To build the convolutional layers, we develop a new signal transform, called theSaab (Subspace approximation with adiusted bias) transform. The blias in fiter weights is chosen to anihilate nonlinearity of the activationfunction. To build the fully connected (FC) layers, we adopt a label-guided linear least squared regression (LSR) method. To generalize theFF design idea furthermore, we present the notion of "successive subspace leaming (SSLY and present a couple of concrete methods forimage and point cloud clasification. Extensive experimental results are given to demonstrate the competitlve performance of the SSL basedsystems. Similanties and dfferences between SSL and deep leaming (DL) are discussed.
Bio: Dr. C-C. Jay Kuo received his Ph.D. degree from theMassachusetts Institute of Technology In 1987. He is now with theUniversity of Southern California (USC) as Director of the MediaCommunications Laboratory and Distiguished Professor ofEletrial Engineering and Computer Sclence. His researchinterests are in the areas of media processing, compression andunderstanding. Dr. Kuo was the Editorin-Chief for the IEEE Trans,on Information Forensics and Security in 2012-2014. Dr. Kuois aFellow of AAAS, IEEE and SPIE. He has guided 150 students totheir Ph.D. degrees and supervised 29 postdoctoral researchfellows. Dr. Kuo is a co-author of 280 joumal papers, 920conference papers and 14 books. Dr. Kuo received the 2016IEEE Computer Society Taylor L. Booth Education Award, the2016 IEEE Circuits and Systems Society John Choma EducationAward, the 2016 IS&T Raymond C. Bowman Award, the 2017EEE Leon K. Kirchmayer Graduate Teaching Award, the 2017 IEEE Signal Processing Society Education Award, and the 2019IEEE Computer Society Edward J. McCluskey TechnicalAchlevement Award.