From convolutional analysis operator learning (CAOL) to convolutional neural network (CNN)MS40

The recent global approach can learn convolutional operators without storing many overlapping patches. We propose 1) a new CAOL framework in the global approach, and 2) a new convergent Block Proximal Gradient method using Majorizer (BPG-M) to solve block multi-nonconvex problems. Numerical experiments show acceleration of CAOL via BPG-M and the effectiveness of tight-frame filters learned via CAOL for sparse-view CT reconstruction. Finally, we illustrate mathematical models for CNN via CAOL.

This presentation is part of Minisymposium “MS40 - Recent Advances in Convolutional Sparse Representations
organized by: Giacomo Boracchi (Politecnico di Milano) , Alessandro Foi (Tampere University of Technology) , Brendt Wohlberg (Los Alamos National Laboratory) .

Il Yong Chun (University of Michigan)
Jeffrey A. Fessler (University of Michigan)
block proximal gradient method using majorizer, computed tomography, convolutional neural network, convolutional operator learning, deep learning, image enhancement, image reconstruction, image representation, inverse problems, machine learning, nonlinear optimization