Dr. C.-C. Jay Kuo is a University of Southern California Distinguish Professor and Directory of its Media Communications Laboratory. Dr. Kuo received his Ph.D. from MIT in 1987. He has served as editor for 12 international journals and co-authored around 250 journal papers, 900 conference papers and 14 books.
He was the guest speaker at a lecture organized by the MGH Gordon Center. Below is the presentation summary provided through the courtesy of Dr. Kuo.
The superior performance of Convolutional Neural Networks (CNNs) has been demonstrated in many applications such as image classification, detection and processing. Yet, the CNN solution has its own weaknesses such as robustness against perturbation, scalability against the class number and portability among different datasets. Furthermore, CNN’s working principle remains a mystery. In this talk, Dr. Kuo first explained the reasons behind the superior performance of CNNs. Then, he presented an alternative solution, which is motivated by CNNs yet allows rigorous and transparent mathematical treatment, based on a data-driven Saak (Subspace approximation with augmented kernels) transform. The kernels of the Saak transform are derived from the second-order statistics of inputs in a one-pass feedforward way. Neither data labels nor backpropagation is needed in kernel determination. The pros and cons of CNNs and multi-stage Saak transforms were compared.